RUNOFF PREDICTION AND ECOHYDROLOGICAL RECOVERY FOR

SMALL CATCHMENTS AFTER FIRE SOUTHERN CALIFORNIA

______

A Thesis

Presented to the

Faculty of

San Diego State University

______

In Partial Fulfillment

of the Requirements for the Degree

Master of Science

in

Civil Engineering

______

by

Brenton A. Wilder

Spring 2021

iii

Copyright © 2021 by Brenton A. Wilder All Rights Reserved

iv

DEDICATION

For my parents, Ian and Melissa, and my sister, Emma.

v

Purpose is an essential element of you. It is the reason you are on the planet at this particular time in history. Your very existence is wrapped up in the things you are here to fulfill … remember, the struggles along the way are only meant to shape you for your purpose. -- Chadwick Boseman

vi

ABSTRACT OF THE THESIS

Runoff Prediction and Ecohydrological Recovery for Small Catchments after Fire in Southern California by Brenton A. Wilder Master of Science in Civil Engineering San Diego State University, 2021

Over the past few decades, increasing fire frequency and severity in southern California - and across the western United States - has posed a concern to the safety and well-being of communities and ecosystems. Increased aridity coupled with water stressed vegetation from prolonged droughts are leading to a higher propensity for larger, more intense fires that directly impact ecohydrological processes such as streamflow and evapotranspiration (ET). Accurate characterization of these processes are required to improve rapid response efforts and resource management to promote resilient communities along the wildland-urban interface. This thesis presents methods to improve emergency rapid predictions of post-fire streamflow and characterization of ecohydrological recovery after fire. A random forest machine learning algorithm with 45 catchment parameters was created to predict post-fire peak streamflow during the period 1920 to 2019. By incorporating additional characteristics about meteorological and catchment properties, the random forest, flood forecasting technique provided more realistic predictions of peak streamflow in relation to Rowe et al. (1949), a commonly used flood frequency method. The time elapsed after fire, peak hourly rainfall intensity, and drainage area were important factors that increased accuracy of the random forest predictions. To improve vegetation assessments and resource management, two satellite-based ET products, ECOsystem Spaceborne Thermal Radiometer Experiment on Space Station (ECOSTRESS) Priestley-Taylor (PT-JPL) algorithm and Operational Simplified Surface Energy Balance Model (SSEBop), were used to evaluate conditions in relation to the 2018 Holy Fire. There was high uncertainty in post-fire ET between ECOSTRESS PT-JPL and SSEBop daily scaled ET due to the coarse spatial resolution of SSEBop and high spatial heterogeneity of the burn severity. To link recovery and hydrology, hydrologic signatures were quantified at the annual timescale for burned and unburned catchments. Post-fire water balance calculation for WY 2020 showed high uncertainty between ECOSTRESS PT-JPL and SSEBop, where differences in storage between catchments varied by over 500-mm depending on the model. Finally, ECOSTRESS PT-JPL was used to differentiate the landscape recovery by soil burn severity, vegetation species, slope aspect, and riparian area. The findings of this thesis improve upon our current methods in hydrologic modeling associated with fire in southern California.

vii

TABLE OF CONTENTS

PAGE

ABSTRACT ...... vi LIST OF TABLES ...... x LIST OF FIGURES ...... xi ACKNOWLEDGEMENTS ...... xiii CHAPTER 1 INTRODUCTION ...... 1 1.1 Motivation ...... 1 1.2 Research Objectives ...... 2 2 IMPROVING RAPID PREDICTIONS OF POST-FIRE PEAK STREAMFLOW FOR SMALL CATCHMENTS IN SOUTHERN CALIFORNIA ...... 4 2.1 Introduction to Post-Fire Hazards in Southern California ...... 4 2.2 Study Area ...... 7 2.2.1 Regional Geology ...... 9 2.2.2 Climate and Wildfire...... 10 2.2.3 Post-fire Processes ...... 10 2.3 Methods...... 12 2.3.1 Data ...... 12 2.3.2 Performance Measurements ...... 12 2.3.3 Peak Streamflow Modeling Methods...... 13 2.3.3.1 Flood Frequency of Historical Flows ...... 13 2.3.3.2 Rowe, Countryman, and Storey ...... 13 2.3.3.3 Random Forest Post-Fire Models ...... 15 2.4 Results ...... 16 2.4.1 Rowe, Countryman, and Storey Model...... 16 2.4.1.1 RCS Performance Before Fire ...... 16

viii

2.4.1.2 RCS Performance After Fire ...... 18 2.4.2 Random Forest Post-fire Model Performance ...... 19 2.4.3 Towards an Analytical Solution for Assessing Post-fire Peak Flows ...... 21 2.5 Discussion ...... 22 2.5.1 The Future Role of Rowe, Countryman, and Storey Model ...... 22 2.5.2 Role of Machine Learning as a Tool for Post-fire Risk Assessment ...... 24 2.5.3 Factors Influencing Post-fire Peak Streamflow ...... 25 2.5.4 Analytical Solution ...... 27 2.6 Conclusion ...... 28 3 CHARACTERIZING FIRE SEVERITY AND ECOHYDROLOGICAL RECOVERY FOR THE 2018 HOLY FIRE IN SOUTHERN CALIFORNIA ...... 29 3.1 Introduction ...... 29 3.2 Materials and Methods ...... 31 3.2.1 Study Area and Hydrologic Data ...... 31 3.2.2 Remote Sensing and Spatial Products...... 33 3.2.3 Climatology...... 35 3.2.4 Ecohydrological Analysis ...... 36 3.2.4.1 Correlation of SSEBop and ECOSTRESS PT-JPL ...... 36 3.2.4.2 Pre-fire Above-ground Biomass and Post-fire Vegetation Recovery ...... 36 3.2.4.3 Hydrologic Signatures ...... 37 3.3 Results ...... 38 3.3.1 Climate and ET Processes ...... 38 3.3.2 Pre-Fire Above-Ground Biomass ...... 40 3.3.3 Correlation of SSEBop and ECOSTRESS PT-JPL with Respect to the Holy Fire ...... 41 3.3.4 Post-fire Ecohydrology and Hydrologic Signatures ...... 42 3.3.5 Spatial and Temporal Recovery of Post-Fire ET ...... 45 3.4 Discussion ...... 46 3.4.1 Role of Climate Extremes ...... 46 3.4.2 High Uncertainty in Post-Fire ET ...... 47

ix

3.4.3 Linking Ecohydrological Recovery After Fire to Observed Data using Hydrologic Signatures ...... 48 3.4.4 Post-fire Revegetation and Recovery Strategies after the Holy Fire ...... 49 3.5 Conclusion ...... 50 4 SUMMARY ...... 52 REFERENCES ...... 54 APPENDIX A SUPPLEMENTARY FIGURES FOR CHAPTER 2 ...... 65 B SUPPLEMENTARY FIGURES FOR CHAPTER 3 ...... 70

x

LIST OF TABLES

PAGE

Table 2.1. Study Catchment Characteristics (Name, Size, Wildfire, and Burn Severity) by Region, Where Burn Severity is Categorized by Unburned‐Low (Un‐Low), Moderate (Mod), and High (High) ...... 8 Table 2.2. Rowe, Countryman, and Storey (RCS) Bias for Peak Streamflow Predictions by Region for the 2‐ and 10‐year Events ...... 18 Table 3.1. Catchment Parameters and Vegetation Types for Paired Catchments Santiago and Coldwater ...... 33 Table 3.2. Annual Hydrologic Signatures (runoff-ratio [RO], and Richards-Baker [RB] index) for WY 2014-2020 for Coldwater and Santiago. Cumulative EVI (ΣEVI) and Annual SSEBop ET (ETAnnual) are Also Noted ...... 45 Table A.1. List of Products used to Find or Derive Data (column 1), Data Type and Motivation (Column 2) ...... 66

xi

LIST OF FIGURES

PAGE

Figure 1.1. Total area burned by county (data source: Moderate Resolution Imaging Spectroradiometer Burned Area Monthly Global at 500-m (2000-2020)...... 1 Figure 2.1. Study area consisted of 33 catchments in 6 different regions within the Transverse Ranges and Peninsular Ranges in southern California, USA...... 7 Figure 2.2. Rowe, Countryman, and Storey (RCS) unburned peak streamflow predictions compared to observed peak streamflow for 2‐year (a) and 10‐year (b) return periods...... 17 Figure 2.3. Sediment‐laden flows and high water mark observations in a concrete lined channel at Toft Dr. near the outlet of Dickey Canyon during 29 November (a and b) and 6 December 2018 storms (c)...... 19 Figure 2.4. Predictions for the RF‐5 random forest model by region...... 20 Figure 2.5. Observed peak streamflow versus predicted streamflow response...... 21 Figure 2.6. Modeled regression for small, steep catchments during the first year after fire in southern California, fitted to historic high flows within the first year after wildfire collected from 1920–2019...... 22 Figure 2.7. Observed peak streamflow per unit area with respect to peak hourly rainfall intensities (a), catchment size (b), and days after fire containment (c)...... 27 Figure 3.1. Location of Santiago (control) and Coldwater (burned) catchments and the 2018 Holy Fire soil burn severity; streamflow and precipitation gauges are also denoted...... 31 Figure 3.2. Standardized precipitation index (SPI) for Upper Silverado station (WY 1991-2020) are shown on the primary axis...... 39 Figure 3.3. Average SSEBop monthly evapotranspiration (ET) at 1-km spatial resolution for the Holy Fire (August 2018) burn area before the fire (WY 2001 to 2018) and for the available record of ET (WY 2001-2011)...... 40 Figure 3.4. Pre-fire average ΣEVI (WY 2014-2017) are shown on the primary axis with respect to their soil burn severity classification, where n represents the average number of pixels for each sample in Coldwater...... 41 Figure 3.5. ECOSTRESS PT-JPL and SSEBop scatter plot (n = 103 pixels) for pre- fire on August 2, 2018 (a), 1-year post-fire on August 17, 2019 (b), and 2- years post-fire on October 3, 2020 (c)...... 42

xii

Figure 3.6. Coldwater (CW) and Santiago (Sant) daily rainfall (a), difference in daily evapotranspiration (ET) from WY 2019-2020 (b), and daily streamflow (c). Difference in ECOSTRESS PT-JPL ET is calculated as Sant ET minus CW ET...... 44 Figure 3.7. Cumulative ECOSTRESS evapotranspiration (ET) collected from WY 2019 (35 images) and WY 2020 (76 images) for the Holy Fire with respect to slope aspect (a), soil burn severity (b), riparian versus hillslope (c), and pre- fire vegetation species (d)...... 46 Figure A.1. Relative importance for random forest calibration...... 67 Figure A.2. Random Forest model with 45 parameters (RF-45) ...... 68 Figure A.3. Random Forest model with 5 most important parameters (RF-5) ...... 69 Figure A.4. Control volume diagram for WY 2020 water balance calculation ...... 71 Figure A.5. ECOSTRESS PT-JPL daily evapotranspiration (ET) of 4-days before the fire on August 2, 2018 (a), SSEBop monthly ET 1-month before the fire in July 2018 converted to daily ET (b), ECOSTRESS approximately 1-year after the fire on August 17, 2019 (c), SSEBop approximately 1-year after the fire in August 2019 (d), ECOSTRESS approximately 2-years after the fire on October 3, 2020 (e), and SSEBop approximately 2-years after the fire in October 2020...... 72 Figure A.6. Field photos from 2018 Holy Fire showing 1 year of revegetation...... 73

xiii

ACKNOWLEDGEMENTS

I would like to thank my advisor, Dr. Alicia Kinoshita, for guiding the methodologies and progression of this work. Her extensive knowledge in the field of post-fire hydrology has provided a great foundation the past two and a half years of which to build from. Also, I would like to thank the CAL FIRE & CGS team including Jeremy Lancaster, Brian Swanson, Don Lindsay, Bill Short, Peter Cafferata, and Drew Coe for their mentorship and contributions to this project. It was a huge honor to conduct research with this group. Also, I would like to thank my thesis committee Dr. Hassan Tavakol-Davani & Dr. Hilary McMillan for their direction throughout my time as a graduate student and for critically reviewing this work. Being introduced to optimization techniques and hydrologic modeling in their graduate lectures was an experience that I am extremely grateful for. Thank you also to Christine Lee and Paa Sey from the ECOSTRESS project for their collaboration. Lastly, I would like to thank all the members of the Disturbance Hydrology Lab and Blue Gold Graduate Group for their insight and feedback during our many research meetings and conferences. I will cherish all of the great memories we shared, and I am so thankful to have been a part of these groups. I would also like to acknowledge the Joint Fire Science Program for partially funding this thesis through the Graduate Research Innovation Award #19-1-01-55. Additionally, thank you to San Diego State University for awarding the Master’s Research Scholarship for this work.

1

CHAPTER 1

INTRODUCTION

1.1 MOTIVATION California is the 3rd largest state in terms of area and has 58 counties that range climatically from semi-arid, to Mediterranean, to desert, and are conducive to wildfire. Across the State of California, the proportion of burned area from 2000-2020 in relation to county area ranged from 0-65% (Figure 1.1).

Figure 1.1. Total area burned by county (data source: Moderate Resolution Imaging Spectroradiometer Burned Area Monthly Global at 500-m (2000-2020).

2

Some counties in northern California had proportions of burned area that were over 50%. This work focuses on understanding the impacts of wildfire in southern California which has experienced some of the most destructive fires in the state including the 2003 Cedar Fire (3rd most destructive). High proportions of total wildfire area (2000 to 2020) with respect to county area in southern California, include Ventura County (50%, 5th highest in the state), San Diego County (37%), Los Angeles County (28%), Santa Barbara County (22%), and Orange County (15%), over the past 20 years. These frequent fires lead to changes in the post-fire landscape that produce elevated peak streamflow in small catchments (Neary et al., 2005) and spatially and temporally variant ecohydrological recovery (Kinoshita & Hogue, 2011). Under changing climate and wildfire regimes, accurate predictions of post-fire peak streamflow rates and ecohydrological recovery are critical for key emergency response and management agencies such as the California Department of Forestry and Fire Protection (CAL FIRE), California Department of Conservation – California Geological Survey (CGS), the U.S. Forest Service (USFS), and local county flood control districts who seek to mitigate risks associated with post-fire flooding and erosion. Models used to predict these processes have varying degrees of sophistication (Atchley et al., 2018; Cannon et al., 2004; Kinoshita et al., 2014; Robichaud et al., 2007; Wilder et al., 2021). Due to the short time period between fire and flood in southern California, simpler models such as Rowe et al. (1949) are prioritized for rapid assessments of risks to downstream communities and ecosystems (Kinoshita et al., 2014; Wilder et al., 2021). These simpler models such as Rowe et al. (1949) remain largely unvalidated. Further, long-term vegetation assessments are essential for agencies such as the USFS to better inform decision making on recreational services infrastructure such as trail and road access to the public. Therefore, there is a need to study these existing methodologies and investigate improved prediction tools to assist rapid and long-term risk assessments to better protect downstream communities and ecosystems.

1.2 RESEARCH OBJECTIVES To address the existing needs and research gaps outlined in the introduction, this thesis will be presented in two parts:

3

Objective 1: It is hypothesized that the Rowe et al. (1949) model, also known as RCS, does not accurately predict post-fire peak streamflow at the small catchment scale (1 to 50 km2), a common scale that is crucial for post-fire risk assessment. To evaluate and improve the accuracy of decision support tools for post-fire hazard prediction and management in southern California, this work will assess the validity of the RCS flood frequency method using observed pre- and post-fire peak streamflow. This work will then identify important parameters that can improve the characterization of post-fire responses using machine learning, and present an analytical tool for post-fire risk assessments. • Hypothesis 1A: If Rowe et al. (1949) look-up table methodology is evaluated against pre- and post-fire streamflow observations when using the 2- and 10-year recurrence interval storm events, then the predictions will have a large degree of error due to the static nature of the model and the disregard of important variable such as soil burn severity and rainfall intensity. • Hypothesis 1B: If literature-based hydrologic parameters are assembled for the study area (parameters=45), aggregated at the catchment scale, and inputted into a random forest machine learning model (n=74), then the random forest model will perform with improved accuracy compared to the Rowe et al. (1949) methodology due to the usage of more important parameters (i.e., rainfall intensity) that better characterize the physical processes. Objective 2: It is hypothesized that evaporative stress was high in areas leading up to the 2018 Holy Fire, increasing the severity of the fire. To evaluate this, as well as the ecohydrological recovery, with respect to space and time, SSEBop (Operational Simplified Surface Energy Balance Model) and ECOSTRESS (ECOsystem Spaceborne Thermal Radiometer Experiment on Space Station) Priestley-Taylor (PT-JPL) models will be used to analyze evapotranspiration for small catchments after fire. • Hypothesis 2A: Based on historical monthly SSEBop evapotranspiration data (1- km resolution) and precipitation, there will be an observed increase in evaporative stress leading up to the Holy Fire due to the prolonged drought, further exacerbating the severity of the fire. • Hypothesis 2B: ECOSTRESS PT-JPL daily evapotranspiration data (70-meter resolution) will be used to differentiate ecohydrological recovery using hydrologic signatures following the Holy Fire to improve vegetation assessments after fire.

4

CHAPTER 2

IMPROVING RAPID PREDICTIONS OF POST- FIRE PEAK STREAMFLOW FOR SMALL CATCHMENTS IN SOUTHERN CALIFORNIA

2.1 INTRODUCTION TO POST-FIRE HAZARDS IN SOUTHERN CALIFORNIA Anthropogenic climate change, human ignited fires, and increased fuel loads resulting from decades of fire suppression in forested areas have contributed to an increase in wildfire severity and occurrence in the western U.S. (FRAP, 2018; Radeloff et al., 2005; Westerling et al., 2006)⁠. These wildfires can significantly impact land cover and soil properties (Debano, 2000; Moody, 2012; Moody & Martin, 2001; Neary et al., 2005; Rowe et al., 1949) and can pose risks to human lives, valued assets, and ecosystems (Kinoshita et al., 2016; Shakesby et al., 2016)⁠. Development in wildlands and at the wildland‐urban interface has further exacerbated the potential to impact communities. In California, extreme increases in peak streamflow after wildfire contribute to catastrophic flooding and debris flows on steep alluvial fan landforms that may be developed with suburbs and infrastructure. For example, the Holy Fire (August 2018) burned about 94 km2 in Orange County and Riverside County in California and affected over 1000 residents. Three months later, the Woolsey and Hill fires (November 2018) burned approximately 392 km2 in Los Angeles County and Ventura County. Combined, these two fires resulted in three fatalities, eight non‐ fatal injuries, and 1665 structures being destroyed. Quickly following these wildfires (December 2018–February 2019), record amounts of rainfall resulted in flooding and debris flows in the burned regions. During this time, the Riverside County Fire Department ordered the evacuation of 300 homes. Furthermore, on January 9th, 2018, while the Thomas Fire was still burning, Santa Barbara County was impacted by post‐fire debris flows that killed 23

5

people, destroyed or damaged 558 structures, and caused severe damage to infrastructure in Montecito and Carpinteria, California (Kean et al., 2019; Lancaster et al., in press). Federal, state, and local agencies rapidly assess burned catchments to produce timely emergency protection measures for nearby communities that may be in the immediate danger from elevated peak flows and debris flows. Federal Burned Area Emergency Response (BAER) teams and California State Watershed Emergency Response Teams (WERTs) are reliant on modeling to make informed decisions (Foltz et al., 2009)⁠ following field verification of soil burn severity categories (Parsons et al., 2010)⁠. The models used to predict these processes have varying degrees of sophistication (Cannon et al., 2004; Kinoshita et al., 2014; Robichaud et al., 2007)⁠. Debris flows, debris yields, and surface erosion estimates are based on either numerical modeling (Water Erosion Prediction Project derivatives) or relatively robust empirical models, where the independent variables are consistent with our physical understanding of geomorphic processes (Gartner et al., 2014; Robichaud et al., 2011; Staley et al., 2017)⁠. WERTs and BAER teams inventory values‐at‐risk (VARs) and suggest emergency protection measures that can be rapidly implemented, including early warning system use, storm patrol, structure protection with sandbags or K‐rails, channel clearance work near crossings, signage to close low water crossings, parks and trails, and road crossing upgrades (Foltz et al., 2009)⁠. Emergency response teams also rapidly communicate VAR locations and emergency protection measures to local agencies such as county department of public works and flood control districts for implementation. Accurate predictions of post‐fire peak streamflow rates are critical for key emergency response and management agencies such as the California Department of Forestry and Fire Protection (CAL FIRE), California Department of Conservation‐California Geological Survey (CGS), the U.S. Forest Service (USFS), and local county flood control districts especially under changing climate and wildfire regimes. Peak fire season in southern California occurs in the fall, creating a short period between fire and winter storms that produce flood events. This pattern requires effective hazard evaluation to minimize risk for downstream communities, infrastructure, and ecosystems. Further, evidence suggests that California's peak fire season could be shifting later in the year to November–December due to climate change affecting coastal temperature, air pressure, and humidity (Miller & Schlegel, 2006)⁠. This emphasizes the urgency for federal, state, and local agencies to rapidly

6 assess burned catchments and produce timely emergency protection measures for nearby communities. Post‐fire peak streamflow can exceed unburned peak streamflow by three orders of magnitude during similar pre‐fire storm events (Moody & Martin, 2001; Wagenbrenner, 2013; Wohlgemuth, 2016). In southern California, there is substantial reliance on the look‐ up‐tables (LUTs) developed by Rowe et al. (1949) to predict post‐fire peak flows. This simpler method is well understood and convenient for rapid assessment in comparison to more complex process‐based hydrologic models, and therefore, is the most widely used method for rapidly predicting post‐fire peak flow rates in southern California (Kinoshita et al., 2014)⁠. Notable fires such as the 2003 Old and Grand Prix Fires, 2009 Station Fire, 2018 Holy Fire, 2018 Woolsey and Hill fires, 2019 Saddle Ridge Fire, and 2019 Cave Fire have utilized the Rowe, Countryman, and Storey (RCS) methods for post‐fire peak streamflow predictions (e.g., Biddinger et al., 2003; Moore et al., 2009; WERT, 2018a, 2018b, 2019a, 2019b). It is hypothesized that the RCS model does not accurately predict post‐fire peak streamflow at the small catchment scale (1 to 50 km2), a common scale needed for post‐fire risk assessment. Building upon the significant advances in post‐fire hydrology since the development of RCS, there is an opportunity to improve current methodologies for prediction of peak streamflow in small catchments. In particular, machine learning techniques can be used for post‐fire applications (Saxe et al., 2018; Schmidt et al., 2020)⁠. Saxe et al. (2018) utilized machine learning to study a vast area with varying catchment scales. This study focuses on small to medium‐sized catchments ranging from 1 to 42 km2 to reduce error associated with scaling, and to improve regional responses when using machine learning. To evaluate and improve the accuracy of decision support tools for post‐fire hazard prediction and management in southern California, we 1) assess the validity of the RCS method using observed pre‐ and post‐fire peak streamflow, 2) identify important parameters that can improve the characterization of post‐fire responses using machine learning, and 3) present an analytical tool for post‐fire risk assessments.

7

2.2 STUDY AREA The study area is within the Transverse Ranges and Peninsular Ranges, which are geomorphic provinces with a wide topographic range from sea level to 3506-m (CGS, 2002). The study area straddles the southern California Coastal and southern California Mountain and Valleys ecoregions (Cleland et al., 2007)⁠. Approximately 60%–70% of these ecoregions are dominated by shrubland (e.g., chaparral) vegetation types, with a lesser proportion occupied by grassland and woodland or forest vegetation types (FRAP, 2015). The median pre‐settlement fire return intervals for these vegetation types are variable, ranging on average from approximately 60 to 100 years for chaparral and coastal scrub vegetation types (Van de Water & Safford, 2011)⁠. This research investigated 33 small catchments with drainage areas ranging between 1.2 to 41.7 km2 (Figure 2.1;Table 2.1).

Figure 2.1. Study area consisted of 33 catchments in 6 different regions within the Transverse Ranges and Peninsular Ranges in southern California, USA.

8

Table 2.1. Study Catchment Characteristics (Name, Size, Wildfire, and Burn Severity) by Region, Where Burn Severity is Categorized by Unburned‐Low (Un‐Low), Moderate (Mod), and High (High) Un‐ Catchment Name (RCS Table Mod Size (km2) Wildfire (% Burned Area) Low High (%) No.) (%) (%) San Bernardino Mountains Devil (60) 14.2 2003 Old (97%) 50% 31% 19% East Twin Creek (57) 22.6 2003 Old (98%) 50% 31% 19% Waterman (58) 11.9 1980 Panorama (86%) n/a n/a n/a Santa Ynez Mountains Maria Ygnacio (204) 16.6 1990 Painted Cave (52%) 65% 26% 9% Franklin (188) 6.9 1971 Romero (56%) n/a n/a n/a Coyote (183) 33.9 1985 Wheeler (98%) 39% 33% 28% Mission (200) 17.1 2009 Jesusita (69%) 60% 32% 8% San Jose (205) 14.4 1955 Refugio (69%) n/a n/a n/a Carpinteria (187) 33.9 1971 Romero (84%) n/a n/a n/a Santa Ana (182) 24.2 1985 Wheeler (100%) 39% 33% 28% Romero (193) 5.14 2017 Thomas Fire (94%) 59% 39% 2% Toro (190) 2.33 2017 Thomas Fire (94%) 59% 39% 2% Arroyo Paredon (191) 2.22 2017 Thomas Fire (83%) 59% 39% 2% San Jacinto Mountains Snow (85) 27.9 1973 One Horse (37%) n/a n/a n/a 1980 Dry Falls (92%) n/a n/a n/a Andreas (82) 22.4 2013 Mountain Fire (67%) 49% 49% 2%

San Gabriel Mountains Little Santa Anita (136) 4.8 1954 Monrovia Peak (100%) n/a n/a n/a Sawpit (131) 13.7 1924 San Gabriel #2 (73%) n/a n/a n/a Fish (129) 18.3 1968 Canyon Inn (100%) n/a n/a n/a Lower Big Dalton (117) 18.8 1919 San Gabriel (100%) n/a n/a n/a Haines (145) 3.3 1933 La Crescenta (51%) n/a n/a n/a Arroyo Seco (141) 41.7 2009 Station (100%) 34% 38% 28% Santa Anita (137) 25.2 1954 Monrovia Peak (97%) n/a n/a n/a Santa Ana Mountains Horsethief (35) 5.6 2018 Holy (100%) 15% 71% 14% Indian (35) 7.3 2018 Holy (100%) 15% 71% 14% Rice (35) 5.0 2018 Holy (100%) 15% 71% 14% Dickey (34) 1.2 2018 Holy (98%) 15% 71% 14% Coldwater (36) 10.7 2018 Holy (99%) 15% 71% 14% Agua Chinon (32) 7.1 2007 Santiago (92%) 61% 28% 11% Santiago (32) 32.3 2007 Santiago (70%) 61% 28% 11% San Diego County Los Coches (14) 31.6 2003 Cedar (57%) 24% 24% 52% Pechanga (28) 34.6 2000 Pechanga (45%) 32% 34% 34% Fallbrook (27) 18.3 2014 Tomahawk (59%) 42% 52% 6% Rainbow (27) 26.7 2007 Rice (20%) 75% 17% 8% Note: Note that burn severity data collected from monitoring trends in burn severity (MTBS) are not available for fires prior to 1984, which are indicated by n/a.

9

2.2.1 Regional Geology The Transverse Ranges are characterized by a 480‐km long east–west‐trending series of steep mountain ranges, which include the Santa Ynez, San Gabriel, and San Bernardino mountains in the study area. They are underlain by diverse rock types of different ages. Older Precambrian to Cretaceous metamorphic and igneous rocks underlie the western San Bernardino and central San Gabriel mountains (CGS, 2018). In contrast, the Santa Ynez Mountains are underlain by younger marine and continental sedimentary sequences of sandstone, siltstone, and shale, deposited during late Cretaceous and Cenozoic times (CGS, 2018). The Peninsular Ranges are characterized by a series of northwest‐trending mountain ranges, including the Santa Ana, San Jacinto, and Laguna mountains, having generally lower average elevation than the Transverse Ranges. They are characterized by igneous and metamorphic rocks of Mesozoic age, including granite in the east and volcanic and metasedimentary rocks in the west (CGS, 2018; Harden & Matti, 1989)⁠. Mountain ranges in these provinces have steep canyons with active hillslope processes including shallow and deep‐seated landslides, rockfall, and debris flow initiation. Pleistocene‐ and Holocene‐age alluvial and debris fan landforms deposited at the mountain front throughout these provinces demonstrate dominant catchment runoff and sedimentation processes ranging from sediment‐laden stream flows to debris flows. At the catchment scale, many geologic factors control sediment availability in the region. These include rock type (lithology), structure, presence of landslides, weathering characteristics, elevation, slope, and tectonic history (DiBiase & Lamb, 2020; Lavé & Burbank, 2004; Scott & Williams, 1978). Rock to regolith conversion is rapid in the study area, thus, shallow landslides, dry ravel, and rill erosion after wildfire may dominate the sediment supply. Conversely, where regolith conversion is slow, deep‐seated landslides may dominate hillslope processes with landslide deposits and their over‐steepened source areas as primary sediment sources. To represent this in the model, soil erodibility factors and landslide susceptibility maps were used as parameters for analysis.

10

2.2.2 Climate and Wildfire The Transverse Ranges and Peninsular Ranges experience some of the highest storm precipitation totals in the nation (Dettinger, 2011; O’Connor & Costa, 2004; Ralph & Dettinger, 2012)⁠. Further, the unique east–west alignment of the Transverse Ranges relative to the Pacific Ocean translates to large amounts of orographic precipitation from atmospheric rivers, where 10%–30% of the annual precipitation can originate from one large storm (Lamjiri et al., 2018)⁠ during cool season months of November–April (Dettinger, 2011; Neiman et al., 2008)⁠. Multi‐year drought interspersed with, or followed by, extreme precipitation or wetter than average years are common in the area, a pattern conducive to growth and then desiccation of the region's fire‐prone vegetation. The combination of steep and complex terrain, combustible fuels, prolonged dry seasons, and strong Santa Ana wind events produce the most intense fire climate in the United States (Keeley et al., 2004; Raphael, 2003; Wells, 1987; Wells, 1981)⁠. Further, population growth proximal to wildlands means that much of southern California has experienced higher fire frequency relative to pre‐settlement return intervals (Safford & Van de Water, 2013; Syphard et al., 2007)⁠. While southern California's chaparral and forested ecosystems are fire‐adapted, climate change is predicted to increase mean annual burned area by 42% for forests in the San Gabriel Mountains under RCP 8.5 scenario (Representative Concentration Pathways) between the initial period of 1950 to 2005 and the projected period of 2006 to 2099 (https://cal-adapt.org/). One study observed no apparent increase in summer wildfire in non‐ forested areas between 1972 to 2018, however, observed a slight increase in fall wildfire probability due to atmospheric aridity (Williams et al., 2019)⁠, which is exemplified by the fall fires in 2017 and 2018, including the 2017 Thomas Fire, the 7th largest wildfire documented in southern California.

2.2.3 Post-fire Processes The regional climatology, geology, soils, and vegetation in southern California equates to an unmatched magnitude of post‐fire hydrogeomorphic response (Moody et al., 2013)⁠. Examples of extreme impacts include the 1934 New Year's Day debris flow in the La Crescenta area (Chawner, 1935)⁠; the 1981 Mill Creek disaster in the San Gabriel Mountains

11

(Shuirman & Slosson, 1992); the December 25th debris flows following the 2003 Old Fire (Cannon & DeGraff, 2009)⁠; and, hyperconcentrated flood flows following the 2018 Holy Fire causing damage to homes, flood control facilities, and highway infrastructure in (and below) the Santa Ana Mountains. For example, observed hyperconcentrated flood flow velocities varied greatly by catchment during the 29 November 2018 storm following the 2018 Holy Fire with flows at Dickey Canyon producing a flow velocity of approximately 12 m/s, Rice, Horethief, and Indian Canyons producing flow velocities between 5–8 m/s, and Coldwater Canyon producing a flow velocity of approximately 3 m/s. Short duration, high intensity precipitation produced from atmospheric rivers and orographic lift is a key factor that generates flash floods and debris flows after wildfire (Cannon et al., 2008; Moody et al., 2013; Oakley et al., 2017)⁠. Moreover, precipitation intensity is anticipated to increase about 7% per degree Celsius of warming, suggesting a potential increase of flash flood and debris flow magnitude after wildfire in a changing climate (Prein et al., 2017)⁠. In southern California, the dominant vegetation types covering the hills and lower mountain slopes are the chaparral and scrub brushes. Further, the closed canopy nature of these vegetation types means that large scale, stand‐replacing fires are common (Keeley, 2009)⁠. After stand‐replacing fires on steep slopes, dry ravel increases, with much of it delivered directly into the channel network resulting in large increases in sediment yield and/or the frequency of debris flows (DiBiase & Lamb, 2020)⁠. Since post‐fire runoff reflects a continuum between clear water flow and debris flows (non‐Newtonian fluid), the recognition of these linked processes is vital for understanding the fire‐flood cycle in southern California (Travis et al., 2012)⁠. The well‐established linkage between chaparral vegetation and the development of soil water repellency (Debano, 2000; Doerr et al., 2000)⁠ influences infiltration, runoff, and erosion (Doerr et al., 2009)⁠. Soil water repellency and soil erosion are enhanced with increasing soil burn severity (Huffman et al., 2001; Lewis et al., 2006)⁠ due to changes in mechanical and hydraulic properties such as changes in relative density, saturated hydraulic conductivity, and sorptivity (Shakesby et al., 2016)⁠. Thus, soil burn severity is crucial for modeling of hydrologic response after wildfire in southern California.

12

2.3 METHODS

2.3.1 Data To develop the random forest (RF) (Breiman, 2001)⁠ algorithm to characterize historical post‐fire peak streamflow in southern California, 45 catchment parameters (Table A.1) were derived from local and national data bases. These sources included streamflow records from the United States Geological Survey (USGS), hourly precipitation data from National Oceanic and Atmospheric Administration (NOAA), the average annual precipitation (USGS StreamStats), burn perimeters and severity from Monitoring Trends in Burn Severity (MTBS), soils data from the USDA Natural Resources Conservation Service's Soil Survey Geographic Database (SSURGO), land cover data from the National Land Cover Database (NLCD), and digital elevation models (DEM) from USGS Elevation Products. Selection of these parameters were based on a review of studies that directly linked the effects of specific parameters on pre‐ and post‐fire processes. Some of the datasets were not available for all post‐fire events, specifically, burn severity datasets were only available after 1984.

2.3.2 Performance Measurements To measure model performance, the following statistics were used to estimate the bias, root mean squared error (RMSE), and coefficient of determination (R2):

Bias = ( ) (2.1) 1 𝑛𝑛 𝑛𝑛 𝑖𝑖=1 𝑖𝑖 𝑖𝑖 ∑ 𝑃𝑃 − 𝑂𝑂

( ) RMSE = (2.2) 2 1 𝑛𝑛 𝑃𝑃𝑖𝑖−𝑂𝑂𝑖𝑖 �𝑛𝑛 ∑𝑖𝑖=1 𝑛𝑛 where n is the number of catchments, i is the i‐th catchment, Oi is observed peak streamflow for i‐th catchment, and Pi is predicted peak streamflow for i‐th catchment. Positive bias represents an overprediction and a negative bias represents an underprediction. Larger RMSE values are associated with higher error than those with lower RMSE values.

( ) 2 1 R = 𝑛𝑛( 2) (2.3) ∑𝑖𝑖=1 𝑒𝑒𝑖𝑖 𝑛𝑛 2 − ∑𝑖𝑖=1 𝑦𝑦𝑖𝑖−𝑦𝑦

13

where ei is error at point i, yi known streamflow at point i, and y̅ is average streamflow. Higher R2 values that are closer to 1, or equal to 1, represent better agreement between the predicted and observed values.

2.3.3 Peak Streamflow Modeling Methods

2.3.3.1 FLOOD FREQUENCY OF HISTORICAL FLOWS RCS‐derived pre‐ and post‐fire predictions were compared to historical streamflow records. The observed peak streamflow values are based on statistical distributions and not individual observations. Weibull and PeakFQ were used to estimate flood frequency for existing streamflow records (Clarke, 2002; Flynn et al., 2006)⁠. Weibull is a generalization of the exponential distribution to calculate peak streamflow for different flood frequencies:

T = (2.4) 𝑛𝑛+1 𝑚𝑚 where T is return period in years, n, is the number of data points and, m, is the rank of highest annual streamflow. The Peak FQ program was developed by the USGS and utilizes a log‐Pearson Type III distribution to model flood frequencies, based on Bulletin 17C (Parrett et al., 2011) procedures:

logQ = + (2.5)

where logQ, is the logarithmic peak streamflow,𝑥𝑥 x̅ , is𝑆𝑆𝑆𝑆 the sample logarithmic mean, S, is the sample logarithmic SD and, k, is the station skew coefficient (Flynn et al., 2006)⁠. The average percent difference between Weibull and PeakFQ results were 33% for the 2‐year return period and 41% for the 10‐year return period. The two methods were averaged to form a statistical representation of the observed peak streamflow for each catchment for the 2‐ and 10‐year return periods for comparison of RCS flood frequency pre-fire predictions.

2.3.3.2 ROWE, COUNTRYMAN, AND STOREY Rowe et al. (1949) undertook extensive observations across southern California catchments (Mexican border to San Luis Obispo) and developed relations for size of peak streamflow events and erosion rates associated with normal (unburned) conditions for 256

14

catchments within five climatic zones. Relations were established between storm precipitation and post‐fire peak discharge for catchments from the years 1869 to 1949 in each specific storm zone. It should be noted that several large storm events occurred after 1950 that were not recorded during the RCS study, which would have likely influenced the model significantly. The changes in these flows for subsequent post‐fire years were determined and are the basis for the 256 LUTs. Since discharge measurements were not available for all catchments, rating curves were established for each storm zone considering size, shape, steepness, stream channel characteristics, infiltration and water storage capacities of various soil‐geologic formations, precipitation, and vegetation characteristics (Rowe et al., 1949). Each of the 256 tables were adjusted according to percent burnable area and vegetation density, and includes post‐fire flood frequency predictions for 1, 2, 3, 7, 15, 30 and 70 years (normal) after burning for different recurrence intervals. The San Bernardino storm zone is the only region where RCS includes the influence of snowpack in streamflow estimates. The RCS normal and burned peak streamflow estimates for the 33 catchments were compared to the logarithmic means of the basic frequency classes (tables A and B of Rowe et al., 1949). A Savitzky–Golay function (Orfanidis, 1996) was used to remove noise from the data and to create a smooth curve for the storm frequency relationship. Data were plotted for 1, 2, 3, 7, 15, 30 and 70 years after the fire. Power functions were fitted to these relationships, where R2 values were greater than 0.95, with the form:

Q = (2.6) 𝑏𝑏 where Q is the predicted peak streamflow from𝑎𝑎 RCS,𝑝𝑝 a and b are coefficients, and p is the exceedance probability. Also, the RCS 1949 linear function was used to calculate the effect of partial burn on peak streamflow by proportioning with the percent of burnable area. The 2‐ and 10‐year recurrence interval predictions to assess catchments are typically used by WERT and BAER teams because there is more confidence in flood flow prediction methods at smaller recurrence intervals compared to larger intervals such as 25‐, 50‐ 100‐ year (Kinoshita et al., 2014)⁠. Additionally, the recovery period for local vegetation in southern California is typically 2–7 years, which influences the hydrologic response (Bell et al., 2009; Keeley & Keeley, 1981; Kinoshita & Hogue, 2011)⁠. Thus, post‐fire observations were derived from available streamflow observations collected within the first 7 years after

15

fire for all of the catchments and compared to the flood frequency model for 2-year and 10- year events.

2.3.3.3 RANDOM FOREST POST-FIRE MODELS Random forests (RF) are an effective machine learning (ML) tool for prediction, both in classification problems and regression (Breiman, 2001)⁠. Two random forest models were developed into flood forecasting models to predict post‐fire peak streamflow. The MATLAB R2019a RF TreeBagger function combines the results of many decision trees (without replacement) and was used to build a regression based on the 45 catchment parameters, RF‐ 45. The sample size was increased from 33 to 74 by selecting one to three post‐fire storm events for each catchment. In general, tree‐structured classifiers are independently distributed and explore different class conditions at each input. One of the main advantages of random forest is reproducibility and high transparency of feature importance. Random forests' feature importance is easy to interpret directly from the model, allowing conclusions to be drawn for the impact of different catchment parameters on the accuracy of predictions. TreeBagger also requires a low amount of hyperparameter tuning and the features do not need to fit a normal distribution, which is present in post‐fire peak flow data. However, a common issue when building an RF model is the tendency for models to over‐fit to the data, translating to a smaller variance in the predictions. This can occur if there are too many features in the regression, which falsely implies that the model is predicting perfectly. TreeBagger helped to reduce overfitting of data by using Bootstrap‐aggregated (bagged) decision trees, which combined the results of many decision trees, where the Strong Law of Large Numbers applied (Pal, 2005; Feller, 1968). Redundant or unimportant parameters were filtered out, resulting in a multi‐variable model with only five important parameters to predict post‐fire peak streamflow (RF‐5). RF‐5 was validated using a k‐fold cross validation with k = 5. The accuracy of RF-45 and RF-5 were assessed across each region and each catchment by using root mean squared error and coefficient of determination [Equations (2.2) and (2.3)] to evaluate model performance in comparison to observed streamflow data. The most important parameters identified by ML were also used to create a regression model using a three‐ dimensional polynomial function.

16

2.4 RESULTS

2.4.1 Rowe, Countryman, and Storey Model

2.4.1.1 RCS PERFORMANCE BEFORE FIRE Only 24 of the 33 catchments had observed pre‐fire streamflow, which are shown as streamflow per unit area for the 2‐ and 10‐year recurrence interval events (Figure 2.2). Pre‐ fire, RCS over‐predicted the 2‐year return period for 17 of the catchments and under‐ predicted the 10‐year return period for 18 of the catchments. In total, the 2‐year return period had a positive bias of 0.162 cms/km2 and the 10‐year return period had a negative bias of −0.791 cms/km2 (Table 2.2). There was substantial variation between catchments that the RCS method did not represent. For example, the SD for the RCS 10‐year return period was 0.31 cms/km2, while the SD of the observed 10‐year return period was 1.28 cms/km2 (Figure 2.2). The Santa Ynez region had the lowest accuracy by region for the 10‐year return period, bias = −2.362 cms/km2, and Franklin Canyon, within the Santa Ynez region, had the lowest accuracy for all catchments, where 10‐year return period peak flow was under‐predicted by a factor of 4.75. Overall, pre‐fire peak streamflow prediction performance was low for the 2‐ and 10‐year recurrence interval events (R2 = 0.24 and RMSE = 0.38 cms/km2; R2 = 0.34 and RMSE = 1.43 cms/km2, respectively).

17

Figure 2.2. Rowe, Countryman, and Storey (RCS) unburned peak streamflow predictions compared to observed peak streamflow for 2‐year (a) and 10‐year (b) return periods.

18

Table 2.2. Rowe, Countryman, and Storey (RCS) Bias for Peak Streamflow Predictions by Region for the 2‐ and 10‐year Events

RCS Bias (cms/km2) Region Q2 Q10 San Bernardino (n = 3) 0.551 0.371 Santa Ynez (n = 7) −0.238 −2.362 San Jacinto (n = 2) 0.134 −0.237 San Gabriel (n = 6) 0.346 0.230 Santa Ana (n = 2) 0.237 −0.188 San Diego (n = 4) 0.272 −0.276 Southern California (n = 24) 0.162 −0.791 Note: Return periods were calculated from at least 21 years of peak flow data and n represents the number of catchments.

2.4.1.2 RCS PERFORMANCE AFTER FIRE RCS provides probabilistic flood frequency predictions of peak streamflow. A large enough statistical sample size was not available after fire to produce 2- and 10-year estimates (commonly used in post-fire risk assessment) from streamflow observations. Thus, post-fire predictions of the flood frequency model for 2-year and 10-year events were compared directly to observed peak streamflow from the first year. For 10‐year return period post‐fire predictions, 22 catchments experienced flooding during the first 5 years after fire and were under‐predicted by RCS. When comparing observed post‐fire flow directly to the probabilistic predictions, R2 and RMSE for RCS 2‐year return periods yielded 0.26 and 16.01 cms/km2, and R2 and RMSE for RCS 10‐year return periods yielded 0.25 and 15.52 cms/km2, respectively. Predictions generally had largely negative bias with 2‐year return periods yielding bias of −8.68 cms/km2 and 10‐year return periods yielding bias of −7.78 cms/km2. Regionally, it was noted that 13 catchments observed post‐fire peak streamflow larger than the RCS 100‐year prediction. Catchments in the San Diego and San Jacinto regions on average had 149% lower post‐fire peak streamflow compared to the other regions. Predictions for catchments in the Santa Ynez, San Gabriel, San Bernardino, and Santa Ana mountains were inaccurate, with errors ranging up to 1720% during the floods at Dickey Canyon that followed the 2018 Holy Fire (Figure 2.3). These flows were rapid, where field observed velocity was measured to be approximately 12 m/s.

19

Figure 2.3. Sediment‐laden flows and high water mark observations in a concrete lined channel at Toft Dr. near the outlet of Dickey Canyon during 29 November (a and b) and 6 December 2018 storms (c). The December 6 peak flow filled the channel (c) and resulted in damage at outlet, where flow exceeded the culvert capacity (d). Photos a and c were extracted from videos provided by Riverside County Flood Control and Water Conservation District personnel and by resident S. Engelhardt, respectively.

2.4.2 Random Forest Post-fire Model Performance The RF‐45 TreeBagger function identified the following parameters as the most important for predicting post‐fire peak streamflow for the study area (in order of importance): days elapsed from end of fire (containment date) to storm, total area of catchment burned, catchment drainage area, catchment perimeter, and peak 1‐hour rainfall intensity (Figure A.1-3). A second RF model was developed using only the five most important parameters. This resulted in a model that could be adequately represented with five parameters, therefore eliminating the issue of model overfitting (Figure 2.4). The RF‐45 (R2 = 0.79 and RMSE = 7.34 cms/km2) and RF‐5 models (R2 = 0.46 and RMSE = 7.89 cms/km2) exhibited modest performance metrics and decreased uncertainty in post-fire predictions relative to RCS method.

20

Figure 2.4. Predictions for the RF‐5 random forest model by region. The sample size (n) is noted for each region.

Accuracy decreased substantially when peak streamflows were above 10 cms/km2, most likely due to uncertainty caused by debris flows and hyperconcentrated flows. The Santa Ynez region had the highest R2 compared to the other regions (Bias = −1.77 cms/km2; R2 = 0.93; RMSE = 7.99 cms/km2). This region also had the largest sample size in the study (n = 27) and the largest amount of available precipitation data to train the RF model. However, the extreme high and low flows within this sample space may contribute to the larger R2. Additionally, the performance of RCS and RF-5 were compared for the largest post‐ fire floods for each of the 33 catchments between 1920–2018 (Figure 2.5). The RCS 2‐year and RCS 10‐year RMSE values were 16.01 cms/km2 and 15.52 cms/km2, respectively, while the RF RMSE was found to be lower at 10.41 cms/km2. RCS is a flood frequency estimation model while RF-5 is a flood forecasting model and cannot be directly compared. However, the ultimate goal was to reduce uncertainty in forecasting extreme floods following fires over the past century in southern California.

21

Figure 2.5. Observed peak streamflow versus predicted streamflow response. The black line represents a perfect prediction. Squares represent RCS (2‐ and 10‐year events) and triangles represent random forest with the five most important parameters (RF‐5). RF‐ 5 is based on the entire training dataset and has no affiliated event magnitude. The extreme conditions or the highest floods for each catchment are shown (n = 33).

2.4.3 Towards an Analytical Solution for Assessing Post-fire Peak Flows In June 2020, 65 out of 116 researchers and professionals in fire science prioritized “assessing and mitigating flash flood and debris flows hazards” over “burn area recovery” (n = 31), “water quality impacts” (n = 17), and “invasive species and/or changes in fire regime” (n = 3) (Coalitions, & Collaboratives, Inc., 2020). This motivated the development of a simple analytical solution for assessing immediate responses for post‐fire peak flows in small catchments. The most important parameters identified by RF‐45 were time after fire (used to distinguish events within the first year), rainfall intensity, and burned area; they were used to create a simple regression (Figure 2.6). Thirty‐one rainfall‐runoff events during the first year after fire from the study area were fitted to a three‐dimensional polynomial function (R2 = 0.82) with the equation:

= 8.316 + 0.4033( ) + 0.9041( 60) 0.04079( )( 60) + 0.0127( 60) (2.7) 2 𝑄𝑄𝑄𝑄𝑄𝑄 − 𝐴𝐴 𝑖𝑖 − 𝐴𝐴 𝑖𝑖 𝑖𝑖 22

where Qpk is peak streamflow (cms/km2), A is burned area (km2), and i60 is peak hourly rainfall intensity (mm/h). Catchment area and catchment perimeter were not used for the regression. On average, burn proportions consisted of 56% moderate to high soil burn severity and 44% unburned to low soil burn severity.

Figure 2.6. Modeled regression for small, steep catchments during the first year after fire in southern California, fitted to historic high flows within the first year after wildfire collected from 1920–2019. This model is limited to catchments ranging between 1 to 20 km2 where more than half of the catchment has been burned at moderate to high soil burn severity.

2.5 DISCUSSION

2.5.1 The Future Role of Rowe, Countryman, and Storey Model RCS is commonly used in southern California after wildfire to model flood risk for affected catchments. In this study, RCS had large inaccuracy for small catchments (1 to 42 km2). This could be due to a multitude of factors including catchment morphology, exclusion of soil burn severity, greater development in the wildland‐urban interface, and increasing frequency in extreme weather events due to climate change. RCS had the largest inaccuracies for catchments in the Santa Ynez Mountains, which is of concern due to the tendency for hazardous (Sections 2.2.1–2.2.3) events in this region, such as the 9 January 2018 Montecito debris flows (Kean et al., 2019)⁠. RCS regressions were developed for flows

23 with low sediment concentrations (non‐bulked flows), therefore, RCS predictions are expected to be limited and unable to predict peak flows associated with debris flows. In comparison to clear water flows, which generally have suspended‐sediment concentrations less than 5% to 10% sediment by volume, hyperconcentrated flows can have suspended‐ sediment concentrations from 5% to 60% sediment by volume, and debris flows can have suspended‐sediment concentrations greater than 60% sediment by volume (Pierson, 2005)⁠. Accurate predictions of post‐fire streamflow along the continuum from flooding to debris flows are needed due to their frequent occurrence following wildfire in southern California (Cannon & DeGraff, 2009)⁠. Significant advances in post‐fire hydrology since the development of Rowe et al. (1949) should be incorporated to improve the accuracy of predictions (Kinoshita et al., 2014)⁠. For example, sediment bulking is implicit in RCS, yet independent variables that are strongly linked to sediment production or sediment yield are not used. Additionally, this observation‐ based method predates development of the soil burn severity metric. Soil burn severity characterizes the fire‐induced changes in soil and ground cover properties that can affect infiltration, runoff, and erosion potential (Parsons et al., 2010)⁠ and is incorporated in post‐fire hydrogeomorphic modeling as an independent variable to predict debris flows and debris yield (Gartner et al., 2014; Staley et al., 2017)⁠. Pre‐fire RCS SDs between catchments (0.17–0.31 cms/km2) were smaller than the observed estimates (0.35–1.28 cms/km2), indicating that RCS predicted less variation than was actually present. This disparity between the SDs highlights the insensitivity of RCS predictions to different conditions between catchments, and ultimately, contributes to additional error in post‐fire predictions. This is most prominent in the Santa Ynez region, likely due to increased development around the wildland-urban interface, agricultural changes, and channelization of flow after the RCS period of record (1860 to 1949). The post‐ fire RCS method utilizes “fire factors” that modify pre‐fire flow. This discrepancy is illustrated in the post‐fire predictions of the Santa Ynez region. Many of the spatially and temporally sensitive parameters such as rainfall intensity and soil burn severity are not adequately represented in the RCS 1949 method. The limitations of RCS are attributed to its conceptualization prior to the development of the Geographic Information System (GIS). While, RCS rating curves were developed using

24

basin size, shape, steepness, stream channel characteristics, infiltration and water storage capacities of various soil‐geologic formations, precipitation, and vegetation characteristics (Rowe et al., 1949), these datasets were based on hand‐drawn maps, field observations, and other historical records with lower spatial and temporal resolution. Further, RCS used a flood frequency technique that assumes a static landscape and climate; however, these assumptions are not valid under dynamic climate (Milly et al., 2008)⁠. For example, studies such as Musselman et al. (2017) and Prein et al. (2017) note that climate is not stationary, where earlier onset of snowmelt, increased air and water temperature, and increased frequency of extreme weather events can directly impact surface runoff. Therefore, a statistical, flood frequency approach to post‐fire modeling, or a model that uses return periods and probabilities of occurrence such as RCS, will decrease in reliability over time, eventually rendering it irrelevant for future post‐fire risk assessment in southern California.

2.5.2 Role of Machine Learning as a Tool for Post-fire Risk Assessment The flood forecasting approach of this method provided an opportunity to better capture the parameters influencing extreme post-fire flood events. The parameters presented in Wilder et al. (2021) are essential in characterizing post‐fire response in small southern California catchments, however, this list is not exhaustive and potentially misses other important parameters. For example, 15‐minute and 30‐minute peak rainfall intensities, vegetation density, vegetation type, soil moisture, fault boundaries, sediment availability in the channels, relative humidity, and drought and climate variations, were not investigated (DiBiase & Lamb, 2020)⁠. Also, RF‐45 did not determine soil burn severity to be an important parameter, which is contrary to several studies that note the importance of soil burn severity in triggering large post‐fire peak flows (Gartner et al., 2014; Huffman et al., 2001; Lewis et al., 2006; Moody, 2012; Shakesby et al., 2016; Staley et al., 2017). This discrepancy is likely due to the lack of soil burn severity data for this modeling exercise. Soil burn severity was not available for 35 of the 74 rainfall‐runoff events, limiting the model's ability to identify the importance of severity. This reinforces the need and importance for consistent collection of field data, particularly high temporal resolution rainfall intensity, soil burn severity

25

mapping, and streamflow data (Staley et al., 2013)⁠ to improve the contributions of ML in post‐fire hydrologic predictions. There is significant research that demonstrates the importance of high temporal resolution rainfall data (e.g., 5‐, 15‐, and 30‐minute rainfall intensities) initiating debris flows and associated high peak flows (Staley et al., 2013)⁠. This study focused on data that were open access (60‐minute rainfall intensity). Further, only one rain gauge was used as a representative rainfall measurement for each catchment, which may contribute to uncertainty in the RF‐5, RF‐45, and polynomial regression predictions. There are also inherent errors in the measured post‐fire peak streamflow, which may have been heavily laden with soil, ash, burned vegetation, large boulders, and other debris (i.e., significantly bulked flow). This study suggests that with sufficient high‐quality data, machine learning can be a valuable procedure for developing predictive tools for post‐fire risk assessment. For example, the Santa Ynez region had the largest sample size and the most available rainfall data, resulting in the highest R2 by region for RF‐5 (n = 27; Bias = −1.77 cms/km2; R2 = 0.93; RMSE = 7.99 cms/km2). The larger correlation may be attributed to the extreme high and low flows. Further, machine learning requires data collection, calibration, and parameterization that should be carried out cautiously. Excluding or missing parameters that have significant importance to model accuracy can lead to highly inaccurate predictions due to insufficient processes being defined by the data.

2.5.3 Factors Influencing Post-fire Peak Streamflow Based upon the ML approach, peak rainfall intensity, catchment size, and time after fire containment have a significant role in determining flow rate per unit area after wildfire (Figure 2.7). It was observed that peak hourly rainfall intensities over 10 mm/h led to larger magnitude floods (Figure 2.7 (a)). These findings may be due to physical catchment processes, whereby larger peak rainfall intensities increase rill erosion and channel incision (Cannon & DeGraff, 2009). Catchments with smaller areas (1–10 km2) are more likely to have larger magnitude runoff per unit area (p < 0.05) (Figure 2.7 (b)), which is similar to Neary et al. (2005). In smaller catchments with predominantly chaparral vegetation type, runoff responses are erratic and have potential to transport large amounts of sediment per unit area after fire (Keller et al., 1997). Neary et al. (2005) also reported much larger magnitudes

26

for small catchments in Western United States (<1 km2), where post-fire peak flows averaged 193 cms/km2, further highlighting the increased potential for higher magnitude peak flows. Finally, storms that occurred closer in time to the fire containment have a higher likelihood of larger magnitude events (p < 0.05) (Figure 2.7 (c)). The passing of time allows hydrophobic soils to normalize and vegetation to recover, reducing rainfall impact on bare soil (Neary et al., 2005). As seen in studies such as Saxe et al. (2018) and Wilder et al. (2021), post-fire flow can be characterized using machine learning at varying scales (United States and southern California) using publicly available data. With increased collection of data related to important factors found in this study (i.e. rainfall intensity and streamflow), there is an opportunity to scale machine learning studies up to larger scales (continental) to develop accessible models that can be used for rapid risk assessment (Saxe et al., 2018; Schmidt et al., 2020).

27

Figure 2.7. Observed peak streamflow per unit area with respect to peak hourly rainfall intensities (a), catchment size (b), and days after fire containment (c).

2.5.4 Analytical Solution Rapid assessment and accurate modeling of post-fire peak streamflow are essential for effective risk management implementation. The use of RCS has persisted in emergency assessments despite its shortcomings due to general acceptance, ease of use, and lack of a better simplified model. Many current models require extensive data acquisition, model setup and testing, and field calibration; these models may generally work well but are not feasible under time constraints required by post-fire assessment teams and emergency management

28

agencies to make rapid decisions for quantifying risk and developing response measures. Thus, it is advocated that more work is needed to provide more data for validation and improve existing methods and tools that better fit the current needs of post-fire assessment teams and local agencies to quantify risk to downstream communities, infrastructure, and ecosystems. Based on the regional analysis using random forest (Figure 2.4; Figure A.1), a simple regression is proposed that can be used for peak streamflow estimates in postfire risk assessment for the first year after fire for catchments between 1 to 20 km2 in southern California [Figure 2.6; Equation (2.7)].

2.6 CONCLUSION Following wildfires in southern California, the probability of flooding and debris flows increase, thus prompting the need for accurate prediction to protect downstream communities, infrastructure, and ecosystems. Rowe et al. (1949) produced large inaccuracies for pre- and post-fire peak streamflow predictions in small sized catchments (1 to 42 km2) in southern California. It is suggested that RCS should be used with extreme caution. The development of two models using machine learning is demonstrated. It is shown that RF-45 can identify relationships and characterize post-fire peak streamflow of small catchments. Through RF-45, it was discovered that days elapsed from end of fire to storm, total area of catchment burned, drainage area, catchment perimeter, and peak 1-hour rainfall intensity are important parameters that contribute to greater model accuracy. The RF-5 model built with these five parameters (Bias = −2.81 cms/km2; R2 = 0.46; RMSE = 7.89 cms/km2) had higher reliability than the RCS Q2 and Q10 regressions (Bias = −8.68 – −7.78 cms/km2; R2 = 0.26– 0.25; RMSE = 15.52–16.01 cms/km2). A simple analytical solution is introduced to provide post-fire assessment teams an interim tool to rapidly predict peak floods for small catchments in southern California (Bias = −2.26 cms/km2; R2 = 0.82; RMSE =6.59 cms/km2). It is concluded that a significant increase in data collection of high temporal and spatial resolution rainfall intensity, streamflow, and sediment loading in channels will help to guide future model development to quantify post-fire flood risk.

29

CHAPTER 3

CHARACTERIZING FIRE SEVERITY AND ECOHYDROLOGICAL RECOVERY FOR THE 2018 HOLY FIRE IN SOUTHERN CALIFORNIA

3.1 INTRODUCTION The 2020 fire season burned 16,907 km2 of land across the state of California, including the 2020 August Complex Fire (4,179 km2), the largest fire documented in California during the period 1932 to 2020 (CALFIRE, 2020). Severe fires such as these affect physical and biogeochemical processes at the catchment scale and influence changes in hydrologic processes for several post-fire seasons (Mayor et al., 2007)⁠ including a notable decrease in root water uptake (Obrist et al., 2004)⁠ and annual above-ground biomass accumulation (Uyeda et al., 2015), and an increase in post-wildfire streamflow from ecological disturbance (Wine et al., 2018; Wohlgemuth, 2016) and sediment transport from channel loaded dry ravel (DiBiase & Lamb, 2020)⁠. The recovery and normalization of these hydrologic processes in Mediterranean systems are relatively resilient, recovering quite rapidly within 3 to 5 years (Wittenberg et al., 2007); however, they tend to vary spatially and temporally across different landscapes with respect to soil burn severity, slope, aspect, pre- fire vegetation species, and revegetation and risk mitigation strategies (Keeley & Keeley, 1981; Kinoshita & Hogue, 2011; Vo & Kinoshita, 2020). Additionally, there may be a notable decrease in evapotranspiration (ET) rates for burned areas in semi-arid regions (Poon & Kinoshita, 2018; Prater & DeLucia, 2006)⁠, which has implications for water yield (Kinoshita & Hogue, 2015; Soulis et al., 2021). To date, research into post-fire ET measurements in small catchments, and the resulting ecohydrological recovery, have been limited to field investigations and small-scale studies due to the unavailability of satellite-based products with appropriate resolution at the small

30 catchment scale. For example, Prater and DeLucia (2006) coupled energy flux station data (Bowen ratio-energy balance method) with normalized difference in vegetation index. Their study found disturbances such as wildfire were converting native sagebrush to invasive grasses in a semi-arid environment (Prater & DeLucia, 2006)⁠. This conversion led to relatively lower soil moisture and lower ET during the dry months. In a larger scale study, Poon and Kinoshita (2018) used Operational Simplified Surface Energy Balance Model (SSEBop) to obtain monthly ET data at 1-km spatial resolution to assess vegetation recovery patterns for a semi-arid region affected by fire. The study found there was a statistically significant decrease in ET following the 2011 Las Conchas Fire in New Mexico (USA) and a potential shift in vegetation type from conifer to grasslands during the post-fire period 2011 to 2014. This research builds upon previous studies by integrating ET at improved resolution to assess disturbance from fire in small catchments, a scale that is not well documented to date, yet is recognized as vital in post-fire land management. In the case of the Holy Fire, this type of information is crucial in deciding when to safely open up trails and roads to the public after fire (USFS, 2020). To help resolve this challenge, this research incorporated the ECOsystem Spaceborne Thermal Radiometer Experiment on Space Station (ECOSTRESS) Level 3 daily Priestley-Taylor (PT-JPL) ET product, which has a high spatial resolution (70- m) and high temporal resolution (1-5 days) to assess ecohydrological effects associated with fire (Stavros et al., 2017). To link vegetation and hydrologic processes, additional data were collected including above-ground biomass indices, vegetation species map, SSEBop monthly ET, soil burn severity map, and local rainfall-runoff data to quantify changes in streamflow using hydrologic signatures at the annual timescale (McMillan, 2020)⁠. Accurate measurements of ET are crucial in our understanding of the long-term impacts to downstream water resources infrastructure, ecosystems, and communities after fire. Thus, the overarching goal of this work is to improve vegetation assessments as it relates to fire by demonstrating the application of ECOSTRESS PT-JPL ET measurements. Using the case of the 2018 Holy Fire this work 1) characterized pre-fire climate and vegetation conditions and 2) characterized recovery in ecohydrological processes (streamflow and ET) at the small catchment scale.

31

3.2 MATERIALS AND METHODS

3.2.1 Study Area and Hydrologic Data The study area is located within Orange County and Riverside County, California (USA) and is within the Cleveland National Forest and Santa Ana Mountains (Figure 3.1). This region is characterized as a Mediterranean climate, Köppen Csa, which consists of hot, dry summers and cool, mildly wet winters (Peel et al., 2007)⁠.

Figure 3.1. Location of Santiago (control) and Coldwater (burned) catchments and the 2018 Holy Fire soil burn severity; streamflow and precipitation gauges are also denoted. The inset shows the general location of the study area in California, United States.

32

On August 6, 2018, a human-ignited fire burned approximately 94 km2 (Figure 3.1). The soil burn severity classifications for this fire were comprised of approximately 14% high, 71% moderate, 8% low, and 7% low to unburned (WERT, 2018a), where typically higher soil burn severities have the greatest effects on ecohydrological processes and recovery time (Lentile et al., 2007; Parsons et al., 2010)⁠. To analyze the effects of this fire, two catchments, Santiago (control) and Coldwater (burned), were selected for a paired analysis. Daily streamflow for Santiago and Coldwater were collected from United States Geological Survey (USGS) gage 11075800 and 11071760, respectively. Daily streamflow were downloaded for Water Year (WY) 2014 to 2020 (October 1, 2013 to September 30, 2020). It was noted that Coldwater streamgage was installed after the fire on December 4, 2018 and does not have streamflow prior to this date. Streamflow results were transformed to provide better visualization of the baseflows of the ephemeral catchments using a Box-Cox transformation (Box & Cox, 1964)⁠:

Box-Cox Q = 𝜆𝜆 (3.1) 𝑄𝑄 −1 𝜆𝜆 where λ = 0.3 has been found to work well for semi-arid regions (Hogue et al., 2000)⁠. Daily precipitation totals were acquired from the Upper Silverado gauge maintained by Orange County Public Works and data were downloaded for WY 1991 to 2020. The neighboring catchments (sharing Santiago Peak as their common border) were similar in size, elevation, topography, and vegetation type, consisting of mostly chaparral, coastal scrub, and montane species (Table 3.1). Catchment parameters were derived from Chapter 2 (Wilder et al., 2021) and include variables that can influence hydrologic properties such as hydrologic soil group (classification of infiltration rates), slope aspect, and vegetation species.

33

Table 3.1. Catchment Parameters and Vegetation Types for Paired Catchments Santiago and Coldwater Catchment Parameters Santiago Coldwater

Drainage area (km2) 32.4 10.9 Estimated slope along longest flow path 14% 19% Lowest elevation (msl) 385 389 Highest elevation (msl) 1735 1735 Average annual precipitation (mm) 650 577 Area with north aspect 22% 40% Area with east aspect 13% 32% Area with south aspect 29% 14% Area with west aspect 36% 14% Area covered by forest 28% 37% Area developed (urban) 2.1% 1.8% Hydrologic soil group A 1% 3% Hydrologic soil group B 3% 1% Hydrologic soil group C 9% 0% Hydrologic soil group D 87% 96% Average Erodibility Factor (k) 0.34 0.31 Mean slope computed from 10-meter DEM 55% 63% Vegetation Type Santiago Coldwater Mixed Chapparal 65.4% 53.5% Coastal Scrub 14.0% 8.2% Chamise-Redshank Chapparal 7.9% 1.5% Montane Hardwood 7.4% 17.9% Montane Hardwood-Conifer 0.3% 13.3% Coastal Oak Woodland 3.1% 4.4% Annual Grassland, Barren, Lacustrine, Montane < 1.4% < 1.2% Chapparal, Sierran Mixed Conifer, Urban, and Valley Foothill Riparian

3.2.2 Remote Sensing and Spatial Products The following satellite-based products were used in the investigation: California Department of Forestry and Fire Protection (CAL FIRE) Fveg, Environmental Protection Agency (EPA) NHDPlusV2 (National Hydrography Dataset) flow accumulation raster, U.S.

34

Forest Service soil burn severity map, Landsat 8 Enhanced Vegetation Index (EVI), SSEBop monthly ET, and ECOSTRESS Level 3 PT-JPL daily ET. CAL FIRE Fveg was used to characterize the vegetation species of the study area. The Fveg dataset was compiled from 1990 to 2014, and is the most detailed vegetation classification for California, containing all the different species found in the region at 30-m spatial resolution (CALFIRE-FRAP, 2015). Secondly, a flow accumulation raster was acquired from the EPA to differentiate the riparian and hillslope areas of the 2018 Holy Fire at 30-m spatial resolution (McKay et al., 2015). After clipping the flow accumulation raster to the fire area, a minimum threshold of 1,500 upstream pixels was used to establish a stream network. Following reclassification and vectorization of the stream network, a 70-meter buffer was selected to be placed around the stream network to approximate the riparian areas for this study area. This buffer size was selected based on the 70-m footprint of the ECOSTRESS pixels and vegetative buffers of approximately 100-m, which has been noted for optimal wildlife habitat (Wenger & Fowler, 2000). A soil burn severity map was acquired from the U.S. Forest Service Burned Area Emergency Response (BAER) Imagery Support. Soil burn severity maps are delineated into 4 categories (High, Moderate, Low, and Low to Unburned) based on differences in near and mid infrared reflectance values (Parsons et al., 2010). In general, these maps are helpful for risk assessment teams in determining areas that may be prone to increased erosion and flooding (Fernández & Vega, 2016; Vieira et al., 2015; WERT, 2018a, 2018b). Additionally, EVI is an optimized vegetation index that can monitor changes in above-ground biomass with a high sensitivity to variations in canopy structure, phenomena critical for post-fire studies (Lentile et al., 2007). Landsat 8 EVI data were collected for WY 2014-2020 at 30-m spatial resolution and 8-day temporal resolution. Next, the SSEBop product is based on the conversion of reference ET to actual ET through the formulation of ecophysiological parameters (land surface temperature, vegetation indices, hot/cold pixels) into a simplified radiation driven energy model (Savoca et al., 2013)⁠. In a study by Chen et al. (2016), they compared SSEBop ET predictions to 42 AmeriFlux sites across the United States during the period 2001 to 2007, yielding high performance metrics of R2 = 0.86 and RMSE = 15 mm/month (sample size = 1,680). For our study, SSEBop monthly ET product at 1-km spatial resolution were collected from WY 2001 to 2020. Finally, higher resolution ECOSTRESS ET data derived from the PT-JPL algorithm were collected from July 2018 to

35

September 2020 (mission launched on June 29, 2018) using the Land Processes Distributed Active Archive Center (LP DAAC) AppEEARS. The PT-JPL algorithm is derived using a series of ecophysiological scalar functions based on atmospheric vapor pressure deficit, relative humidity, and vegetation indices to translate potential ET into actual ET (Fisher et al., 2020)⁠. Preliminary validation of ECOSTRESS PT-JPL ET yielded a cumulative R2 = 0.88 and RMSE = 41.3 W/m2 (sample size = 502) when compared to 82 eddy covariance sites from across the world (Fisher et al., 2020), and in a more local study, yielded a cumulative R2 = 0.89 and RMSE = 0.10 mm/hr (sample size = 143) when compared to five California Irrigation Management Information System (CIMIS) Stations in Riverside County, California (Kohli et al., 2020). ET data was converted from W/m2 to mm/day using the latent heat of vaporization and density of water at 20°C. ET data were also checked for quality using the accompanying ECOSTRESS L3/L4 Ancillary Data Quality Assurance (QA) Flags. No data were withheld for QA flags, as all data had acceptable amounts of pixels for spatial analysis. However, the images collected on August 26, 2018, August 23, 2019, September 23, 2019, April 22, 2020, May 20, 2020, June 23, 2020, and July 10, 2020 measured ET that were greater than 3 standard deviations. These images were removed from the analysis.

3.2.3 Climatology A standardized precipitation index (SPI) analysis was used to note any climate anomalies or patterns in the study area (Bonaccorso et al., 2003; World Meteorological Organization, 2012). Daily rainfall data from Upper Silverado precipitation gauge, spanning from WY 1991 to 2020, were used to calculate annual SPI values using the following equation:

SPI = (3.2) 𝑃𝑃𝑦𝑦−𝑃𝑃 𝜎𝜎 where y denotes the year of interest, Py is the annual precipitation for a year, and and are

the long-term average and standard deviation, respectively. SPI values greater than𝑃𝑃 2.0 were𝜎𝜎 classified as extremely wet, values between 1.5 to 1.99 were classified as very wet, values between 1.0 to 1.49 were classified as moderately wet, values between -0.99 to 0.99 were classified as near normal, values between -1.0 to -1.49 were classified as moderately dry,

36

values between -1.5 to -1.99 were classified as severely dry, and values less than -2.0 were classified as extremely dry (World Meteorological Organization, 2012)⁠.

3.2.4 Ecohydrological Analysis

3.2.4.1 CORRELATION OF SSEBOP AND ECOSTRESS PT-JPL Google Earth Engine (GEE) is a cloud computing platform and was used to analyze spatial trends and create time series for the satellite-based products (Gorelick et al., 2017)⁠. SSEBop and ECOSTRESS PT-JPL ET were directly compared using two post-fire images, August 17, 2019 and October 3, 2020, and one pre-fire image, August 4, 2018. SSEBop monthly ET images were scaled to daily time scale by dividing the total actual ET (mm) by the number of days for the selected month. A regression analysis was performed in GEE to assess coefficient of determination and slope of the best-fit between ECOSTRESS PT-JPL and SSEBop daily ET. ECOSTRESS PT-JPL was reduced to 1-km resolution by taking the mean of each pixel within each SSEBop footprint to allow a direct comparison. Sample pixels for the regression analysis were selected using the GEE “Image.sampleRegions,” which resulted in an average of 103-pixel samples from each SSEBop image for the Holy Fire burn scar.

3.2.4.2 PRE-FIRE ABOVE-GROUND BIOMASS AND POST-FIRE VEGETATION RECOVERY The relationship between EVI, vegetation type, and soil burn severity were examined. To estimate EVI for each catchment, the median value of each scene within each catchment were calculated. Data were processed using a Savitzky-Golay function (degree of 1) in MATLAB to smooth noise and a moving average function to fill in dates with missing data. EVI was summed for each WY (ΣEVI) to represent the annual accumulation of above- ground biomass, a metric that has been used to assess post-fire recovery of catchments in semi-arid regions (Kinoshita & Hogue, 2011)⁠. The ΣEVI data was then differentiated with respect to vegetation type and soil burn severity during the pre-fire period WY 2014 to 2017 and tested for statistical significance. Lastly, ECOSTRESS PT-JPL ET images were

37

compiled for WY 2019 and 2020 to analyze spatial and temporal trends with respect to different vegetation types, slope aspect, soil burn severity, and riparian/hillslope areas.

3.2.4.3 HYDROLOGIC SIGNATURES Hydrologic signatures at the annual timescale were calculated using the runoff-ratio (RO) and Richards-Baker (R-B) flashiness index to quantify streamflow responses to changes in landcover and ET after fire (Baker et al., 2004; McMillan, 2020)⁠. The runoff- ratios were calculated for each WY using the following equation:

RO = (3.3) 𝑄𝑄𝑦𝑦 𝑃𝑃𝑦𝑦 where y denotes the year of interest, Py is the annual precipitation for a year, and Qy is the annual streamflow for the year. In general, RO at the annual timescale is useful in distinguishing losses to deep groundwater and surface ET. In this specific case, it is assumed both catchments were similar geologically and therefore differences in RO could be directly comparable. Exact values of RO are site dependent, however in general, higher RO values indicate relatively higher direct surface water runoff while lower RO values indicate larger losses to either deep groundwater or surface ET. The R-B indices were calculated for each WY using the following equation:

R-B index = 𝑛𝑛 (3.4) ∑𝑖𝑖=1 𝑞𝑞𝑖𝑖−𝑞𝑞𝑖𝑖−1 𝑛𝑛 ∑𝑖𝑖=1 𝑞𝑞𝑖𝑖 where q is the measured daily streamflow for day, i. In general, R-B index at the annual timescale is useful in distinguishing variability between daily flows due to landcover change and disturbance, where higher R-B values indicate greater amounts of precipitation translated to direct surface water runoff and lower R-B values indicate greater amounts of interception and infiltration. These hydrologic signatures, as well as SSEBop ET and ΣEVI for each WY, were compared between the burned and control catchments. Also, differences in ECOSTRESS PT- JPL ET at the bi-monthly timescale were computed by using a 60-day rolling average fitting algorithm in Python (Reback et al., 2020) to assess trends between the control and burned catchments. Finally, a simple water balance calculation (Figure A.4) was carried out using

38

the differenced daily ET generated from the rolling average fitting algorithm to estimate difference in storage volume for WY 2020:

ΔS = + + (3.5)

Where ΔS was annual difference in storage𝛥𝛥𝛥𝛥 between𝛥𝛥𝛥𝛥𝛥𝛥 Santiago𝛥𝛥𝛥𝛥 and Coldwater, ΔR was the difference in annual runoff, ΔET was annual difference in ET, and ΔG was annual difference in groundwater contributions. It is noted that ΔG was neglected for this water balance calculation (assumed to be approximately zero for small catchments in proximity with high relief).

3.3 RESULTS

3.3.1 Climate and ET Processes For this site SPI classified wet years appeared intermittently, typically following 4-6 moderately dry to near normal years. Further, there were no classified wet years leading up to the fire from the period 2012 to 2018, where all SPI values were less than 1.0. The next classified wet year occurred later after the fire in WY 2019 (moderately wet classification). Timeseries analysis of SSEBop ET and SPI demonstrated the relation between drought variations and vegetation health for the area affected by the 2018 Holy Fire (Figure 3.2). The average annual pre-fire ET for the area burned by the fire was 722-mm for WY 2001 to 2017. The three years prior to the fire had the lowest annual ET for the area burned during this period, ranging from 610 to 625-mm. Additionally, the six years prior to the fire were all were below the pre-fire average. Further, a modest correlation existed between pre- fire annual ET and SPI (R2 = 0.41; n=17), where generally the extreme wet and dry years had relatively high ET and low ET, respectively. Immediately following the fire in WY 2019, a moderately wet year occurred, with an SPI value of 1.17 (Figure 3.2). There were several intense storms during the first post-fire season with 60-minute rainfall intensities ranging from 14-21 mm/hr, 30-minute rainfall intensities ranging from 18-28 mm/hr, and 15-minute rainfall intensities ranging from 23-40 mm/hr (Riverside County Flood Control and Water Conservation District). This moderately wet year translated to a large resurgence in ET (44% increase) the following year during WY 2020.

39

Figure 3.2. Standardized precipitation index (SPI) for Upper Silverado station (WY 1991-2020) are shown on the primary axis. SSEBop annual evapotranspiration (ET) measurements (WY 2001-2020) for the 2018 Holy Fire burn area are shown on the secondary axis.

Further, SSEBop monthly ET timeseries was used to compare the six drought years (WY 2012-2018) leading up to the Holy Fire with respect to the rest of the available data set (WY 2001 to 2011) using the average of the 103 pixels (Figure 3.3). The two groups were statistically different (p < 0.05) for all months (except for November), where the six-year period before the fire had significantly lower ET. Monthly ET typically peaked in this region during the dry period between June to July, with values ranging from 64 to 151 mm/month for the pre-fire period. Most notably, one month prior to the fire in July of WY 2018, ET was 64 mm/month, the lowest value for the pre-fire record (2001-2018) for the month of July.

40

Figure 3.3. Average SSEBop monthly evapotranspiration (ET) at 1-km spatial resolution for the Holy Fire (August 2018) burn area before the fire (WY 2001 to 2018) and for the available record of ET (WY 2001-2011). Average number of pixels in the burned area equal to 103. Note that WY 2018* does not include post-fire months, August and September.

3.3.2 Pre-Fire Above-Ground Biomass Pre-fire (WY 2014-2017), the two catchments had comparable (but statistically different) annual ΣEVI, where the difference between average ΣEVI was 0.23%. The difference in pre-fire EVI is attributed to the control catchment having a larger proportion of montane hardwood and montane hardwood conifers (Table 3.1). Santiago was comprised of 8% montane hardwood and montane hardwood conifers, and Coldwater was comprised of 31%, which was the largest difference between the two catchments’ initial conditions. To further characterize the pre-fire fuel conditions, pre-fire ΣEVI is differentiated by different soil burn severity categories. Areas with high concentrations of montane hardwood and montane hardwood conifers (areas with typically higher pre-fire EVI) were found within the high soil burn severity classification and were statistically different from moderate, low, and low to unburned (Unb) areas (p < 0.05) (Figure 3.4). In general, there were no strong correlations with mixed chaparral, coastal scrub, and other vegetation types. Areas with high proportions of montane hardwood and montane hardwood conifers (MH+MHC), greater than 30% by area, appeared to be moderately correlated with high soil burn severity for this fire.

41

Figure 3.4. Pre-fire average ΣEVI (WY 2014-2017) are shown on the primary axis with respect to their soil burn severity classification, where n represents the average number of pixels for each sample in Coldwater. The proportion of area that had montane hardwood and montane hardwood conifer (MH+MHC) are shown as brown triangles.

3.3.3 Correlation of SSEBop and ECOSTRESS PT- JPL with Respect to the Holy Fire SSEBop and ECOSTRESS PT-JPL ET pixels were compared and resulted in comparatively higher uncertainty of the two ET products after fire (Figure 3.5; Figure A.5), where pre-fire R2 = 0.47 and post-fire R2 ranged from 0.00 to 0.04. Further, the pre-fire slope of the linear regression between the two datasets was equal to 0.36 (below the 1-to-1 line) and decreased during post-fire to values ranging from -0.10 to 0.12.

42

Figure 3.5. ECOSTRESS PT-JPL and SSEBop scatter plot (n = 103 pixels) for pre-fire on August 2, 2018 (a), 1-year post-fire on August 17, 2019 (b), and 2-years post-fire on October 3, 2020 (c). Shading around the regression line represents the standard error of the estimate.

3.3.4 Post-fire Ecohydrology and Hydrologic Signatures In the pre-fire period, control catchment RO values ranged from 0.01 to 0.16 and R-B index ranged from 0.52 to 0.81. In the post-fire period, it was found that comparatively higher dry-season baseflows increased the runoff-ratio for the burned catchment for both post-fire seasons (RO = 0.30), WY 2019 and WY 2020 (Figure 3.6; Table 3.2). Runoff events following the fire during the 1st post-fire wet season had noticeably larger response for the burned catchment. The R-B index of the burned catchment was higher than the control

43

catchment, 1.68, for the first year after fire, indicating comparatively higher variation between daily flows. While the increase in R-B index (0.91) in WY 2019 for the control catchment reflects the increased precipitation during that year. Finally, the daily streamflow data for WY 2020 (2nd year post-fire) were differenced, resulting in 131-mm more water translated to direct runoff for the burned catchment. This was a 127% increase in annual direct runoff when compared to similar pre-fire levels of the control catchment. ECOSTRESS PT-JPL ET measurements varied between the two catchments depending on the season (Figure 3.6c). ET differences between the two catchments were highly sinusoidal after the fire, with the largest difference in magnitude (2 to 3-mm) between the control and burned catchments in the dry months (June to September) and the smallest difference in magnitude (0 to 1-mm) during the wet months (November to March). In general, based on ECOSTRESS PT-JPL measurements, the burned catchment had less ET during the post-fire study period than the control catchment for all seasons. The daily streamflow and ET were differenced to estimate storage volume (EQ 3.5). For ECOSTRESS, the calculated difference in total storage volume between the two catchments for WY 2020 (392-mm) showed more storage in the burned catchment relative to the control catchment. This contradicted the SSEBop (1-km) calculation which resulted in more storage in the control catchment relative to the burned catchment. This is a direct result of SSEBop predicting higher annual ET for the burned catchment in WY 2020 relative to the

control (Table 3.2). In summary for WY 2020, ΔETAnnual (Santiago ETAnnual – Coldwater

ETAnnual) measured by SSEBop was -103-mm and 523-mm when measured by ECOSTRESS

PT-JPL. Due to the small scale of the fire processes, ETAnnual for SSEBop (1-km x 1-km) may present error due to sub-pixel contamination in post-fire period. Finally, ΣEVI is shown for WY 2014 to 2020 for both the burned and unburned catchments. Percent difference in ΣEVI between control and burned catchments decreased from 1st year after fire (65%) to 2nd year after fire (26%), indicating rapid recovery of annual accumulation of above-ground biomass following the moderately wet WY 2019.

44

Figure 3.6. Coldwater (CW) and Santiago (Sant) daily rainfall (a), difference in daily evapotranspiration (ET) from WY 2019-2020 (b), and daily streamflow (c). Difference in ECOSTRESS PT-JPL ET is calculated as Sant ET minus CW ET. The dashed green line represents a 60-day rolling average (approximately 2 months) of the difference in ET measurements.

45

Table 3.2. Annual Hydrologic Signatures (runoff-ratio [RO], and Richards-Baker [RB] index) for WY 2014-2020 for Coldwater and Santiago. Cumulative EVI (ΣEVI) and Annual SSEBop ET (ETAnnual) are Also Noted Coldwater (burned) Santiago (control)

WY RO R-B index ΣEVI RO R-B index ΣEVI

(mm/mm) (mm/mm) (ETAnnual (mm/mm) (mm/mm) (ETAnnual

in mm) in mm)

2014 N/a N/a 12.6 (627) 0.01 0.56 12.3 (290) 2015 N/a N/a 13.6 (635) 0.01 0.81 13.8 (319) 2016 N/a N/a 13.9 (615) 0.03 0.54 13.8 (353) 2017 N/a N/a 14.7 (633) 0.16 0.52 14.2 (406) 2018 N/a N/a 11.7 (563) 0.01 0.67 12.4 (336) 2019* 0.30 1.68 7.3 (589) 0.28 0.91 14.3 (589) 2020 0.30 0.52 11.5 (607) 0.13 0.54 14.9 (504) *Runoff data for WY 2019 not available prior December 4, 2018

3.3.5 Spatial and Temporal Recovery of Post-Fire ET Cumulative ECOSTRESS PT-JPL ET collected from February 2019 to October 2020 (111 images) varied with respect to slope aspect, soil burn severity, riparian/hillslope, and vegetation type (Figure 3.7). ET measured on west facing aspects had the highest cumulative total, 599-mm, relative to other directions during this period (Figure 3.7a). North facing aspects had 2nd most with 594-mm, and south and east aspects recorded 585-mm. When compared by soil burn severity classification, areas with high soil burn severity had the fastest recovery and recorded 596-mm across the 111 images, while moderate soil burn severity recorded 591-mm, and low soil burn severity recorded 574-mm (Figure 3.7b). The unburned/very low category was ignored for this analysis due to the area not being affected by the fire. The riparian area was compared to the hillslope (any area outside of the 70-meter defined buffer). Interestingly, riparian area (574-mm) recorded lower cumulative total than hillslope (584-mm) (Figure 3.7c). When compared to vegetation species before the fire, ET had the highest cumulative total for Montane Hardwood which recorded 568-mm (Figure 3.7d). It is noted that that due to the smaller sub-areas for Figure 3.7c/d, the totals were lower from not accumulating the same number of images.

46

Figure 3.7. Cumulative ECOSTRESS evapotranspiration (ET) collected from WY 2019 (35 images) and WY 2020 (76 images) for the Holy Fire with respect to slope aspect (a), soil burn severity (b), riparian versus hillslope (c), and pre-fire vegetation species (d).

3.4 DISCUSSION

3.4.1 Role of Climate Extremes Prior to the fire, WYs 2012-2016 and 2018 ranged from near normal to moderately dry (Figure 3.2). This period was a dry stretch of years, where the previous classified wet year occurred in 2011. Montane hardwood species burned at highest soil burn severity for the

47

Holy Fire, more so than any other vegetation type. Montane hardwoods are typically well adapted to drought, however, can experience high water stress and increased tree mortality during longer drought periods (Guarín & Taylor, 2005), thus leading to elevated potential for fire (Taylor et al., 2008). ECOSTRESS data are only available since June 29, 2018. Thus, this work was unable to evaluate evaporative stress before the fire; however, based on the available records, there is modest agreement (R2 = 0.47; n = 103) with the pre-fire image on August 2, 2018 between ECOSTRESS PT-JPL and SSEBop. Recognizing the limitations of this study and the preliminary results presented, it is advocated that higher resolution products such as ECOSTRESS will be able to address science questions for future fires once more data becomes available. In the meantime, by utilizing SSEBop, a lower resolution product, the monthly ET for the month prior to the fire was measured to be the lowest in the pre-fire study period for the month of July (64 mm/month). Since no other fires or landcover change occurred during this time, the high intensity and severity of the Holy Fire is attributed to the extended dry period preceding the fire event. In the aftermath of the fire, high intensity and short duration storms translated to large flooding and hyperconcentrated flows in burned areas during the first storm season following the Holy Fire (Guilinger et al., 2020)⁠. It is likely that high intensity storms will continue to become more common over this next century, where studies have reported that rainfall intensities will increase proportionately with global warming in the Pacific Southwest (Prein et al., 2017)⁠. Thus, more work is needed to link post-fire vegetation recovery to hydrologic processes to better predict post-fire flooding and erosion under extreme climate scenarios.

3.4.2 High Uncertainty in Post-Fire ET Prior to the fire, there was a modest correlation between SSEBop (scaled to daily) and ECOSTRESS PT-JPL daily ET (R2 = 0.47; Slope = 0.36). After the fire however, there was a large degree of uncertainty measured between the two ET products (R2 = 0.00 to 0.04; Slope = -0.10 to 0.12). This is attributed to the high spatial heterogeneity of soil burn severity for this fire. For example, Chen et al. (2016) found SSEBop model to be most sensitive to land surface temperature and reference ET inputs, especially in the non-growing season in dry areas. SSEBop at 1-km spatial resolution may not be suitable for capturing fine-scale post- fire processes of semi-arid and dry regions and is noted to be most reliable for large-area ET

48

estimation (Chen et al., 2016). Therefore, utilizing SSEBop in a post-fire setting at the sub- catchment scale can introduce sub-pixel contamination within the burn scar that can affect highly sensitive input parameters such as land surface temperature and reference ET. ECOSTRESS PT-JPL ET post-fire images appear to have more agreeable results with the study site; however, more work needs to be done to field validate ECOSTRESS PT-JPL in post-fire settings across diverse ecoregions and longer timespans. Along with ground-based validation efforts, the use of Unmanned Aircraft Systems (UAS) may present an opportunity to improve certainty in post-fire assessments of satellite-based ET (Fernández-Guisuraga et al., 2018).

3.4.3 Linking Ecohydrological Recovery After Fire to Observed Data using Hydrologic Signatures Coldwater had an R-B index of 1.68 for the 1st year after fire, which was higher than any value recorded by Santiago for the period WY 2014-2020. The RB-index decreased for the 2nd year after fire to a value of 0.52, returning to control levels measured at Santiago of 0.54. This highlights the decrease in flashy floods for the 2nd year after fire for the burned catchment, agreeing with a recent study of the sediment and runoff processes of the Holy Fire (Guilinger et al., 2020)⁠. The authors documented rapid reestablishment of vegetation during the wet period of 2019, likely driven by the moderately wet WY 2019 observed in the SPI analysis (Figure 3.2). This also agrees with the SSEBop ET timeseries, which showed a rapid resurgence of ET during the following year in WY 2020 (Figure 3.2). The runoff-ratio (RO) highlighted the dramatic increase in dry season low flows after fire, where Coldwater RO were the same (0.30) for both WY 2019 and 2020. This is similar to observations by Kinoshita and Hogue (2015), who showed elevated RO in burned catchments for several post-fire seasons following the 2003 Old Fire (Kinoshita & Hogue, 2015)⁠. The Coldwater and Santiago RO appear to be similar for WY 2019, however, this could be due to the above average amount of rainfall during WY 2019, 1,036 mm, resulting in more saturation excess runoff and muting the comparative effects from fire. In general, increase in dry season low flows highlight the decrease in deep root vegetation pathways to stop through-flow (Atchley et al., 2018; Kinoshita & Hogue, 2011)⁠.

49

To link the runoff measurements to the vegetation recovery, a simple water balance calculation using ECOSTRESS PT-JPL ET was carried out for WY 2020 to approximate the difference in storage between a control and burned catchment. ECOSTRESS PT-JPL ET measured comparatively lower ET in the burned catchment after fire. Decreased ET in similar regions after fire has also been observed in other studies such as Poon and Kinoshita (2018), Ma et al. (2020), and Soulis et al. (2021). This reduction in ET in the burned catchment, coupled with the increase in streamflow, likely controlled the storage component, which was found to be 392-mm higher than the control catchment. This analysis demonstrates that ECOSTRESS PT-JPL ET can be useful in future modeling and calibration efforts of spatially distributed processes (Figure 3.6; Figure 3.7). This study highlights the importance of higher spatial resolution products such as ECOSTRESS PT-JPL (70-m) to make advances in modeling and predicting coupled ecohydrological processes. The high uncertainty observed between SSEBop and ECOSTRESS post-fire ET at the small catchment scale for both post-fire years were likely due to the complex spatial heterogeneity of the burned conditions, leading to sub-pixel contamination of the SSEBop pixels (Table 3.2; Figure A.5). ECOSTRESS PT-JPL ET, having a smaller footprint, can better account for spatially heterogeneity present in post-fire parameters such as land surface temperature and reference ET (Chen et al., 2016). This was demonstrated by differentiating the post-fire ECOSTRESS ET patterns with respect to slope aspect, soil burn severity, riparian/hillslope, and vegetation type (Figure 3.7).

3.4.4 Post-fire Revegetation and Recovery Strategies after the Holy Fire Burned Area Emergency Response (BAER) treatment recommended strategies to mitigate post-fire risks and promote healthy regrowth included identification of noxious weed suppression points/areas, signage and gates to close dangerous areas, hazmat stabilization areas for residential areas, road stabilization treatments, storm proofing of trails, and identification of threatened species habitat area (California Gnatcatcher, or Polioptila californica). There were no re-seeding efforts after the Holy Fire, and thus, ET measurements captured natural recovery following a moderately wet WY 2019. It is noted that ET recovery rates cannot differentiate between vegetation species present before and after fire without

50

field-based validation. This is relevant, as fast-growing grasses were observed in field visits in several regions (see supplementary Figure A.6).

3.5 CONCLUSION Ecohydrological processes such as streamflow and evapotranspiration (ET) are highly variable after fire in Mediterranean systems with respect to time and space, therefore prompt the need for sensitive monitoring of the recovery. In this study, a suite of remote sensing products were used to quantify effects related to fire using hydrologic signatures of paired catchments Santiago (control) and Coldwater (burned) after the 2018 Holy Fire in Orange County and Riverside County, California. The burned areas in the 2018 Holy Fire were more likely to be classified as high soil burn severity if pre-fire ΣEVI were large (14 to 16) and contained high proportion of montane hardwood species (>30%). Based on SSEBop monthly ET and SPI analyses, it was observed that actual ET had a sustained decline leading up to the fire accompanied by low amounts of rainfall during WYs 2012-2016 and 2018. While measurements from ECOSTRESS were not available for the pre-fire period, this study highlights useful future applications for fire management and identifying specific areas that have high evaporative stress and dry conditions conducive to fire. Post-fire, there was high uncertainty in ET measurements between ECOSTRESS PT-JPL and SSEBop satellite-based products. This is attributed to the high spatial heterogeneity of this fire and coarse spatial resolution of SSEBop (1-km x 1-km), which cannot adequately capture the variability of sensitive parameters like land surface temperature and reference ET inherent in the burned area. On the other hand, ECOSTRESS PT-JPL supported other recent studies in semi-arid regions which showed reductions in post-fire ET when compared to the control catchment. Also, ECOSTRESS PT-JPL was successfully able to highlight variability in ET recovery for fine scale features at 70-m resolution such as specific slope aspects, soil burn severity, and vegetation species. Based on ECOSTRESS PT-JPL, areas that were recovering quickly included west and north facing slopes, areas burned at high soil burn severity, and areas that contained high proportions of montane species before the fire. It is concluded that ECOSTRESS PT-JPL daily ET product appears to be effective in monitoring ecohydrological recovery for small catchments (< 50 km2) after fire; however, more ground-

51 based field research is needed to further validate its uses in fire applications globally across diverse ecoregions.

52

CHAPTER 4

SUMMARY

This thesis presented methods to improve rapid assessment and long-term prediction tools for ecohydrological processes such as streamflow and evapotranspiration (ET) in small catchments in southern California affected by fire. This study focused mainly on small, rural catchments (<50 km2) that were contained within the Santa Ynez Mountains, San Gabriel Mountains, San Bernardino Mountains, Santa Ana Mountains, and San Jacinto Mountains. In Chapter 2, a random forest machine learning algorithm was built with 45 catchment parameters to predict post-fire peak streamflow and compared these results to an existing methodology, Rowe et al. (1949) look-up table method, and observed streamflow (Wilder et al., 2021). It was demonstrated that Rowe et al. (1949), a flood frequency model, overgeneralized catchment processes and did not adequately represent the spatial and temporal variability in systems affected by wildfire and extreme weather events and often underpredicted peak streamflow without sediment bulking factors. Towards improving rapid post-fire predictions, this work developed a random forest flood forecasting model, which performed well. The improvement was expected, given the type of input data and difference in models, however, the modeling exercise demonstrated the importance and reliance on data availability of important storm dependent parameters such as rainfall intensity and time after fire. The important parameters identified by the machine learning techniques were used create a simple regression to calculate post‐fire peak streamflow in small catchments (less than 20 km2) in southern California during the first year after fire (R2 = 0.82; RMSE = 6.59 cms/km2) which can be used as an interim tool by post‐fire risk assessment teams. Future efforts using machine learning to predict post-fire flooding would be improved greatly if there are significant increases in high resolution rainfall intensity data, sub-hourly streamgaging, and sediment loading in channels. Future efforts may also consider using a classification-based machine learning algorithm to predict the outcome. For example, using

53

various catchment parameters as inputs, and designing the model to output a classified outcome such as debris flow, mud flow, or flooding. Finally, the prediction accuracy for future models can be improved by collecting more samples to train the model. In Chapter 3, ECOsystem Spaceborne Thermal Radiometer Experiment on Space Station (ECOSTRESS) PT-JPL ET measurements were used to characterize the 2018 Holy Fire in southern California to improve vegetation assessments and resource management. Using a pixel-by-pixel analysis it was demonstrated that Operational Simplified Surface Energy Balance Model (SSEBop) daily scaled ET had lower correlation to ECOSTRESS PT- JPL post-fire (R2 = 0.00 to 0.04; slope = -0.10 to 0.12; sample size =103) than pre-fire (R2 = 0.47; slope = 0.36; sample size = 103). This highlights the higher uncertainty in post-fire SSEBop ET caused by the large spatial heterogeneity of the burn conditions. Further, ECOSTRESS PT-JPL daily ET was reduced for the burned catchment compared to the control, which agreed with previous studies for semi-arid regions. When comparing spatial recovery using ECOSTRESS PT-JPL, areas burned at highest soil burn severity had the largest increase in cumulative post-fire ET. Also, areas that consisted of montane species before the fire, had the highest cumulative post-fire ET. This work demonstrated the potential advantages of using ECOSTRESS PT-JPL to improve vegetation assessments with the added benefit of increased spatial and temporal resolution. More work is needed in identifying areas of high-water stress before fire, as well as validating post-fire ET across diverse ecoregions using a combination of ground-based stations and Unmanned Aerial Systems (UAS) in the future. In conclusion, this study used machine learning to develop flood forecasting models to improve prediction methods for floods that immediately follow fires in southern California. As demonstrated in Chapter 3, above-ground biomass and vegetation production plays a role in recovery of hydrologic processes following fires in southern California. There is the potential to incorporate ECOSTRESS PT-JPL ET measurements as a predictor in future machine learning models to help improve monitoring of annual and seasonal flows during the recovery period. Future work may also look to incorporate satellite-based products such as ECOSTRESS PT-JPL ET and others to improve prediction accuracy and improve scalability for worldwide applications.

54

REFERENCES

Atchley, A. L., Kinoshita, A. M., Lopez, S. R., Trader, L., & Middleton, R. (2018). Simulating surface and subsurface water balance changes due to burn severity. Vadose Zone Journal, 17(1), 1-13. https://doi.org/10.2136/vzj2018.05.0099 Baker, D. B., Richards, R. P., Loftus, T. T., & Kramer, J. W. (2004). A new flashiness index: Characteristics and applications to Midwestern rivers and streams. Journal of the American Water Resources Association, 40(2), 503-522. https://doi.org/10.1111/j.1752-1688.2004.tb01046.x Bell, C. E., Ditomaso, J. M., & Brooks, M. L. (2009). Invasive plants and wildfires in Southern California. University of California, Division of Agriculture and Natural Resources. https://doi.org/10.3733/ucanr.8397 Biddinger, T., Gallegos, A., Janeki, A., TenPas, J., & Weaver, R. (2003). BAER watershed assessment report, 2003 Grand Prix and Old Fire. San Bernardino National Forest, unpublished report. Bonaccorso, B., Bordi, I., Cancelliere, A., Rossi, G., & Sutera, A. (2003). Spatial variability of drought: An analysis of the SPI in Sicily. Water Resources Management, 17(4), 273-296. https://doi.org/10.1023/A:1024716530289 Box, G. E. P., & Cox, D. R. (1964). An analysis of transformations. Journal of the Royal Statistical Society: Series B (Methodological), 26(2), 211-243. https://doi.org/10.1111/j.2517-6161.1964.tb00553.x Breiman L. (2001). Random forests. Machine Learning, 45(1), 5–32. CALFIRE. (2020). Top 20 largest California wildfires. https://www.fire.ca.gov/media/4jandlhh/top20_acres.pdf CALFIRE-FRAP. (2015). Vegetation (fveg) - CALFIRE FRAP [ds1327]. https://map.dfg.ca.gov/metadata/ds1327.html Cannon, S. H., & DeGraff, J. (2009). The increasing wildfire and post-fire debris-flow threat in western USA, and implications for consequences of climate change. In Landslides– disaster risk reduction (pp. 177-190). Springer. Cannon, S. H., Gartner, J. E., Rupert, M. G., & Michael, J. A. (2004). Emergency assessment of debris-flow hazards from basins burned by the Padua fire of 2003, Southern California (US Geological Survey Open-File Report 1072). US Geological Survey. Cannon, S. H., Gartner, J. E., Wilson, R. C., Bowers, J. C., & Laber, J. L. (2008). Storm rainfall conditions for floods and debris flows from recently burned areas in

55

southwestern Colorado and southern California. Geomorphology, 96(3-4), 250-269. https://doi.org/10.1016/j.geomorph.2007.03.019 CGS [California Geological Survey]. (2002). California geomorphic provinces (Note 36). California Department of Conservation, California Geological Survey. CGS [California Geological Survey]. (2018). Geology of California. Author. Chawner, W. D. (1935). Alluvial fan flooding: The Montrose, California, Flood of 1934. Geographical Review, 25(2), 255-263. https://doi.org/10.2307/209600 Chen, M., Senay, G. B., Singh, R. K., & Verdin, J. P. (2016). Uncertainty analysis of the Operational Simplified Surface Energy Balance (SSEBop) model at multiple flux tower sites. Journal of Hydrology, 536, 384-399. https://doi.org/10.1016/j.jhydrol.2016.02.026 Clarke, R. T. (2002). Estimating trends in data from the Weibull and a generalized extreme value distribution. Water Resources Research, 38(6). https://doi.org/10.1029/2001wr000575 Cleland, D. T., Freeouf, J. A., Keys, J. E., Nowacki, G. J., Carpenter, C. A., & McNab, W. H. (2007). Ecological subregions: Sections and subsections for the conterminous United States (General Technical Report WO-76D). U.S. Department of Agriculture, Forest Service. Coalitions & Collaboratives, Inc. (Producer). (2020). After the flames [Video]. https://aftertheflames.com/science-sessionresources/ Debano, L. F. (2000). The role of fire and soil heating on water repellency in wildland environments: A review. Journal of Hydrology, 231-232, 195-206. https://doi.org/10.1016/S0022-1694(00)00194-3 Dettinger, M. (2011). Climate change, atmospheric rivers, and floods in California - a multimodel analysis of storm frequency and magnitude changes. Journal of the American Water Resources Association, 47(3), 514-523. https://doi.org/10.1111/j.1752-1688.2011.00546.x DiBiase, R. A., & Lamb, M. P. (2020). Dry sediment loading of headwater channels fuels post-wildfire debris flows in bedrock landscapes. Geology, 48(2), 189-193. https://doi.org/10.1130/G46847.1 Doerr, S. H., Shakesby, R. A., & Macdonald, L. H. (2009). Soil water repellency: A key factor in post-fire erosion. In A. Cerda & P. R. Robichaud (Eds.), Fire effects on soils and restoration strategies. Routledge. Doerr, S. H., Shakesby, R. A., & Walsh, R. P. D. (2000). Soil water repellency: Its causes, characteristics and hydro-geomorphological significance. Earth Science Reviews, 51(1-4), 33-65. https://doi.org/10.1016/S0012-8252(00)00011-8 Feller, W. (1968). An introduction to probability theory and its applications (Vol. 1). John Wiley & Sons.

56

Fernández, C., & Vega, J. A. (2016). Modelling the effect of soil burn severity on soil erosion at hillslope scale in the first year following wildfire in NW Spain. Earth Surface Processes and Landforms, 41(7), 928-935. https://doi.org/10.1002/esp.3876 Fernández-Guisuraga, J. M., Sanz-Ablanedo, E., Suárez-Seoane, S., & Calvo, L. (2018). Using unmanned aerial vehicles in postfire vegetation survey campaigns through large and heterogeneous areas: Opportunities and challenges. Sensors, 18(2), 586. https://doi.org/10.3390/s18020586 Fisher, J. B., Lee, B., Purdy, A. J., Halverson, G. H., Dohlen, M. B., Cawse-Nicholson, K., Wang, A., Anderson, R. G., Aragon, B., Arain, M. A., Baldocchi, D. D., Baker, J. M., Barral, H., Bernacchi, C. J., Bernhofer, C., Biraud, S. C., Bohrer, G., Brunsell, N., Cappelaere, B., … Hook, S. (2020). Ecostress: NASA’s next generation mission to measure evapotranspiration from the International Space Station. Water Resources Research, 56(4), e2019WR026058. https://doi.org/10.1029/2019WR026058 Flynn, K., Kirby, W., & Hummel, P. (2006). User’s manual for program PeakFQ, annual flood-frequency analysis using bulletin 17B guidelines (Chapter 4 of Book 4, Section B). U.S. Department of the Interior, U.S. Geological Survey Foltz, R. B., Robichaud, P. R., & Rhee, H. (2009). A synthesis of post-fire road treatments for BAER teams: methods, treatment effectiveness, and decision making tools for rehabilitation (General Technical Report RMRS-228). U.S. Department of Agriculture, Forest Service, Rocky Mountain Research Station. FRAP [Fire and Resource Assessment Program]. (2015). State fire perimeter database. California Department of Forestry and Fire Protection. FRAP [Fire and Resource Assessment Program]. (2018). California's forests and rangelands: 2017 assessment. California Department of Forestry and Fire Protection. Gabet, E. J. (2003). Post‐fire thin debris flows: Sediment transport and numerical modelling. Earth Surface Processes and Landforms: The Journal of the British Geomorphological Research Group, 28(12), 1341-1348. Gartner, J. E., Cannon, S. H., & Santi, P. M. (2014). Empirical models for predicting volumes of sediment deposited by debris flows and sediment-laden floods in the transverse ranges of southern California. Engineering Geology, 176(24), 45-56. https://doi.org/10.1016/j.enggeo.2014.04.008 Goodrich, D. C., Burns, I. S., Unkrich, C. L., Semmens, D. J., Guertin, D. P., Hernandez, M., & Levick, L. R. (2012). KINEROS2/AGWA: model use, calibration, and validation. Transactions of the ASABE, 55(4), 1561-1574. Goodrich, D., Canfield, H. E., Burns, I. S., Semmens, D., Miller, S., Hernandez, M., Levick, L. R., Guertin, W. G., & Kepner, W. (2005, July). Rapid post-fire hydrologic watershed assessment using the AGWA GIS-based hydrologic modeling tool. In G. E. Moglen (Ed.), Managing watersheds for human and natural impacts: engineering, ecological, and economic challenges (pp. 1-12). ASCE.

57

Gorelick, N., Hancher, M., Dixon, M., Ilyushchenko, S., Thau, D., & Moore, R. (2017). Google e1arth engine: Planetary-scale geospatial analysis for everyone. Remote Sensing of Environment, 202, 18-27. https://doi.org/10.1016/j.rse.2017.06.031 Guarín, A., & Taylor, A. H. (2005). Drought triggered tree mortality in mixed conifer forests in Yosemite National Park, California, USA. Forest Ecology and Management, 218(1-3), 229-244. https://doi.org/10.1016/j.foreco.2005.07.014 Guilinger, J. J., Gray, A. B., Barth, N. C., & Fong, B. T. (2020). The evolution of sediment sources over a sequence of postfire sediment-laden flows revealed through repeat high-resolution change detection. Journal of Geophysical Research: Earth Surface, 125(10), e2020JF005527. https://doi.org/10.1029/2020JF005527 Harden, J. W., & Matti, J. C. (1989). Holocene and late Pleistocene slip rates on the San Andreas fault in Yucaipa, California, using displaced alluvial-fan deposits and soil chronology. Geological Society of America Bulletin. https://doi.org/10.1130/0016- 7606(1989)101<1107:HALPSR>2.3.CO;2 Hogue, T. S., Sorooshian, S., Gupta, H., Holz, A., & Braatz, D. (2000). A multistep automatic calibration scheme for river forecasting models. Journal of Hydrometeorology, 1(6). https://doi.org/10.1175/1525- 7541(2000)001<0524:AMACSF>2.0.CO;2 Horton, R. E. (1945). Erosional development of streams and their drainage basins; hydrophysical approach to quantitative morphology. Geological Society of America Bulletin, 56(3), 275-370. Huffman, E. L., MacDonald, L. H., & Stednick, J. D. (2001). Strength and persistence of fire-induced soil hydrophobicity under ponderosa and lodgepole pine, Colorado Front Range. Hydrological Processes, 15, 2877-2892. https://doi.org/10.1002/hyp.379 Johansen, M. P., Hakonson, T. E., & Breshears, D. D. (2001). Post‐fire runoff and erosion from rainfall simulation: Contrasting forests with shrublands and grasslands. Hydrological Processes, 15(15), 2953-2965. Kean, J. W., Staley, D. M., Lancaster, J. T., Rengers, F. K., Swanson, B. J., Coe, J. A., Hernandez, J. L., Sigman, A. J., Allstadt, K. E., & Lindsay, D. N. (2019). Inundation, flow dynamics, and damage in the 9 January 2018 Montecito debris-flow event, California, USA: Opportunities and challenges for post-wildfire risk assessment. Geosphere. https://doi.org/10.1130/GES02048.1 Keeley, J. E. (2009). Fire intensity, fire severity and burn severity: A brief review and suggested usage. International Journal of Wildland Fire, 18(1), 116-126. https://doi.org/10.1071/WF07049 Keeley, J. E., & Keeley, S. C. (1981). Post-fire regeneration of southern California Chaparral. American Journal of Botany, 68(4), 524-530. https://doi.org/10.2307/2443028 Keeley, J. E., Fotheringham, C. J., & Moritz, M. A. (2004). Lessons from the October 2003 wildfires in Southern California. Journal of Forestry, 102(7), 26-31. https://doi.org/10.1093/jof/102.7.26

58

Keller, E. A., Valentine, D. W., & Gibbs, D. R. (1997). Hydrological response of small watersheds following the Southern California painted cave fire of June 1990. Hydrological Processes, 11(4), 401-414. Kinoshita, A. M., & Hogue, T. S. (2011). Spatial and temporal controls on post-fire hydrologic recovery in Southern California watersheds. Catena, 87(2), 240-252.. https://doi.org/10.1016/j.catena.2011.06.005 Kinoshita, A. M., & Hogue, T. S. (2015). Increased dry season water yield in burned watersheds in Southern California. Environmental Research Letters, 10(1). https://doi.org/10.1088/1748-9326/10/1/014003 Kinoshita, A. M., Chin, A., Simon, G. L., Briles, C., Hogue, T. S., O’Dowd, A. P., Gerlak, A. K., & Albornoz, A. U. (2016). Wildfire, water, and society: Toward integrative research in the “Anthropocene.” Anthropocene, 16(2016), 166-27. https://doi.org/10.1016/j.ancene.2016.09.001 Kinoshita, A. M., Hogue, T. S., & Napper, C. (2014). Evaluating pre- and post-fire peak discharge predictions across western U.S. watersheds. Journal of the American Water Resources Association, 50(6), 1540-1557. https://doi.org/10.1111/jawr.12226 Kohli, G., Lee, C. M., Fisher, J. B., Halverson, G., Variano, E., Jin, Y., Carney, D., Wilder, B. A., & Kinoshita, A. M. (2020). ECOSTRESS and CIMIS: A Comparison of Potential and Reference Evapotranspiration in Riverside County, California. Remote Sensing, 12(24), 4126. https://doi.org/10.3390/rs12244126 Lamjiri, M. A., Dettinger, M. D., Ralph, F. M., Oakley, N. S., & Rutz, J. J. (2018). Hourly analyses of the large storms and atmospheric rivers that provide most of California’s precipitation in only 10 to 100 hours per year. San Francisco Estuary and Watershed Science, 16(4), 1-17. https://doi.org/10.15447/sfews.2018v16iss4art1 Lancaster, J. T., Swanson, B. J., Lukashova, S., Oakley, N., Lee, J. B., Spangler, E., Hernandez, J. L., Olson, B. P. E., DeFrisco, M. J., Lindsay, D. J., Schwartz, Y. J., McCrea, S. E., Roffers, P. D., & C. M. Tran. (in press). Observations and analyses of the 9 January 2018 debris flow disaster, Santa Barbara County. Environmental and Engineering Geoscience. Lavé, J., & Burbank, D. (2004). Denudation processes and rates in the Transverse Ranges, southern California: Erosional response of a transitional landscape to external and anthropogenic forcing. Journal of Geophysical Research: Earth Surface, 109(F1). https://doi.org/10.1029/2003jf000023 Lentile, L. B., Morgan, P., Hudak, A. T., Bobbitt, M. J., Lewis, S. A., Smith, A. M., & Robichaud, P. R. (2007). Post-fire burn severity and vegetation response following eight large wildfires across the western United States. Fire Ecology, 3(1), 91-108. https://doi.org/10.4996/fireecology.0301091 Lewis, S. A., Wu, J. Q., & Robichaud, P. R. (2006). Assessing burn severity and comparing soil water repellency, Hayman Fire, Colorado. Hydrological Processes, 20(1), 1-16. https://doi.org/10.1002/hyp.5880

59

Ma, Q., Bales, R. C., Rungee, J., Conklin, M. H., Collins, B. M., & Goulden, M. L. (2020). Wildfire controls on evapotranspiration in California’s Sierra Nevada. Journal of Hydrology, 590, 125364. https://doi.org/10.1016/j.jhydrol.2020.125364 MacDonald, L. H., & Huffman, E. L. (2004). Post-fire soil water repellency. Soil Science Society of America Journal, 68(5), 1729-1734. Mayor, A. G., Bautista, S., Llovet, J., & Bellot, J. (2007). Post-fire hydrological and erosional responses of a Mediterranean landscpe: Seven years of catchment-scale dynamics. Catena, 71(1), 68-75. https://doi.org/10.1016/j.catena.2006.10.006 McKay, L., Bondelid, T., Dewald, T., Rea, A., Johnston, C., & Moore, R. (2015). NHDPlus version 2: User guide (data model version 2.1). Horizon Systems. McMillan, H. (2020). Linking hydrologic signatures to hydrologic processes: A review. Hydrological Processes, 34(6), 1393-1409. https://doi.org/10.1002/hyp.13632 Miller, J. D., Knapp, E. E., Key, C. H., Skinner, C. N., Isbell, C. J., Creasy, R. M., & Sherlock, J. W. (2009). Calibration and validation of the relative differenced Normalized Burn Ratio (RdNBR) to three measures of fire severity in the Sierra Nevada and Klamath Mountains, California, USA. Remote Sensing of Environment, 113(3), 645-656. Miller, J. D., Nyhan, J. W., & Yool, S. R. (2003). Modeling potential erosion due to the Cerro Grande Fire with a GIS-based implementation of the Revised Universal Soil Loss Equation. International Journal of Wildland Fire, 12(1), 85-100. Miller, N. L., & Schlegel, N. J. (2006). Climate change projected fire weather sensitivity: California Santa Ana wind occurrence. Geophysical Research Letters, 33(15). https://doi.org/10.1029/2006GL025808 Miller, V. C. (1953). Quantitative geomorphic study of drainage basin characteristics in the Clinch Mountain area, Virginia and Tennessee [Technical report]. Columbia University, Department of Geology. Milly, P. C. D., Betancourt, J., Falkenmark, M., Hirsch, R. M., Kundzewicz, Z. W., Lettenmaier, D. P., & Stouffer, R. J. (2008). Climate change: Stationarity is dead: Whither water management? Science, 319(5863), 573-574. https://doi.org/10.1126/science.1151915 Moody, J. A. (2012). An analytical method for predicting postwildfire peak discharges. USGS Investigations Report. Moody, J. A., & Martin, D. A. (2001). Initial hydrologic and geomorphic response following a wildfire in the Colorado front range. Earth Surface Processes and Landforms, 26(10), 1049-1070. https://doi.org/10.1002/esp.253 Moody, J. A., Martin, D. A., Haire, S. L., & Kinner, D. A. (2008). Linking runoff response to burn severity after a wildfire. Hydrological Processes: An International Journal, 22(13), 2063-2074. Moody, J. A., Shakesby, R. A., Robichaud, P. R., Cannon, S. H., & Martin, D. A. (2013). Current research issues related to post-wildfire runoff and erosion processes. Earth-

60

Science Reviews, 122(July 2013), 10-37. https://doi.org/10.1016/j.earscirev.2013.03.004 Moody, J. A., Shakesby, R. A., Robichaud, P. R., Cannon, S. H., & Martin, D. A. (2013). Current research issues related to post-wildfire runoff and erosion processes. Earth- Science Reviews, 122, 10-37. Moore, M., Bigginger, T., Thornton, J., Wright, K., & Stewart, C. (2009). BAER watershed assessment report, Station Fire [Unpublished report]. Musselman, K. N., Molotch, N. P., & Margulis, S. A. (2017). Snowmelt response to simulated warming across a large elevation gradient, southern Sierra Nevada, California. The Cryosphere, 11(6), 2847–2866. Neary, D. G., Ryan, K. C., & DeBano, L. F. (2005). Wildland fire in ecosystems: Effects of fire on soils and water (General Technical Reports. RMRS-GTR42-vol. 4.). US Department of Agriculture, Forest Service, Rocky Mountain Research Station. Neiman, P. J., Ralph, F. M., Wick, G. A., Kuo, Y. H., Wee, T. K., Ma, Z., Taylor, G. H., & Dettinger, M. D. (2008). Diagnosis of an intense atmospheric river impacting the pacific northwest: Storm summary and offshore vertical structure observed with COSMIC satellite retrievals. Monthly Weather Review, 136(11). https://doi.org/10.1175/2008MWR2550.1 O’Connor, J. E., & Costa, J. E. (2004). The world’s largest floods, past and present: Their causes and magnitudes. US Geological Survey Circular. https://doi.org/10.3133/cir1254 Oakley, N. S., Lancaster, J. T., Kaplan, M. L., & Ralph, F. M. (2017). Synoptic conditions associated with cool season post-fire debris flows in the Transverse Ranges of southern California. Natural Hazards, 88, 327-354. https://doi.org/10.1007/s11069-017-2867-6 Obrist, D., Yakir, D., & Arnone, J. A. (2004). Temporal and spatial patterns of soil water following wildfire-induced changes in plant communities in the Great Basin in Nevada, USA. Plant and Soil, 262, 1-12. https://doi.org/10.1023/B:PLSO.0000037026.93675.a2 Orfanidis, S. J. (1996). Introduction to signal processing. Prentice Hall. Pal, M. (2005). Random forest classifier for remote sensing classification. International Journal of Remote Sensing, 26(1), 217-222. https://doi.org/10.1080/01431160412331269698 Parrett, C., Veilleux, A., Stedinger, J. R., Barth, N. A., Knifong, D. L., & Ferris, J. C. (2011). Regional skew for California, and flood frequency for selected sites in the Sacramento-San Joaquin River Basin, based on data through water year 2006. U. S. Geological Survey. Parsons, A., Robichaud, P. R., Lewis, S. A., Napper, C., & Clark, J. T. (2010). Field guide for mapping post-fire soil burn severity (Gen. Tech. Rep. RMRS-GTR-243). U.S. Department of Agriculture, Forest Service, Rocky Mountain Research Station.

61

Peel, M. C., Finlayson, B. L., & McMahon, T. A. (2007). Updated world map of the Köppen- Geiger climate classification. Hydrology and Earth System Sciences, 11(5), 1633- 1644. https://doi.org/10.5194/hess-11-1633-2007 Pierson, T. (2005). Distinguishing between debris flows and floods from field evidence in small watersheds. U.S. Geological Survey. Poon, P. K., & Kinoshita, A. M. (2018). Spatial and temporal evapotranspiration trends after wildfire in semi-arid landscapes. Journal of Hydrology, 559, 71-83. https://doi.org/10.1016/j.jhydrol.2018.02.023 Prater, M. R., & DeLucia, E. H. (2006). Non-native grasses alter evapotranspiration and energy balance in Great Basin sagebrush communities. Agricultural and Forest Meteorology, 139(1-2), 154-163. https://doi.org/10.1016/j.agrformet.2006.08.014 Prein, A. F., Rasmussen, R. M., Ikeda, K., Liu, C., Clark, M. P., & Holland, G. J. (2017). The future intensification of hourly precipitation extremes. Nature Climate Change, 7, 48- 52. https://doi.org/10.1038/nclimate3168 Radeloff, V. C., Hammer, R. B., Stewart, S. I., Fried, J. S., Holcomb, S. S., & McKeefry, J. F. (2005). The wildland-urban interface in the United States. Ecological Applications, 15(3), 799-805. https://doi.org/10.1890/04-1413 Ralph, F. M., & Dettinger, M. D. (2012). Historical and national perspectives on extreme west coast precipitation associated with atmospheric rivers during December 2010. Bulletin of the American Meteorological Society, 93(6), 783–790. https://doi.org/10.1175/BAMS-D-11-00188.1 Raphael, M. N. (2003). The Santa Ana Winds of California. Earth Interactions, 7(8), 1-13. https://doi.org/10.1175/1087-3562(2003)007<0001:tsawoc>2.0.co;2 Reback, J., McKinney, W., jbrockmendel, Van den Bossche, J., Augspurger, T., Cloud, P., gfyoung, Sinhrks, Klein, A., Roeschke, M., Hawkins, S., Tratner, J., She, C., Ayd, W., Petersen, T., Garcia, M., Schendel, J., Hayden, A., MomIsBestFriend, … Mehyar, M. (2020, March 18). pandas-dev/pandas: Pandas 1.0.3 (Version v1.0.3) [Software]. Zenodo. http://doi.org/10.5281/zenodo.3715232 Robichaud, P. R., Elliot, W. J., Pierson, F. B., Hall, D. E., & Moffet, C. A. (2007). Predicting postfire erosion and mitigation effectiveness with a web-based probabilistic erosion model. Catena, 71(2), 229-241. https://doi.org/10.1016/j.catena.2007.03.003 Robichaud, Peter R., Elliot, W. J., & Wagenbrenner, J. W. (2011, September 18-21). Probabilistic soil erosion modeling using the Erosion Risk Management Tool (ERMiT) after wildfires [Paper presentation]. International Symposium on Erosion and Landscape Evolution, Anchorage, Alaska. Rowe, P. B., Countryman, O. M., & Storey, H. C. (1949). Probable peak discharges and erosion rates from southern California watersheds as influenced by fire. US Department of Agriculture, Forest Service. Safford, H. D., & Van de Water, K. M. (2013). Using Fire Return Interval Departure (FRID) analysis to map spatial and temporal changes in fire frequency on National Forest

62

lands in California (Res. Pap. PSW-RP-266). Department of Agriculture, Forest Service, Pacific Southwest Research Station. Savoca, M. E., Senay, G. B., Maupin, M. A., Kenny, J. F., & Perry, C. A. (2013). Actual evapotranspiration modeling using the operational simplified surface energy balance (SSEBop) approach (U.S Geological Survey Scientific Investigations Report 2013– 5126). U.S. Geological Survey. Saxe, S., Hogue, T. S., & Hay, L. (2018). Characterization and evaluation of controls on post-fire streamflow response across western US watersheds. Hydrology and Earth System Sciences, 22(2), 1221-1237. https://doi.org/10.5194/hess-22-1221-2018 Schmidt, L., Heße, F., Attinger, S., & Kumar, R. (2020). Challenges in applying machine learning models for hydrological inference: A case study for flooding events across Germany. Water Resources Research, 56(5). https://doi.org/10.1029/2019WR025924 Schumm, S. A. (1956). Evolution of drainage systems and slopes in badlands at Perth Amboy, New Jersey. Geological society of America Bulletin, 67(5), 597-646. Scott, K. M., & Williams, R. P. (1978). Erosion and sediment yields in the Transverse Ranges, southern California. US Government Printing Office. Shakesby, R. A., Moody, J. A., Martin, D. A., & Robichaud, P. R. (2016). Synthesising empirical results to improve predictions of post-wildfire runoff and erosion response. International Journal of Wildland Fire, 26(3), 257-261. https://doi.org/10.1071/WF16021 Shuirman, G., & Slosson, J. E. (1992). Forensic engineering – Environmental case histories for engineers and geologists. Academic Press. Soulis, K. X., Generali, K. A., Papadaki, C., Theodoropoulos, C., & Psomiadis, E. (2021). Hydrological response of natural mediterranean watersheds to forest fires. Hydrology, 8(1), 15. https://doi.org/10.3390/hydrology8010015 Staley, D. M., Kean, J. W., Cannon, S. H., Schmidt, K. M., & Laber, J. L. (2013). Objective definition of rainfall intensity-duration thresholds for the initiation of post-fire debris flows in southern California. Landslides, 10, 547-562. https://doi.org/10.1007/s10346-012-0341-9 Staley, D. M., Negri, J. A., Kean, J. W., Laber, J. L., Tillery, A. C., & Youberg, A. M. (2017). Prediction of spatially explicit rainfall intensity–duration thresholds for post- fire debris-flow generation in the western United States. Geomorphology, 278, 149- 162. https://doi.org/10.1016/j.geomorph.2016.10.019 Stavros, E. N., Bloom, A. A., Brown, T., Coen, J., Dennison, P., Giglio, L., Green, R., Hinkley, E., Holden, Z., Hook, S., Johnson, W., Miller, M. E., Petersen, B., Quayle, B., Ramirez, C., Randerson, J., Schimel, D., Schroeder, W., Soja, A., & Tosca, M. (2017). The role of fire in the Earth System. NASA. Syphard, A. D., Radeloff, V. C., Keeley, J. E., Hawbaker, T. J., Clayton, M. K., Stewart, S. I., & Hammer, R. B. (2007). Human influence on California fire regimes. Ecological Applications, 17(5), 1388-1402. https://doi.org/10.1890/06-1128.1

63

Tarboton, D. G. (2003). Terrain analysis using digital elevation models in hydrology [Conference paper]. 23rd ESRI International Users Conference, San Diego, California. Taylor, A. H., Trouet, V., & Skinner, C. N. (2008). Climatic influences on fire regimes in montane forests of the southern Cascades, California, USA. International Journal of Wildland Fire, 17(1), 60-71. https://doi.org/10.1071/WF07033 Travis, B., Teal, M., & Gusman, J. (2012, May 20-24). Best methods and inherent limitations of bulked flow modeling with HEC-RAS [Paper presentation]. World Environmental and Water Resources Congress 2012, Albuquerque, New Mexico. USDA Forest Service. (2013). Burned area emergency response tools. https://forest.moscowfsl.wsu.edu/BAERTOOLS/ROADTRT/Peakflow/CN/suppleme nt.html USFS. (2020). Round table discussion. Holy Fire Research Symposium at United States Forest Service Pacific Southwest Research Station, Riverside, California Uyeda, K. A., Stow, D. A., & Riggan, P. J. (2015). Tracking MODIS NDVI time series to estimate fuel accumulation. Remote Sensing Letters, 6(8), 587-596. https://doi.org/10.1080/2150704X.2015.1063736 Van de Water, K. M., & Safford, H. D. (2011). A summary of fire frequency estimates for California vegetation before Euro-American settlement. Fire Ecology, 7, 266-58. https://doi.org/10.4996/fireecology.0703026 Vieira, D. C. S., Fernández, C., Vega, J. A., & Keizer, J. J. (2015). Does soil burn severity affect the post-fire runoff and interrill erosion response? A review based on meta- analysis of field rainfall simulation data. Journal of Hydrology, 523, 452-464. https://doi.org/10.1016/j.jhydrol.2015.01.071 Vo, V. D., & Kinoshita, A. M. (2020). Remote sensing of vegetation conditions after post- fire mulch treatments. Journal of Environmental Management, 260, 109993. https://doi.org/10.1016/j.jenvman.2019.109993 Wagenbrenner, J. W. (2013). Post-fire stream channel processes: Changes in runoff rates, sediment delivery across spatial scales, and mitigation effectiveness [Doctoral dissertation, Washington State University]. Washington State University Libraries. http://hdl.handle.net/2376/4931 Wells, W. G. (1981, January 25-31). Some effects of brushfires on erosion processes in coastal southern California: Erosion and sediment transport in Pacific Rim steeplands [Paper presentation]. Proceedings of the Christchurch Symposium, Christchurch, New Zealand. Wells, W. G. (1987). The effects of fire on the generation of debris flows in southern California. GSA Reviews in Engineering Geology, 7, 105-114. https://doi.org/10.1130/REG7-p105 Wenger, S. J., & Fowler, L. (2000). Protecting stream and river corridors: creating effective local riprian buffer ordinances. Carl Vinson Institute of Government, The University of Georgia.

64

WERT (Watershed Emergency Response Team). (2018a). Holy Fire— Watershed Emergency Response Team final report, CA-RRU-100160. ACWI. WERT (Watershed Emergency Response Team). (2018b). Woolsey and Hill Fires— Watershed Emergency Response Team final report, CA-VNC091023 and CA-VNC- 090993. ACWI. WERT (Watershed Emergency Response Team). (2019a). Cave Fire— Watershed Emergency Response Team final report, CA-LPF-002908. ACWI. WERT (Watershed Emergency Response Team). (2019b). Saddle Ridge Fire—Watershed Emergency Response Team final report, CA-LFD00001582. ACWI. Westerling, A. L., Hidalgo, H. G., Cayan, D. R., & Swetnam, T. W. (2006). Warming and earlier spring increase Western U.S. forest wildfire activity. Science, 313(5789), 940- 943. https://doi.org/10.1126/science.1128834 Wilder, B. A., Lancaster, J. T., Cafferata, P. H., Coe, D. B., Swanson, B. J., Lindsay, D. N., Short, W. R. & Kinoshita, A. M. (2021). An analytical solution for rapidly predicting post‐fire peak streamflow for small watersheds in southern California. Hydrological Processes, 35(1), e13976. https://doi.org/10.1002/hyp.13976 Williams, A. P., Abatzoglou, J. T., Gershunov, A., Guzman-Morales, J., Bishop, D. A., Balch, J. K., & Lettenmaier, D. P. (2019). Observed impacts of anthropogenic climate change on wildfire in California. Earth’s Future, 7(8), 892-910. https://doi.org/10.1029/2019EF001210 Wills, C. J., Perez, F. G., & Gutierrez, C. I. (2011). Susceptibility to deep-seated landslides in California. California Geological Survey Map Sheet. Wine, M. L., Cadol, D., & Makhnin, O. (2018). In ecoregions across western USA streamflow increases during post-wildfire recovery. Environmental Research Letters, 13(1), 014010. https://doi.org/10.1088/1748-9326/aa9c5a Wittenberg, L., Malkinson, D., Beeri, O., Halutzy, A., & Tesler, N. (2007). Spatial and temporal patterns of vegetation recovery following sequences of forest fires in a Mediterranean landscape, Mt. Carmel Israel. Catena, 71(1), 76-83. https://doi.org/10.1016/j.catena.2006.10.007 Wohlgemuth, P. (2016). Long-term hydrologic research on the San Dimas Experimental Forest, southern California: Lessons learned and future directions [Conference proceedings]. The Fifth Interagency Conference on Research in the Watersheds. North Charleston, SC, United States. Wohlgemuth, P. (2016, March). Long-term hydrologic research on the San Dimas Experimental Forest, southern California: Lessons learned and future directions. In C. E. Stringer, K. W. Krauss, & J. S. Latimer (Eds.), Headwaters to estuaries: Advances in watershed science and management-Proceedings of the Fifth Interagency Conference on Research in the Watersheds (Vol. 211; pp. 227–232). US Department of Agriculture Forest Service, Southern Research Station. World Meteorological Organization. (2012). Standardized precipitation index user guide (M. Svoboda, M. Hayes and D. Wood). Author.

65

APPENDIX A

SUPPLEMENTARY FIGURES FOR CHAPTER 2

66

Table A.1. List of Products used to Find or Derive Data (column 1), Data Type and Motivation (Column 2)

Product Derived Variable, Resolution, Method and Sources

10-m DEM from USGS Region, area, perimeter, circulatory ratio, elongation ratio, drainage density, longest flow path, slope along longest flow path (%), lowest point, highest point, slope aspect (North, East, South, West), stream order (Stream Drop Analysis), Area with slopes over 23° (total area), Area with slopes over 23° (% area), and mean basin slope (Horton, 1945; Miller, 1953; Schumm, 1956; Keeley & Keeley, 1981; Gabet, 2003; Tarboton, 2003; Cannon et al., 2004; Kinoshita & Hogue, 2011; Moody et al., 2013; Kinoshita et al., 2014)

USGS StreamStats Average annual precipitation, area covered by forest (%), and area covered by developed (urban) land (%) (Keeley & Keeley, 1981; Johansen et al., 2001; Cannon & DeGraff, 2009).

USDA SSURGO Hydrologic Soil Group (%), Pre-fire curve number, estimated post-fire curve number, ΔCN, pre-fire lag time, post-fire lag time, Δlag time, and weighted average soils erodibility factor (USDA Forest Service, 2013; Miller et al., 2003; Goodrich et al., 2005; Kinoshita et al. 2014)

California Geological Landslide susceptibility Classes 9-10 (% area) (Wills et al., 2011). Survey Landslide Susceptibility Maps

MTBS (Monitoring Fire year, unburned to low burn severity (%), moderate burn Trends in Burn Severity) severity (%), high burn severity (%), burned area (total area), burned area (%), Mean ΔNBR, and Max ΔNBR (Cannon et al., 2004; Moody et al., 2008; Miller et al., 2009; Kinoshita et al., 2014).

USGS Water Data and Time elapsed from end of fire to storm and Qpk-median (which is News Reports median of all annual high flows on record for each year) (MacDonald & Huffman, 2004).

NOAA/ RCFCWCD/ Total rainfall during post-fire storm event and Peak 1-hr rainfall County of Santa Barbara intensity during post-fire storm event (Moody & Martin, 2001; Public Works Goodrich et al., 2012).

67

Figure A.1. Relative importance for random forest calibration. Five (black) parameters were used for RF-5 regression.

68

Figure A.2. Random Forest model with 45 parameters (RF-45).

69

Figure A.3. Random Forest model with 5 most important parameters (RF-5).

70

APPENDIX B

SUPPLEMENTARY FIGURES FOR CHAPTER 3

71

Figure A.4. Control volume diagram for WY 2020 water balance calculation.

72

Figure A.5. ECOSTRESS PT-JPL daily evapotranspiration (ET) of 4-days before the fire on August 2, 2018 (a), SSEBop monthly ET 1-month before the fire in July 2018 converted to daily ET (b), ECOSTRESS approximately 1-year after the fire on August 17, 2019 (c), SSEBop approximately 1-year after the fire in August 2019 (d), ECOSTRESS approximately 2-years after the fire on October 3, 2020 (e), and SSEBop approximately 2-years after the fire in October 2020.

73

Figure A.6. Field photos from 2018 Holy Fire showing 1 year of revegetation.