Estimating Fractional Cover of Tundra Vegetation at Multiple Scales Using
Total Page:16
File Type:pdf, Size:1020Kb
Remote Sensing of Environment 224 (2019) 119–132 Contents lists available at ScienceDirect Remote Sensing of Environment journal homepage: www.elsevier.com/locate/rse Estimating fractional cover of tundra vegetation at multiple scales using T unmanned aerial systems and optical satellite data ⁎ Henri Riihimäkia, , Miska Luotoa, Janne Heiskanena,b a Department of Geosciences and Geography, University of Helsinki, 00014, PO Box 64, Helsinki, Finland b Institute for Atmospheric and Earth System Research, Faculty of Science, University of Helsinki, Finland ARTICLE INFO ABSTRACT Keywords: Fractional cover of green vegetation (FCover) is a key variable when observing Arctic vegetation under a Unmanned aerial vehicles changing climate. Vegetation changes over large areas are traditionally monitored by linking plot-scale mea- UAV surements to satellite data. However, integrating field and satellite data is not straightforward. Typically, the Drones satellite data are at a much coarser scale in comparison to field measurements. Here, we studied how Unmanned Modifiable Area Unit Problem (MAUP) Aerial Systems (UASs) can be used to bridge this gap. We covered three 250 m × 250 m sites in Fennoscandian Upscaling tundra with varying productivity and FCover, ranging from barren vegetation to shrub tundra. The UAS sites Resolution Arctic were then used to train satellite data-based FCover models. High-latitude First, we created a binary vegetation classification (absent, present) by using UAS-derived RGB-orthomosaics Monitoring and logistic regression. Secondly, we used the classification to calculate FCover to Planet CubeSat (3m), Sentinel-2A MSI (10 m, 20 m), and Landsat 8 OLI (30 m) grids, and examined how well FCover is explained by various spectral vegetation indices (VI) derived from satellite data. The overall classification accuracies for the UAS sites were ≥90%. The UAS-FCover were strongly related to the tested VIs (D2 89% at best). The explained deviance was generally higher for coarser resolution data, in- dicating that the effect of data resolution should be taken into account when comparing results from different sensors. VIs based on red-edge (at 740 nm, 783 nm), or near-infrared and shortwave infrared (SWIR) had the highest performance. We recommend wider inspection of red-edge and SWIR bands for future Arctic vegetation research. Our results demonstrate that UASs can be used for observing FCover at multiple scales. Individual UAS sites can serve as focus areas, which provide information at the finest resolution (e.g. individual plants), whereas a sample of several UAS sites can be used to train satellite data and examine vegetation over larger extents. 1. Introduction feedbacks, for example surface albedo and temperatures (Blok et al., 2011; Farbrot et al., 2013). Due to rapidly advancing climate change, there is a great need for Satellite-based spectral vegetation indices (VIs) are known to be monitoring Arctic vegetation. Substantial changes in vegetation cover sensitive to FCover and are commonly used to examine vegetation and composition have been observed in many areas across the Arctic change (Carlson and Ripley, 1997; Epstein et al., 2012; Jia et al., 2003; (e.g. Myers-Smith et al., 2011; Sturm et al., 2001). One of the key Tucker et al., 2001; Zeng et al., 2003). It is well recognized that more variables characterizing Arctic vegetation is the fractional cover of effort is needed to evaluate circumpolar VI(Guay et al., 2014). Col- green vegetation (hereafter FCover). Monitoring FCover can reveal the lecting ground reference data for coarse resolution data is challenging dynamics of vegetation expansion or loss (Epstein et al., 2013; Lara due to differences in scale between field observations and satellite data et al., 2018; Phoenix and Bjerke, 2016). These changes can be related to (Epstein et al., 2012; Walker et al., 2016). Vegetation survey plot size disturbance caused by herbivores, extreme weather events or earth typically varies from 1 to 104 m2 (Walker et al., 2016), but even smaller surface processes, or recovery from earlier disturbance (Virtanen et al., plots have been used. For example, Liu and Treitz (2016) used a subplot 2010; Phoenix and Bjerke, 2016). Furthermore, vegetation is crucially size of 0.125 m2 when collecting reference data for FCover. Obtaining linked to permafrost dynamics, and earth surface-atmosphere reference data even for a single pixel usually requires sampling and ⁎ Corresponding author. E-mail address: [email protected] (H. Riihimäki). https://doi.org/10.1016/j.rse.2019.01.030 Received 11 May 2018; Received in revised form 24 January 2019; Accepted 26 January 2019 0034-4257/ © 2019 The Authors. Published by Elsevier Inc. This is an open access article under the CC BY license (http://creativecommons.org/licenses/BY/4.0/). H. Riihimäki, et al. Remote Sensing of Environment 224 (2019) 119–132 Fig. 1. Scales of field observation and commonly used remote sensing data. UAS-assisted ground reference data collection can be used to bridge the gapbetweenplot data and medium and coarse resolution satellite data. multiple field observations, i.e. a plot consisting of several point ob- to vegetation monitoring remains largely unexplored. servations or subplots. Thus, novel methods are needed in order to Thus far, several studies have examined Arctic FCover based on dif- bridge this gap between scales. ferent remote sensing data sources (e.g. Laidler et al., 2008; Liu et al., During the past decade, advances in both the development of un- 2017; Liu and Treitz, 2016). Laidler et al. (2008) investigated the re- manned aerial systems (UASs) and photogrammetric methods have lationship between FCover and Normalized Difference Vegetation Index opened new avenues for many fields of research (Anderson and Gaston, (NDVI), and its relationship derived from three data sources; field 2013; Colomina and Molina, 2014; Cruzan et al., 2016; Pajares, 2015). spectrometer, IKONOS, and Landsat 7 ETM+ satellite data. The study UASs can be used as personal remote sensing systems, which can map showed a strong linear relationship between FCover and NDVI for all of the environment with unprecedented accuracy over large areas (Fig. 1). the instruments (R2 = 0.72–0.78). Liu and Treitz (2016) modelled Arctic Several studies have demonstrated that UASs are excellent tools for FCover with WorldView-2-derived VIs, and found similar results mapping and monitoring different vegetation properties (Dandois and (R2 = 0.74–0.81). Liu et al. (2017) modelled FCover for five different Ellis, 2013; Dunford et al., 2009; Wallace et al., 2016). Furthermore, Arctic vegetation types with hyperspectral Hyperion and multispectral UAS-based vegetation classification studies have been successfully WorldView-3 VIs, and found that narrowband indices performed better executed in different environments (Ahmed et al., 2017; Cruzan et al., than broadband indices. The optimal bands were located at the absorp- 2016; Laliberte et al., 2010), including the Arctic (Fraser et al., 2016; tion areas of leaf pigments (681.20 nm; red), water (1174.77 nm and Juszak et al., 2017; Palace et al., 2018). 1184.87 nm; near-infrared [NIR]), cellulose (2072.65 nm; shortwave in- A completely new set of scales is now available for analysis since the frared [SWIR]) and lignin (2102.94 nm; SWIR). In addition, the red-edge emergence of UASs (Anderson and Gaston, 2013; Tømmervik et al., slope (721.90 nm and 732.07 nm) was important for modelling FCover. 2014). Traditional Earth observation is limited by the predefined scale Nevertheless, the application of Landsat 8, Sentinel-2 and Planet of the observations, as the only scales available for analysis are those in CubeSat satellites for FCover modelling remains untested. The Sentinel- which the instruments record the data – regardless of whether these 2 and Planet CubeSat constellations could advance Arctic observation scales are appropriate for the analysis (Dark and Bram, 2007). Because due to their higher revisit time and resolution compared to the Landsat of their low operating altitude, UASs can produce ultrahigh-resolution series. Since the availability of cloud-free imagery in the Arctic is low data, which can be upscaled (i.e. aggregated) to a desired resolution. compared to other regions (Wulder et al., 2015), more frequent ob- The data can be matched with global-coverage satellite data, enabling servation can greatly advance Arctic monitoring. Sentinel-2 offers ad- FCover estimation at multiple scales (Fig. 1). ditional bands, also within the red-edge, which may improve the re- Most of the pan-Arctic vegetation studies are based on coarse-re- mote sensing of vegetation attributes, including those related to FCover solution satellite data, such as Advanced Very-High Resolution (Liu et al., 2017). On the other hand, the Landsat series offers a unique Radiometer (AVHRR, 1.09 km resolution at nadir) or Moderate time series extending back to the 1970s, which is essential for long-term Resolution Imaging Spectrometer (MODIS, resolution ≥ 250 m) change monitoring (Wulder et al., 2012). (Epstein et al., 2013; Guay et al., 2014). However, gradual changes in In this study, we demonstrate the utilization of UASs for estimating vegetation are hard to detect from coarse-resolution data due to land- FCover at multiple scales. Our main motivation was to improve the scape heterogeneity and lack of pure pixels (Jia et al., 2009). Medium- integration of field and satellite data, and to advance the monitoring resolution satellites (e.g. Landsat, Sentinel-2, SPOT) offer much higher methods of Arctic vegetation by using modern remote sensing methods. resolution to reveal more of the fine-scale variability of the Arctic ve- Our first aim was to create and test a simple workflow for estimating getation (Ju and Masek, 2016; Lara et al., 2018). Sentinel-2 and Landsat FCover from UAS-derived RGB-orthomosaics created by low-cost, easy- data are particularly relevant for vegetation monitoring, since both are to-operate UASs. Orthomosaics were classified as presence or absence freely distributed and have global coverage.