Estimation of Mountain Glacier Retreat from Landsat Images
Armin Schwartzman Division of Biostatistics, UC San Diego
June 2016 Mountain Glacier Retreat
Muir Glacier, Alaska
1941 2004
Qori Kalis Glacier, Perú
1978 2004 2/35 Landsat Images
N
N
Franz Joseph Glacier
Southern New Zealand, Feb 2009 3/35 Landsat Images over Time
4/35 Image Analysis of Glaciers • The scientific problem: – Estimate trend of mountain glacier retreat • Motivation: – Climate change indicators – People depend on glacial melt water (Andes, Himalayas, Rockies) – Ground measurements for (mostly) Alps only • Proposal: – Analysis of Landsat images – They are free!
5/35 Difficulties • Difficulties – Clouds, debris, snow – Shadows from adjacent mountains – Landsat 7 defects – Irregular shapes • Standard approach – Segmentation of glacial surface – Carefully delineate glacier boundary – Requires substantial manual input – Accurate but limited
6/35 Our Approach • Goal: – Automatic processing and analysis – Scale up to hundreds of glaciers • Approach: – Estimate glacial flowline – Extract intensity profile along flowline – Use time/space functional data techniques to estimate trend in terminus location
7/35 Gorner Glacier 30 m resolution 30
8/35 Processing and Analysis Pipeline
Data Pre- Image Estimate processing Analysis Terminus
1. Geographical 1. Classify cloudy 1. Spatial smoothing bounding box images 2. Path Search 2. Select frequency 2. Estimate flow line Algorithm band 3. Extract intensity 3. Temporal 3. Generate image profile smoothing stack 4. Obtain DEM
9/35 Processing and Analysis Pipeline
Data Pre- Image Estimate processing Analysis Terminus
1. Geographical 1. Classify cloudy 1. Spatial smoothing bounding box images 2. Path Search 2. Select frequency 2. Estimate flow line Algorithm band 3. Extract intensity 3. Temporal 3. Generate image profile smoothing stack 4. Obtain DEM
10/35 Data Pre-Processing
Bounding box
Global Land Ice Measure- Digital elevation Landsat ments From model (DEM) images Space Database
11/35 Landsat Frequency Bands
B61
Normalized Difference B20 B50 NDSI Snow Index: B20 B50 12/35 Landsat Frequency Bands
Franz Josef glacier B20 (Optical, 30 m) B61 (Thermal, 60 m) New Zealand
Gorner glacier B20 (Optical, 30 m) NDSI (Processed, 30 m) Switzerland
13/35 Processing and Analysis Pipeline
Data Pre- Image Estimate processing Analysis Terminus
1. Geographical 1. Classify cloudy 1. Spatial smoothing bounding box images 2. Path Search 2. Select frequency 2. Estimate flow line Algorithm band 3. Extract intensity 3. Temporal 3. Generate image profile smoothing stack 4. Obtain DEM
14/35 Classify Cloudy Images Clear Cloudy
15/35 Estimate glacial flowline
Digital elevation Starting model (DEM) point
Smoothing
16/35 Terra Flow Algorithm • Existing method for tracking rivers: – Move to lowest 8-connected neighbor
Flow goes up at step 3 Flow goes left at step 3 and reaches boundary and stops at local minimum 17/35 Idea: Follow the Gradient • Let f(x,y) be a surface, and write the
flowline as a curve (gx(t), gy(t)) parametrized by t. • The flowline should satisfy