Colorado Front Range · Sentinel-1 · 2021–2024
How avalanche zones were identified from satellite SAR imagery
Sentinel-1 C-band (5.6 cm wavelength) SAR imagery was collected via Google Earth Engine for the Colorado Front Range. IW mode imagery in VV and VH polarizations was filtered to descending orbit passes, which provide optimal illumination geometry for east-facing mountain slopes.
A focal mean speckle filter (5×5 kernel) was applied to VV and VH bands only the incidence angle band was never smoothed. Backscatter was normalized to gamma-naught (γ⁰) using local incidence angle from SRTM DEM: γ⁰ = σ⁰ − 10·log₁₀(cos θ). This makes backscatter comparable across varying slope orientations.
A per-pixel summer baseline (July–September) was computed as the median and Median Absolute Deviation (MAD) of gamma-naught values. MAD was chosen over standard deviation for its resistance to outliers from summer storm events. Each winter image was scored as: z = (γ⁰_winter − median) / MAD, with a MAD floor of 0.5 dB to prevent division instability in very stable pixels.
A weighted combined z-score was computed per pixel: VV_z × 0.50 + VH_z × 0.35 + ratio_z × 0.15. VV carries most of the avalanche debris signal (surface roughness), while VH adds sensitivity to debris texture and the VV/VH ratio discriminates when channels change asymmetrically. The per-season anomaly was taken as the pixel-wise maximum z-score across all winter images.
Pixels with combined z-score ≥ 3.0 were flagged as candidates. Two static masks were applied: a terrain mask (slope 20–70°, all aspects retained including south-facing slopes) and a JRC Global Surface Water mask (occurrence < 50%) to suppress false positives from freeze-thaw cycles on lakes and reservoirs. Note: a Hansen forest canopy mask was not applied in this iteration, meaning detections in areas with > 30% canopy cover may include false positives from snow loading and unloading on forest canopies.
Candidate rasters were vectorized to polygons using connected-components analysis at 20 m resolution. Each polygon was attributed with area, mean/max z-score, mean slope, aspect, and centroid coordinates. Confidence was assigned by combined z-score threshold: polygons with a peak z-score of 3.0–6.0 were classified as lower confidence; polygons with a peak z-score > 6.0 were classified as higher confidence. Cross-season hotspots represent polygons detected in two or more seasons within 100 m of each other.