Which bands should I use for mapping snow over complex mountain terrain? Data selection using machine learning models and VHR satellite imagery

Abstract:

Observing seasonal snow processes has entered a new age with commercial satellites delivering on-demand, sub-meter-scale panchromatic (PAN) and meter-scale multispectral (MS) imagery. Fine spatial and radiometric resolution combined with accurate image geolocation offer precise and accurate measurements of snow-covered area (SCA) in regions of complex land cover and terrain. These sensors can resolve snow between individual trees for direct SCA measurements without needing to account for large mixed pixels or fractional snow-covered area (fSCA).

While products from Planet SkySat and Airbus Pléaides constellations provide 4-band MS imagery (visible and NIR), MAXAR’s WorldView-3 (WV-3) satellite provides 8-band MS (397–1039 nm at 1.2 m) and 8-band SWIR (1184–2373 nm at 3.7 m). The WV-3 bands allow for calculation of the traditional Normalized Difference Snow Index (NDSI), which is useful for differentiating snow from clouds.

Using varying combinations of WorldView-3’s spectral bands, indices, and ancillary data (e.g., DEM derivatives), we built a collection of Random Forest models for land cover classification of snow, vegetation, exposed rock, surface water, and clouds over snow monitoring sites in the Western U.S. From these classification maps, we created VHR snow-covered area products and compared them with existing co-located coarser-resolution fSCA products (e.g.Landsat-8 at 30 m, and MODIS at 250 m).

Devising flexible, adaptable and accurate models is important for leveraging other available satellite imagery with more limited bands. To assess the relative importance of the 17 WV-3 bands, band difference indices, and DEM derivatives, we conducted systematic permutation tests during model development.

Our preliminary results show that the NIR 2 band (857–1039 nm) and band difference indices are most important for model construction and the resulting land cover classification. Comparisons between fSCA from coarse-resolution imagery and our SCA measurements from VHR imagery show a 5-20% difference, emphasizing the importance of image resolution for accurate SCA mapping.