Intelligent processing of SmallSat Imagery to improve snow mapping in complex and forested terrain

Abstract:

Improving high resolution (m scale) mapping of snow covered areas in complex and forested terrain is crucial to understanding responses of species and water systems to climate change. Planet Labs, Inc. (Planet) is a promising new source of commercial high-resolution imagery that can be used in environmental science, as it has both high spatial (0.7-3.0 m) and temporal (1-2 day) resolution. However, deriving snow cover areas from Planet imagery using traditional radiometric techniques (such as the Normalized Difference Snow Index, NDSI) has limitations due to the near-infrared band placement that is too close to the visible bands to fully exploit the difference in reflectance. Recent research (Cannistra et al., 2019) has demonstrated that snow cover can be successfully mapped from Planet data using a machine learning approach based on convolutional neural networks (CNN) using only 4-band (R, G, B and NIR) reflectance data as input. Here, we demonstrate that the performance of the Cannistra et al., 2019) CNN model for snow mapping can be improved by using additional input data such as vegetation, elevation, slope and aspect. We train the CNN model in the Tuolumne River watershed, California using lidar-derived snow observations, and test it over the Gunnison River watershed, Colorado. We analyze how the model performance varies as a function of training datasets paying particular attention to snow mapping performance across areas with sparse trees and forest gaps.