Downscaling Daily Remotely Sensed Snow Cover Fraction Based on a Two Stage Machine Learning Model

Estimation of snow cover in remote areas is critical to water resource management and understanding long term climate change signals. However, direct in situ measurements are difficult or impossible to obtain in such mountainous terrain, and we must rely on remote sensing estimates. This research develops a machine learning downscaling algorithm that statistically downscales MODIS estimates of fractional snow covered area (fSCA) over the Southern Sierra Nevada Mountain range to the spatial resolution of Landsat imagery. Landsat estimates of fSCA follow directly from a spectral mixture analysis model with adjustments for canopy covered area where not all snow is viewable. MODIS data is also adjusted for canopy and is gap filled to account for cloud cover and weighted by sensor view angle to account for off-nadir viewing. Daily, 500 m MODIS estimates are produced for the period 2001 to 2018. Coincident Landsat estimates can suffer from cloud cover, and are only available at 16 day intervals, but at higher spatial resolution of 30 m. We propose a two-stage random forest algorithm that incorporates both MODIS and physiographic information to predict daily, 30 m fSCA. The first stage random forest is a classification step that predicts a class of zero, one hundred percent, or non-boundary values, while conditionally on a non-boundary classification, the second random forest predicts values between zero and one hundred percent fSCA. We also consider relationships that vary because of saturation with Landsat 5 in visible bands, and the fact that canopy adjusted values are not direct observations. Cross-validation results are presented, and the approach shows low bias and good predictive ability over the study domain. We explore cross-validation through a range of training data from 1% to 10% of pixels in available imagery to test efficiency and accuracy. In testing, the proposed two-stage random forest approach outperforms both multinomial logistic regression and single random forest downscaled predictions.