Towards ML-driven Snow Water Equivalence Prediction

Abstract

With recent trends in climate change, access to water derived from snowpack runoff is decreasing, affecting the millions of people whose livelihoods depend on it. While scientists often have historical Snow Water Equivalent (SWE) data, it is neither easy to estimate for the casual observer, nor is it accessible. We look to solve this through ML-driven SWE predictions, which can be accessed by users for free using a mobile app. By applying real-time climate data pulled from climate monitoring satellite projects such as Sentinel-1 and 2, MODIS, and gridMET, and ground-based measurements via historical SNOTEL data, we developed a dataset for Machine Learning model development. We implemented multiple Machine Learning models such as Random Forest and XGBoost, along with several deep learning algorithms such as GhostNet, LSTM, and GRU. These models were built to predict SWE at any given location, which had varying degrees of success. We found that XGBoost performed the best, achieving a mean squared error of 2.886 and an r^2 of 0.981. Our GRU model performed second-best with a mean squared error of 6.989 training over 200 epochs. Random forest performed significantly worse with a mean squared error of 45.583 and an r^2 of 0.698. In addition, we created numerous station-specific LSTM models that used solely historical SWE data to predict future SWE in order to reduce processing times on the mobile app. These models were significantly quicker given the lack of a need to collect real-time data from numerous sources, cutting down on the waiting time for mobile users. We then use our Flask server to deploy our models into service, providing accurate, real-time predictions of snow water equivalent.