Evaluating AI Models in SWE Prediction

Abstract

Snow Water Equivalent (SWE) is an essential metric for analyzing water supply and climate in specific areas. Over the years, various AI models have been used to predict SWE. These AI models allow us to monitor SWE from a distance; however, the accuracy of each type of model is still up for debate. In this study, we built and tested 4 different types of AI models to measure their accuracy in SWE prediction: a Graph Neural Network (GNN), a Transformer, a Long Short-Term Memory (LSTM) model, and a Random Forest (RF) model. Currently, GNN, which uses graph nodes to generate predictions, and Transformer, which uses self-attention to determine the importance of variables, are considered to be among the most cutting-edge models for SWE prediction. LSTM, which uses memory to learn prediction patterns, and RF, which uses decision trees, are considered to be older models. The models were trained on historical data sets from 2018 to 2021, which included variables such as precipitation, air temperature, and humidity. Model accuracy was measured based on Mean Squared Error (MSE) and the R² metric. Our current findings show that the GNN model had the lowest MSE and the highest R² overall, showing the highest accuracy. This suggests that models that can represent data as graphs can be more effective for SWE prediction.