Evaluating different machine learning algorithms for snow water equivalent prediction

The purpose of current study is to predict Snow Water Equivalent (SWE) in Sohrevard watershed, Iran, using different machine learning algorithms such as Bayesian Artificial Neural Network (BANN), Support Vector Machine (SVM), Cubist and Random Forest (RF) with Latin Hypercube Sampling (LHS). In this regard, nine geo-environmental variables—altitude, slope, eastness, profile curvature, plan curvature, solar radiation, Topographic Position Index (TPI), Topographic Wetness Index (TWI) and wind exposition index—were used as SWE influencing factors. Based on the results obtained from the error metrics, the RF algorithm (train and testing stages, r = 0.96 and 0.76; Root Mean Square Error (RMSE) = 2.54 and 5.46 cm; Mean Absolute Error (MAE) = 1.74 and 4.05 cm; Percent BIAS (PBIAS) = 0.4 and 2.3 respectively) was selected as the best model. Based on our findings, the highest amount of SWE was concentrated in the eastern part of the watershed. SWE modeling is a useful tool for optimal and integrated management of water resources.