Authors: Mousumi Ghosh, Aatish Anshuman, Mukesh Kumar – University of Alabama
Title: Influence of loss function on prediction accuracy of soil temperature in cold regions
Abstract: The importance of soil temperature on hydrological processes, agricultural water management, and land-atmosphere interactions is well recognized, as it is one of the critical variables affecting the physical, biological, and chemical processes in an ecosystem. Despite its significance, long term and reliable datasets of soil temperature remain limited. Numerous soil temperature models, ranging from physically based to empirical have been developed to predict soil temperature at ungauged locations. The efficacy of these methods, especially in cold settings, where the presence of snowpack inhibits the interaction of surface air temperature and solar radiation with the soil temperature, remains challenging. Using two machine learning models, long short-term memory (LSTM) and Artificial neural network (ANN) models, this study assesses the influence of the choice of loss function on model performance, during both snow-free and snow-covered cold periods. This study provides valuable insights into the selection of loss function for accurate prediction of soil temperature and other similar hydrometeorological variables.