Daily soil temperature simulation at different depths in the Red River Basin: a long short-term memory approach

Soil temperature impacts a variety of biotic and abiotic processes such as carbon cycling, soil microbial activity, soil moisture dynamics, and agricultural management within ecosystems. In this study, we evaluated the performance of Long Short-Term Memory (LSTM) networks for modeling soil temperature at different depths (i.e., 10, 50, and 100 cm) by incorporating a set of climatic and soil-related variables, including air temperature, relative humidity, snow depth, solar radiation, and soil moisture for two stations in the Red River Basin, Fargo and Grand Forks, North Dakota. The SHapley Additive exPlanations (SHAP) framework was used to identify the key features that have a significant impact on soil temperature and the identified significant features were subsequently used to design and test the LSTM model. It was found that average air temperature is the most consequential feature affecting soil temperature at all depths examined. However, during colder periods, an increase in the significance of other factors such as snow depth and soil moisture at greater depths was observed. The performance of the LSTM model was assessed using the Nash-Sutcliffe Efficiency (NSE) and Kling-Gupta Efficiency (KGE). The LSTM model achieved high efficiency scores, with NSE values ranging from 0.977 to 0.994 and KGE values between 0.937 and 0.980 across all depths and both stations, confirming the model’s ability to accurately simulate soil temperature. However, we found limitations in the model’s ability to capture soil temperature variations during winter months, indicating challenges in modeling under frozen conditions. This study offers insights into the intricate dynamics that regulate soil temperature across different depths and offers a soil temperature simulation framework applicable to local agricultural, hydrological, and environmental studies.