Daily soil temperature prediction using hybrid deep learning and SHAP for sustainable soil management

The thermal conditions within soil layers represent essential information across diverse fields including sustainable agriculture, power generation, biological research, ecological studies, forest management, and thermal energy systems. Therefore, this study assessed prediction capabilities for subterranean thermal measurements using different advanced deep learning models at six distinct depth points across dual climate monitoring locations, Darbandikhan and Chamchamal within Iraq's Kurdistan territory. Furthermore, various configurations of the input variables were examined across seven distinct observational scenarios to identify the most significant predictive factors. Based on the results of this study, various deep learning models demonstrated optimal performance for soil temperature prediction at different depths across the two stations. At Chamchamal station, the hybrid deep learning model that combines bidirectional gated recurrent unit (BiGRU) and convolutional neural network (CNN), denoted as BiGRU-CNN achieved the best result for the 05 cm depth (RMSE = 1.298°C), while the hybrid model based on gated recurrent unit (GRU) and convolutional neural network (CNN), referred to as GRU–CNN yielded the best performance at 10 cm (RMSE = 1.333°C). BiLSTM-CNN provided the most accurate predictions at 20 cm (RMSE = 1.489°C), and BiGRU-CNN generated superior results at 30 cm (RMSE = 1.267°C). For deeper soil layers, GRU-CNN delivered the most accurate results at 50 cm (RMSE = 1.058°C), and LSTM-CNN performed the best at 100 cm (RMSE = 1.171°C). Additionally, at Darbandikhan station, BiLSTM-CNN generated the most accurate predictions at 50 cm depth (RMSE = 1.506°C). The results indicate that the proposed methodologies offer viable solutions for soil temperature forecasting when utilizing the meteorological input variables. These models demonstrate substantial potential as reliable tools for accurately estimating subsurface thermal conditions based on available parameters.