Deep learning prediction of multi-depth soil temperature using climatic variables in eastern Hungary

Abstract

Soil temperature is a critical soil-health indicator that regulates biogeochemical processes, water balance, and carbon cycling under a changing climate. Accurate prediction of soil temperature enhances our ability to manage soil functions, sustain crop production, and mitigate climate-related risks. In this study, along with statistical exploration, we applied three predictive approaches: linear regression, random forest (RF), and a deep neural network (DNN) to predict soil temperature at depths of 5 and 50cm in Central Europe, specifically eastern Hungary. General linear models explained a very high proportion of soil-temperature variability at both depths, with R2 =0.964 for ST5 and R2 =0.913 for ST50. Minimum air temperature was the strongest climatic predictor at both depths, followed by maximum air temperature and solar radiation, whereas wind speed had a significant negative effect and relative humidity was not significant. Season also significantly affected soil temperature during 2017–2025, while the Group ×Season interaction was not significant at either depth. Although RF achieved the strongest training fit, DNN showed the most consistent validation and independent test performance. Under temporal testing at 5 cm, DNN achieved R2= 0.954, RMSE= 1.777 °C, and MAE= 1.368 °C. When applied at 50cm, DNN retained good predictive performance, with R2 =0.870, RMSE =2.416 °C, and MAE= 1.850 °C. These findings highlight soil temperature as a sensitive indicator of environmental change and demonstrate the value of deep learning for soil-health monitoring. The results provide a methodological and interpretive framework for integrating predictive modelling into sustainable soil and environmental management in climate-vulnerable regions.