High-Resolution Forecasting of Soil Thermal Regimes Using Different Deep Learning Frameworks under Climate Change

Soil temperature is a critical parameter influencing ecological and hydrological processes, yet its accurate projection under climate change remains challenging due to coarse-resolution climate models and complex soil-atmosphere interactions. This study develops a deep learning framework to downscale soil temperature (5 cm depth) in western Iran, under climate change scenarios. Using Kendall’s Tau statistic, we identified optimal predictors from 26 CanESM5 (CMIP6) variables, selecting those with strong monotonic relationships (τ > 0.4) with observed soil temperature. Four deep learning models—CNN, LSTM, GRU, and a hybrid CNN-LSTM—were evaluated for downscaling performance using historical data (1980–2014). The hybrid CNN-LSTM model outperformed others, achieving the highest accuracy (NSE > 89%, RMSE < 3.94°C) by capturing spatial and temporal dependencies in soil thermal dynamics. Validation (2015–2020) revealed regional climate patterns: western stations, more arid and warming-sensitive, aligned with SSP245/SSP370, while eastern stations, influenced by the Zagros Mountains, showed cooling and precipitation feedbacks favoring SSP119/SSP126. Future projections under SSP126, SSP245, and SSP585 scenarios indicated nonlinear soil temperature responses, with high-emission pathways (SSP585) causing initial cooling (-4.11°C by 2040) followed by accelerated warming (+ 2.09°C by 2100). Elevation played a critical role, with high-altitude stations exhibiting persistent warming (+ 4.02°C) contrary to cooling trends at lower elevations. The model’s robust validation (Kendall's τ = 0.72–0.80) confirms its reliability for climate projection studies, bridging the resolution gap between global models and local conditions to support agricultural and ecosystem management.