Amin Sharififar, Alex McBratney, Budiman Minasny
Soil temperature is a key regulator of biogeochemical processes, plant growth, and soil–atmosphere interactions, yet spatially explicit estimates of subsurface temperature is limited at large spatial scales. This study evaluates the performance of a physics-based mechanistic model for mapping annual average soil temperature (ST) across Australia and compares its predictions with two machine-learning (ML) approaches, Support Vector Machine and Extreme Gradient Boosting. The mechanistic model is based on a steady-state analytical solution of soil heat conduction to estimate ST at multiple depths. Model predictions were evaluated using soil temperature observations and compared with ML models trained on a broad set of environmental covariates. The mechanistic model achieved predictive performance comparable to the ML approaches, with an average R 2 of 0.94 and RMSE values between 5.5 and 6.0 °C, while providing better physical interpretability and spatial consistency. The resulting maps reveal a clear continental-scale gradient in soil temperature, with northern Australia exhibiting substantially warmer soils than southern coastal regions. Although surface air temperature strongly influences ST patterns, soil thermal properties, particularly thermal conductivity, volumetric heat capacity, and soil moisture, modulate both vertical and lateral temperature variations. On average, annual soil temperature at 1 m depth was approximately 4 °C higher than mean annual surface air temperature, indicating the buffering effect of soil thermal stability. The results suggest that simplified mechanistic models with sensor calibrated approximation can provide annual mean soil temperature patterns at continental scales. Combining physically based approaches with machine-learning methods may further improve soil temperature prediction and support large-scale assessments of ecosystem processes, soil carbon dynamics, and land–atmosphere interactions.