Depth-dependent predictability of soil temperature under Mediterranean climatic conditions: a frequency-based perspective

Soil-temperature predictability is inherently depth-dependent because atmospheric forcing is progressively attenuated and transformed into smoother, lower-frequency thermal dynamics with increasing depth. Under Mediterranean climatic conditions, where pronounced seasonality and strong surface forcing coexist, this depth-dependent transformation provides an informative basis for evaluating model behaviour. This study examines soil-temperature prediction at depths of 5, 10, 20, 50, and 100 cm using a multivariate dataset comprising 5,285 retained daily observations from the Mersin meteorological station on Türkiye’s eastern Mediterranean coast. Precipitation, air temperature, relative humidity, wind speed, and evaporation were used as predictors. Support vector regression, random forest, and LightGBM were evaluated using R², RMSE, and MAE. All three models performed strongly in the near-surface layers, with test R² values ranging from 0.948 to 0.954 at depths of 5–20 cm. Predictive performance declined with increasing depth as short-period variability weakened and the thermal signal became smoother and more seasonally structured. RF and LightGBM showed a greater numerical decline at 50–100 cm, whereas SVR achieved test R² values of 0.923 at 50 cm and 0.847 at 100 cm; however, inter-model differences remained modest and were interpreted descriptively. Lomb–Scargle spectral analysis, applied using the actual observation dates, indicated that short-period variability became relatively less pronounced with depth while seasonal-scale and low-frequency components remained prominent. Overall, the findings suggest that soil-temperature prediction should be interpreted in relation to depth-dependent temporal signal characteristics rather than as a search for a universally superior algorithm.