Spatially adaptive estimation of multi-layer soil temperature at a daily time-step across China during 2010–2020

Abstract. Soil temperature (T_s) is critical in regulating agricultural production, ecosystem functions, hydrological cycling and climate dynamics. However, the inherent spatial and temporal heterogeneity of soil thermal regimes constitutes a persistent challenge in obtaining high-resolution, continuous gridded T_s datasets along vertical profiles. To address this issue, we propose a spatially adaptive layer-cascading Extreme Gradient Boosting (XGBoost) algorithm to generate daily multi-layer T_s data (0, 5, 10, 15, 20, and 40 cm) at a spatial resolution of 1 km in China from 2010 to 2020. The methodology dynamically partitions non-uniformly distributed measuring sites (2,093 sites across the country) to quadtrees and incorporates thermal coupling effects propagated between neighbor soil layers. Multi-source data, including satellite retrievals of land surface temperature and vegetation index, and ERA5 reanalysis climate variables were used as inputs. Independent tests demonstrated high robustness and accuracy of our model, with depth-specific values of coefficients of determination (R²) being 0.94~0.98 and root mean square errors (RMSE) values ranging 1.75~2.21K. It is noted the model’s performance was lower in summers and winters than in springs and autumns. Compared to existing global or regional T_s products, the dataset developed here is characterized by its fine spatio-temporal patterns and high reliability, enabling it to provide supports for precision agriculture, ecosystem modeling and understanding climate-land feedback. Free access to the dataset can be found at https://doi.org/10.11888/Terre.tpdc.302333 (Wang et al., 2025).