Second-dimension-enhanced spatial context-aware modeling for spatial prediction

Spatial prediction is a fundamental task in geographic information science and environmental resource assessment. However, under complex terrain and heterogeneous covariates, models relying only on point-scale attributes or a single global relationship often have difficulty representing spatial context and regional heterogeneity. This study proposes a second-dimension-enhanced spatial context-aware modeling framework (SDA-SCA) that integrates first-dimension covariates, neighborhood-based second-dimension features, covariate screening, spatially constrained stratification, and local model training. Using two independently constructed annual spatial-prediction cases based on surface solar radiation observations from 670 stations in China in 2022 and 2023, five modeling schemes—Ordinary Kriging, Geographically Weighted Random Forest (GWRF), first-dimension attribute modeling (FDA), second-dimension augmented modeling (SDA), and SDA-SCA—were evaluated under spatial cross-validation. SDA-SCA showed a consistent predictive advantage across the two independent annual cases and the three evaluated model families. Relative to FDA, SDA-SCA increased R2 by 0.11–0.16 in 2022 and 0.17–0.21 in 2023, while reducing RMSE by 12.4%–17.7% and 24.3%–33.7%, and MAE by 22.8%–35.5% and 31.5%–50.4%, respectively. Stage-wise ablation showed that the FDA-to-SDA step produced consistent but moderate improvements, whereas the SDA-to-SDA-SCA step provided larger and more stable gains. These results indicate that second-dimension features improve the representation of neighborhood context, and their predictive value is further strengthened when coupled with spatial stratification. SHAP analysis further confirmed the substantial contribution of second-dimension features to the integrated model. The proposed framework provides an extensible pathway for spatial prediction tasks requiring the joint representation of local attributes, neighborhood context, and spatial heterogeneity.