To achieve the objective of a “15 min living circle” for educational services, this study develops an integrated method for primary school site selection in Tianjin, China, by combining multi-source data and ensemble learning techniques. At a 500 m grid scale, a suitability prediction model was constructed based on the existing distribution of primary schools, utilizing Random Forest (RF) and Extreme Gradient Boosting (XGBoost) models. Comprehensive evaluation, feature importance analysis, and SHAP (SHapley Additive exPlanations) interpretation were conducted to ensure model reliability and interpretability. Spatial overlay analysis, incorporating population structure and the education supply–demand ratio, identified highly suitable areas for primary school construction. The results demonstrate: (1) RF and XGBoost achieved evaluation metrics exceeding 85%, outperforming traditional single models such as Logistic Regression, SVM, KNN, and CART. Validation against actual primary school distributions yielded accuracies of 84.70% and 92.41% for RF and XGBoost, respectively. (2) SHAP analysis identified population density, proximity to other educational institutions, and accessibility to transportation facilities as the most critical factors influencing site suitability. (3) Suitable areas for primary school construction are concentrated in central Tianjin and surrounding areas, including Baoping Street (Baodi District), Huaming Street (Dongli District), and Zhongbei Town (Xiqing District), among others, to meet high-quality educational service demands.

Selenium (Se) and cadmium (Cd) have contrasting implications for agricultural land quality. We integrated 25 unpaired rock samples and 149 soil records, including 106 paired soil–crop observations from Zhijin County, Guizhou, China, to characterise element distributions and evaluate static predictive models under spatially blocked cross-validation. Crop Se values below detection were treated as LOD/√2; models used 88 numerical Se records and 106 Cd records across five spatial KMeans blocks. Rock, soil and crop data showed element-specific concentration patterns, bioconcentration factors and soil associations. The highest pooled out-of-fold performance was obtained for Crop Se (R² = 0.831, RMSE = 0.220 log10 mg kg⁻¹), whereas Crop Cd showed weaker performance (R² = 0.233, RMSE = 0.495 log10 mg kg⁻¹). Crop Se robustness analyses showed sensitivity to censoring treatment and crop-group structure, and fold-contained within-crop residual prediction was weak (R² = 0.007). Fitted associations involved soil Se, soil Cd, pH, MgO, organic C and crop group, conditional on correlated environmental and management factors. These findings support crop-stratified sampling across representative geological and soil settings for Se monitoring, while Cd surveillance remains centred on direct crop measurement. The models provide a transparent basis for sampling design and targeted field verification.