Spatial variability of soil geochemical elements using a novel ecological factor-incorporated interpretable machine...

Rapid monitoring of soil geochemical elements (SGEs) in farmlands is crucial for the sustainable development of coastal agriculture. Traditional environmental factors are relatively uniform on a regional scale, leading to low sensitivity and accuracy for detecting geochemical variations. We developed novel ecological factors (ecological quality and soil erosion) combined with a tree-structured Parzen estimator extreme gradient boosting (TPE-XGBoost) model to predict the spatial variability of SGEs (i.e., Co, Mn, Sr, V, Zr) in coastal saline-alkali land in China. The results of TPE optimization revealed that key hyperparameters, including lambda, gamma, colsample_bynode, and colsample_bytree, significantly influenced XGBoost model performance, and the prediction accuracy (R²) for all five geochemical elements exceeded 0.6 when using ecological factors. High concentrations of Co, V, and Zr were primarily found in the southeast, while Sr was distributed outside this region. Mn was mainly concentrated near the Yellow River Shapley additive explanations (SHAP) analysis indicated that climate and ecological factors were the strongest drivers of SGEs spatial variability. Structural equation modeling (SEM) results indicate indirect pathways: ecological quality influences temperature and precipitation, while soil erosion affects temperature, both of which significantly impact the concentrations of geochemical elements (Co, Sr, V, Zr). Moreover, the synergistic effect of ecological and climate factors promoted the absorption or decomposition of geochemical elements by crops, thereby modifying geochemical elements concentrations in soil. Therefore, low-cost ecological factors can predict SGEs spatial variability to support efficient pollution diagnosis in coastal agro-ecosystems.