Field-scale mapping of soil available phosphorus in cropland via interpretable machine learning and multispectral remote sensing

Abstract

Soil available phosphorus (SAP) is a key indicator for assessing crop fertilization requirements. Accurately and quickly mapping the spatial distribution of SAP in cropland is crucial for guiding variable fertilization decisions. To explore a suitable model for estimating SAP at the field scale and enhance its credibility and interpretability, this study focused on multiple croplands at the Jianshan Farm in Heilongjiang Province. Soil samples were collected during the autumn bare soil period, and synchronous Sentinel-2 satellite imagery was acquired. A dataset was constructed by integrating spectral remote sensing factors with topographic variables. Random Forest Regression (RFR), Categorical Boosting (CatBoost), and Back Propagation Neural Network (BPNN) were employed to establish models for estimating SAP content in cropland. The zebra optimization algorithm (ZOA) was used to perform hyperparameter optimization and performance analysis on the best-performing model. On this basis, the Shapley additive explanations (SHAP) method was adopted to quantitatively analyze the contribution and impact of environmental variables on model predictions. Spatial distribution mapping of SAP content was conducted for both the experimental and newly added plots to evaluate the model's applicability. The results indicated that the ZOA-BPNN-G model, developed using raw bands, soil indices, and topographic variables, achieved the highest accuracy in estimating SAP (R²=0.6229, r = 0.8385, RMSE=8.4359 mg/kg, MAE=5.9816 mg/kg, MAPE=15.5123 %). Compared with the estimation model before optimization, R² increased by 9.55 %age points, and r increased by 7.91 %age points, while the values of other evaluation indicators decreased. The raw bands and soil indices are the primary explanatory variables for estimating SAP content, with key features including NDSI, B8A, BI, and B3. Interpretability analysis and uncertainty analysis improved the transparency and credibility of the model. The proposed estimation model achieved good SAP mapping results on the experimental plots and had a strong correspondence with the SAP measured data. The MAPE for the newly added plot reached 13.1288 %, confirming the application feasibility of the proposed method to the autumn bare soil in the study area. It can provide theoretical and technical support for the rapid detection of SAP content and the generation of variable fertilization prescription maps.