Mapping key soil properties in low relief areas using integrated machine learning and geostatistics

• The optimal variables influencing the distribution of soil properties were explored. • The RF-RK is the best model when spatial autocorrelation is strong. • The study provides a method for mapping soil properties in low relief areas. Digital soil mapping based on the soil-landscape model can predict soil information using readily available environmental covariates such as topography and vegetation. However, its application in low-relief areas where topographic factors are relatively uniform and vegetation conditions are similar, is challenging. Meanwhile, geostatistical models often have better performance in low relief areas compared to topographically complex areas. Therefore, we hypothesized that the method of combining geostatistical modeling with soil-landscape modelling can achieve higher prediction accuracy. We comprehensively selected multiple environmental covariates suitable for the study area and compared four models: Inverse Distance Weighting (IDW), Ordinary Kriging (OK), Random Forest (RF), and Random Forest Regression Kriging (RF-RK). These models were used to predict six soil properties (clay, silt, and sand contents, pH, cation exchange capacity, and soil organic matter content) in the study area and evaluate their prediction accuracies. The results were as follows: (1) When the spatial autocorrelation of soil property data was weak, the RF model, which does not consider spatial autocorrelation, yielded more accurate predictions for clay content and soil pH, with R 2 of 0.62 and 0.30, and a NRMSE of 0.46 and 0.08, respectively. (2) When the spatial autocorrelation was strong, the RF-RK model, which accounts for both the soil-environment relationship and spatial autocorrelation, provided more accurate results for sand and silt contents, cation exchange capacity, and soil organic matter content. The RF-RK model achieved R 2 values of 0.81, 0.82, 0.79, and 0.69, and NRMSE values of 0.14, 0.09, 0.14, and 0.26, respectively. (3) Soil type and distance from the Yangtze River were the most important environmental variables explaining the spatial distribution of soil properties. The spatial heterogeneity of soil type and the geographical influence of the Yangtze River explained soil property variations better than other variables. This study highlights the potential of the integrated modeling approach for digital soil mapping in low relief areas. High-precision spatial maps of key soil attributes in the study area can provide critical data support for land planning and sustainable agricultural development.