Remote sensing inversion and mapping of soil total nitrogen based on an enhanced spectral feature set and machine learning model

Background Soil total nitrogen (TN) is an important nutrient source for plant growth and one of the key indicators for evaluating soil quality. Rapid and accurate acquisition of soil TN information in farmland is important for cultivated land quality assessment and precision agricultural management. Purpose In this study, the area on the north bank of the Yellow River was selected as the study area. The aim was to improve the accuracy of regional soil TN estimation and mapping by introducing a new triangular spectral index and combining it with machine learning methods. Methods A total of 82 soil samples were collected, their hyperspectral data were measured, and soil TN contents were determined. A new triangular spectral index (SMI) was introduced to build an enhanced spectral feature set. Then, a soil TN estimation model was developed using the Random Forest (RF) algorithm to achieve regional-scale inversion and mapping of soil TN content. The performances of the RF, Extreme Gradient Boosting (XGBoost), and Support Vector Machine (SVM) models were compared using the ratio of performance to deviation (RPD), ratio of performance to interquartile distance (RPIQ), and root mean square error (RMSE). Results The results showed that the enhanced spectral feature set constructed with the SMI effectively improved the prediction accuracy of soil TN estimation. The RF model achieved the best performance, with the ratio of performance to deviation (RPD) increasing to 1.31, the ratio of performance to interquartile distance (RPIQ) increasing to 1.46, and root mean square error (RMSE) decreasing to 0.13. The Extreme Gradient Boosting (XGBoost) model ranked second, with RPD, RPIQ, and RMSE values of 1.25, 1.40, and 0.14, respectively. The Support Vector Machine (SVM) model showed the weakest estimation ability, with RPD, RPIQ, and RMSE values of 1.01, 1.14, and 0.17, respectively. Compared with the XGBoost model, the RF model increased RPD and RPIQ by 0.06 and reduced RMSE by 0.01, indicating better model fitting and prediction performance. The spatial distribution of soil TN showed that the study area did not have clear continuous high-value or low-value zones. Instead, it was mainly characterized by medium TN values, with high- and low-value patches distributed in a mosaic pattern. This indicates a certain degree of spatial heterogeneity in regional soil nutrient content. Conclusions By introducing the new triangular spectral index to construct an enhanced spectral feature set and combining it with the RF algorithm, this study produced a spatial distribution map of soil TN content on the north bank of the Yellow River. The results can help improve the accuracy of regional soil TN mapping and provide a reference for cultivated land quality assessment and precision agricultural management.