Incorporating an improved spectral feature set into random forest model to enhance inversion and spatial mapping of soil total phosphorus

Soil total phosphorus (P) is an essential macronutrient for plant growth and an important indicator for evaluating soil fertility and cultivated land quality. Rapid and accurate mapping of the spatial distribution of soil P in farmland is important for cultivated land protection and sustainable agricultural development. In this study, the Yellow River alluvial–proluvial plain was selected as the study area. A total of 82 topsoil samples were systematically collected, and soil P content and hyperspectral data were obtained. To address the limitation that traditional single band and two-dimensional spectral indices have difficulty capturing the weak spectral response of soil P, triangular spectral indices were introduced. An improved spectral feature set was then constructed by integrating single-band features, two-dimensional spectral indices, and triangular spectral indices. Feature selection was performed using Pearson correlation coefficients combined with the variance inflation factor. Three machine learning algorithms, namely random forest, support vector regression, and extreme gradient boosting, were compared to develop remote sensing inversion models for soil P content. The results showed that the improved spectral feature set with triangular spectral indices significantly improved model accuracy. The coefficient of determination (R 2 ) of the validation set increased from 0.36 to 0.56, while the root mean square error (RMSE) decreased from 0.12 to 0.10. Among the models, the random forest model based on the improved spectral feature set achieved the best performance, with a validation R 2 of 0.56, an RMSE of 0.10, and a ratio of residual prediction deviation of 1.56, indicating good generalization ability and stability. The spatial distribution map of soil P generated by the optimal model showed that soil P content in the study area was generally at a medium-to-high level, ranging from 0.88 to 1.05 g/kg. No large continuous clusters of high or low values were observed. Instead, the distribution was dominated by medium-to-high values, with scattered patches of high and low values, indicating a certain degree of spatial heterogeneity in soil nutrient content. The hyperspectral inversion method for soil P proposed in this study, which combines the improved spectral feature set with the random forest algorithm, can provide scientific evidence and data support for land-use management, cultivated land protection, and precision fertilization in farmland.