Mapping of Soil pH Based on SVM-RFE Feature Selection Algorithm

Abstract

The explicit mapping of spatial soil pH is beneficial to evaluate the effects of land-use changes in soil quality. Digital soil mapping methods based on machine learning have been considered one effective way to predict the spatial distribution of soil parameters. However, selecting optimal environmental variables with an appropriate feature selection method is key work in digital mapping. In this study, we evaluated the performance of the support vector machine recursive feature elimination (SVM-RFE) feature selection methods with four common performance machine learning methods in predicting and mapping the spatial soil pH of one urban area in Fuzhou, China. Thirty environmental variables were collected from the 134 samples that covered the entire study area for the SVM-RFE feature selection. The results identified the five most critical environmental variables for soil pH value: mean annual temperature (MAT), slope, Topographic Wetness Index (TWI), modified soil-adjusted vegetation index (MSAVI), and Band5. Further, the SVM-RFE feature selection algorithm could effectively improve the model accuracy, and the extreme gradient boosting (XGBoost) model after SVM-RFE feature selection had the best prediction results (R2 = 0.68, MAE = 0.16, RMSE = 0.26). This paper combines the RFE-SVM feature selection with machine learning models to enable the fast and inexpensive mapping of soil pH, providing new ideas for predicting soil pH at small and medium scales, which will help with soil conservation and management in the region.

Keywords

• digital soil mapping
• spatial prediction
• environment variables
• feature selection
• machine learning