Assessment and Prediction of Soil Fertility in Urban Areas of the Loess Plateau Based on Machine Learning Methods

Abstract

The Loess Plateau, a vital ecological region in China, suffers from severe soil pollution and erosion. The soil fertility index (SFI) is a key indicator for assessing soil conditions, and understanding its spatial distribution and influencing factors is crucial for effective soil management. Machine learning methods, capable of analyzing complex and high-dimensional data, offer potential for large-scale SFI prediction. This study focuses on Lanzhou, a representative city on the Loess Plateau, using soil samples and the data of five key factors screened from environmental big data to train three machine learning models (random forest [RF], LightGBM, and XGBoost) for SFI prediction. The results show that all models effectively matched reference data trend, with XGBoost achieving the highest performance (R² > 0.81). Notably, normalized difference vegetation index (NDVI) and soil organic carbon density (SOCD) emerged as the dominant predictors, collectively contributing over 80% to SFI prediction accuracy. Predicted SFI values in Lanzhou ranged from 0.09 to 0.91, with medium and lower quality soils predominantly located in central and north-central regions, highlighting the need for soil quality improvement. This study provides a theoretical basis and scientific support for large-scale SFI prediction.