Assessment of soil fertility degradation affected by mining disturbance and land use in a coalfield via machine learning

The degradation of soil fertility in mining areas poses great risks to agricultural production and the ecological environment, and has increasingly become a worldwide concern. In this study, soil assessments were conducted to evaluate the spatial and temporal variations of soil fertility indicators and characteristics of soil fertility degradation under different types of land use and mining disturbances in a coalfield on the Loess Plateau of China, where soil fertility degradation caused by mining activities has become a serious environmental issue. Soil samples (depth: 0–20 cm) were collected twice from the same 50 sampling points in 2017 and 2019. The sampling points covered three land use types (cropland, shrubland, and grassland) and three years of mining disturbance (2011, 2013, and 2016). Soil organic matter (SOM), total nitrogen (TN), soil-available phosphorus, soil-available potassium, and the fine soil particles in topsoils were measured for each sample. The spatial distributions of the properties and degradation of soil fertility were analysed using kriging interpolation, and the degree to which fertility degraded was analysed via density-based spatial clustering of applications with noise (DBSCAN) and validated using SoftMax and random forest algorithms. The study revealed that the intensity of the degradation of soil fertility could be classified into three clusters (i.e. severely degraded, moderately degraded, and slightly degraded), as indicated by the DBSCAN results, and based on the variation in soil fertility indicators. Validation using random forest and SoftMax suggested that the accuracy of clustering was over 95%. Land use types and coal mining years significantly affected the degree of degradation, and total nitrogen and soil organic matter had the most noticeable impacts on the classification of soil fertility degradation.