Abstract:

Soil erosion is a serious challenge that constrains agricultural production and sustainable development. Its types are diverse, and its driving factors are complex, making it difficult to predict its occurrence and development. This study focuses on  driving factors and occurrence characteristics of soil erosion, extracting a feature variable set from multi-source data, and selecting the optimal feature subset using the simulated annealing (SA) algorithm. A soil erosion rate (SER) prediction model is built using random forest regression (RFR), and spatial distribution maps of SER from 1990 to 2020 (at 5 a intervals) are generated. Finally, the spatiotemporal heterogeneity of SER is analyzed and discussed through correlation analysis and statistical methods. The results are as follows: The optimal feature subset consisted of climate, remote sensing, tasseled cap transformation, and topographic variables, including precipitation, vegetation and soil indices, surface moisture, and terrain metrics. The RFR model constructed based on these feature variables has a root mean square error of 0.048 mm/a and a coefficient of determination of 0.662. The SER from 1990 to 2020 fluctuated, mainly following a decrease-increase-decrease trend. The fluctuations in 2005 and 2020 were mainly influenced by increased annual precipitation. The SER fluctuated gently, with a median value ranging from 1.06 mm/a to 1.22 mm/a, primarily light  erosion, with relatively few areas experiencing moderate or severe erosion. The spatial distribution of SER showed a low-west and high-east pattern, exhibiting significant spatial autocorrelation and a clustered spatial distribution. Soil properties and topographic factors significantly influenced the SER, with notable spatial variations in certain local areas.

Key words: remote sensing data, soil erosion rate, machine learning, spatiotemporal heterogeneity analysis, prediction model