Identification of potentially contaminated areas of soil microplastic based on machine learning: A case study in...

Soil microplastic (MP) pollution has recently become increasingly aggravated, with severe consequences being generated. Understanding the spatial distribution characteristics of soil MPs is an important prerequisite for protecting and controlling soil pollution. However, determining the spatial distribution of soil MPs through a large number of soil field sampling and laboratory test analyses is unrealistic. In this study, we compared the accuracy and applicability of different machine learning models for predicting the spatial distribution of soil MPs. The support vector machine regression model with radial basis function (RBF) as kernel function (SVR-RBF) has a high prediction accuracy (R² = 0.8934). Among the six ensemble models, random forest (R² = 0.9007) could better explain the significance of source and sink factors affecting the occurrence of soil MPs. Soil texture, population density, and MPs point of interest (MPs-POI) were the main source-sink factors affecting the occurrence of soil MPs. Furthermore, the accumulation of MPs in soil was significantly affected by human activity. The spatial distribution map of soil MP pollution in the study area was drawn based on the bivariate local Moran's I model of soil MP pollution and the normalized difference vegetation index (NDVI) variation trend. A total of 48.74 km² of soil was in an area of serious MP pollution, mainly concentrated in urban soil. This study provides a hybrid framework that includes spatial distribution prediction of MPs, source-sink analysis, and pollution risk area identification, providing scientific and systematic methods and techniques for pollution management in other soil environments.