Predicting the effects of microplastics on soil properties using machine learning

Abstract

Microplastics (MPs) are widely distributed in soils, posing a significant threat to the health of soil ecosystems. MPs can affect soil properties, including physicochemical characteristics and microbial communities. However, predicting the impacts of MPs on soil is challenging due to the variability in experimental conditions and the diversity of soil types. Based on the data collected from peer-reviewed literatures, this study predicted the impacts of MPs on soil properties using machine learning. After PyCaret's integrated model evaluates the current mainstream models, it was found that the CatBoost regression model is the most suitable for this dataset under multi-index evaluation, demonstrating high predictive accuracy with R² values exceeding 0.8. Feature importance analysis revealed that dissolved organic carbon (DOC), nitrate nitrogen (NO₃⁻-N), and available phosphorus are the most affected soil properties, with MP size and the exposure time in soil being the most influential factors. As exposure time increases, key soil fertility indicators—such as DOC, ammonium nitrogen (NH₄⁺-N), and available phosphorus—show a significant decline, while pH and microbial indicators increase. This indicates that the long-term presence of MPs in the soil may lead to a reduction in soil fertility. Overall, our study successfully establishes a predictive model for analyzing and predicting changes in soil caused by MPs, providing valuable technical support for assessing the impact of MPs on soil properties.