Assessment and prediction of soil heavy metal pollution in cotton fields by multi-source data feature fusion combined with machine learning
Xianglong Fan, Pan Gao, Jingang Wang, Mengli Zhang, Hong Shuai, Shizhe Qin, Qiang Zhang, Lulu Ma, Lifu Zhang, Zhenxi Zhang, Xin Lv
Abstract
To reveal the key variables affecting soil heavy metal pollution in cotton fields and the predictive power of fusion of important features derived from multi-source data on soil heavy metal content, in this study, a multi-source data feature fusion combined with machine learning (ML) was proposed. The soil available heavy metal (AHM) contents in cotton fields in northern Xinjiang, China were determined. Then, importance analysis was conducted on remote sensing , natural environment, and human activity data to extract important feature variables. By fusing important feature variables, heavy metal pollution prediction models were constructed by multiple machine learning methods (LightGBM, RF, PLSR , XGBoost), and the best one was selected. The results showed that there were severe nickel pollution and mild lead and chromium pollution in the study area. Fertilizers and plastic films were important variables affecting the AHM content in cotton fields. The ML models constructed based on the fusion of important variables had the highest accuracy, increasing by 19.48 %-30.96 % and 31.19 %-46.63 % compared with that of the full dataset model and the remote sensing dataset model, respectively. Nickel pollution was severe in the central region, chromium pollution was obvious in the eastern region, and lead pollution was mostly detected in the eastern and central regions. This study will provide a scientific basis for soil heavy metal pollution monitoring, prediction, and control and contribute to the sustainable agricultural development .