Early warning study on ecological risk of heavy metals in soil based on XGBoost-IER coupled model - A case study of black soil area in Jilin province

Abstract The heavy metals (HMs) content of arable soils in the black soil zone of Jilin Province has exceeded the background thresholds of the soil environment in Jilin Province, and localised contaminated areas have appeared as a result of the influence of factors such as neighbouring industrial discharges and anthropogenic activities. In this study, we innovatively combined the Extreme Gradient Boosting Algorithm (XGBoost) with the Index of Ecological Risk Early Warning ( I ER ) method. We constructed the XGBoost-I ER composite model for predicting the concentration of soil HMs (Cd, Cu, Zn, Pb, and Mn) and for ecological risk early warning assessment. This paper compares the performance of three spectral preprocessing methods (Standard Normal Variable Transform (SNV), Multivariate Scattering Correction (MSC), and Savitzky-Golay Convolutional Smoothing First-order Derivative (SG-FD)), as well as the Competitive Adaptive Re -weighted Sampling (CARS) feature selection method. The results show that SG-FD preprocessing combined with CARS feature selection can extract feature information more effectively. The XGBoost model is further compared with Least Squares Support Vector Machine (LS-SVM) and Partial Least Squares Regression (PLS), and the results show that the XGBoost model has a significant advantage in both prediction accuracy and stability. The ecological risk early warning analysis based on the model prediction shows that Cd and Zn present a moderate pollution warning in the central part of the study area, and Cu, Pb and Mn are all mildly warned; From the comprehensive I ER , there is a heavy pollution warning area in the central and southern part of the study area (Jilin, Changchun and Siping). The ecosystem structure has suffered severe damages. The agreement between the model prediction results and the actual detection data reached: Mn (98.21 %), Cu (94.64 %), Cd (91.96 %), Pb(91.07 %), which fully verified the reliability of the method. • Hyperspectral technology enables rapid, large-scale prediction of heavy metal concentrations. • XGBoost outperforms PLS and LS-SVM models in the hyperspectral inversion of soil HMs. • Combined XGBoost with I ER to propose a novel ecological risk model. • The spatial visualisation of HM ecological risk pinpoints regions exhibiting ecological vulnerability.