Monitoring the Centennial Variation of Heavy Metals in Lake Sediments and Influencing Factors Using Environmental Magnetism and Machine Learning Methods

Abstract

The association between the magnetic properties of lake sediments and heavy metal(loid)s (HMs) is well-documented; however, their correlation with the chemical fractions of HMs remains under-investigated. Developing a robust workflow for predicting HMs risk utilizing various machine learning techniques in conjunction with magnetic analysis presents a complex challenge. This study assessed the predictive efficacy of nine machine learning models for determining the chemical fractions of HMs, employing magnetic parameters derived from sediment cores in a large, shallow lake. These models encompassed random forest, support vector machine, relevance vector machine, extreme gradient boosting, principal component regression, multivariate adaptive regression splines, gradient boosting with component-wise linear models, and lasso and elastic-net regularized generalized linear models. The support vector machine model demonstrated superior performance, achieving coefficient of determination values surpassing 0.8 in both training and testing phases. Through interpretable machine learning approaches, key drivers of HMs were identified among magnetic and physicochemical indicators. Magnetic susceptibility values, high coercivity remanent magnetization, ratios of anhysteretic remanent magnetization to magnetic susceptibility, and anhysteretic remanent magnetization to saturation isothermal remanent magnetization within specific ranges exhibited a positive correlation with Cd, Hg, and Sb. This research significantly advances our understanding of HMs risk assessment in lake sediments by leveraging accessible magnetic measurements within an interpretable machine learning framework.