High-Resolution Estimation of Cadmium Input Flux to Soil at a Regional Scale Through Hydrological Pathways: Coupling Process Models and Machine Learning

Abstract

Existing models for predicting heavy metal input fluxes to soil by hydrological pathways often suffer from limitations in spatiotemporal resolution and the event input flux from flooding, particularly in capturing complex hydrological dynamics and accurately linking pollutant sources to sinks. To overcome these challenges, we developed a novel integrative model that combines emission inventories, hydrological transport mechanisms, and machine learning (ML) correction to quantify cadmium (Cd) input fluxes through irrigation and flood pathways at a regional scale. The model integrates: (1) an enhanced SWAT model coupled with advection-diffusion equations for high-resolution Cd transport simulation in river networks, (2) MIKE FLOOD for estimating Cd input during flood events, and (3) an XGBoost-based ML correction module to align simulated fluxes with empirical soil Cd accumulation data. Applied to Hengnan County, the model estimates annual Cd input fluxes of 0.91 mg·m⁻² via irrigation and 4.63–7.95 mg·m⁻² during flood events. Sediment plays a crucial role in Cd transport, contributing 16.82% and 32.17% of the total flux through irrigation and flooding pathways, respectively. The corrected cadmium input flux via hydrological pathways averages 2.71 mg·m⁻²·a⁻¹, ranging from 0.01 mg·m⁻²·a⁻¹ to 109.72 mg·m⁻²·a⁻¹, with 28.08% of farmland affected by combined inputs from both irrigation and flood inundation. Model validation demonstrates high predictive accuracy (  = 0.89 for ML correction). This integrated approach provides a robust, scalable framework for assessing cadmium fluxes, offering valuable insights for regional soil cadmium risk management and pollution control strategies.

Keywords: Cadmium, Input flux, Hydrological pathways, Process models, Machine learning