Mapping soil cadmium content using multi-spectral satellite images and multiple-residual-stacking model:...

Soil cadmium (Cd) contamination significantly threatens ecosystems and human health. Traditional geochemical investigation, although accurate, is impractical for wide-area and frequent monitoring applications. Multi-spectral satellite images combined with the homologous pollution information (HPI) and the spectral and content information of soil organic matter (SOMSCI) is an unconventional and promising approach for large-scale, dynamic soil heavy metal (SHM) monitoring. Based on a novel Multiple-Residual-Stacked (MRS) machine-learning framework, the study estimated the soil Cd content in Yueyang City, China, during the past decade (2014-2023) using Landsat 8 images. Within it, three feature construction methods and four models were employed. The experimental results indicate that the XGB-MRS model incorporating HPI and SOMSCI significantly improved the estimation performance (RPD exceeded 90 %, R², RMSE, and MAE exceeded 40 %). Moreover, against 243 ground samples during 2016-2022, the average overall estimation accuracy exceeded 80 %, validating the model's robustness and practicality. Furthermore, the descending order of contribution in the modelling is environmental auxiliary variables (55 %), HPI and SOMSCI (26 %), and spectral information (19 %). The fertilizer usage has direct (up to 2 years) and delayed (3-5 years) effects on soil Cd accumulation. Overall, our study provides a scalable framework for monitoring global SHM pollution using open-source multi-spectral satellite data.