Explainable deep learning unveils critical scenarios driving soil cadmium pollution in a coastal industrial city in China: A geospatial AI approach.

Traditional methods for identifying heavy metal pollution sources in soil typically rely on physical and chemical analyses or statistical surveys. However, these approaches face challenges in effectively integrating spatial data, and their high material and time requirements limit their capacity for precise, long-term pollution source analysis. To address these limitations, this study proposes an explainable deep learning model that integrates remote sensing (RS) technology to identify heavy metal pollution sources in soil. By utilizing a lightweight image classification network (SI-LICNet) that preserves large-scale spatial information, the model successfully detected features related to cadmium (Cd) contamination in RS images from six regions in Dongguan City, China, achieving receiver operating characteristic (ROC) area under the curve (AUC) values of 0.848, 0.905, 0.800, 0.802, 0.747, and 0.772, respectively. The application of occlusion sensitivity analysis provided interpretability, highlighting areas with high sensitivity within each region. The results revealed that industrial and residential land types exhibited heightened sensitivity, while variations in natural conditions and economic factors led to distinct occlusion sensitivity patterns across regions. This research introduces an explainable deep learning approach for identifying heavy metal pollution sources, offering a more accurate and scientifically grounded methodology for soil environmental management.

Explainable deep learning unveils critical scenarios driving soil cadmium pollution in a coastal industrial city in China: A geospatial AI approach. Journal Articles