Improving heavy metal prediction using hybrid feature selection and ensemble learning with multi-source remote sensing in Red River Delta mangroves, Vietnam

Mangrove ecosystems play a critical role in supporting coastal aquaculture by providing essential ecosystem services and reducing HM exposure to adjacent aquatic environments. Despite their importance, mangroves have declined sharply worldwide due to conversion for aquaculture and urban development, leading to substantial HM release and transport from sediments into surrounding waters. Quantitative assessment of mangroves’ capacity to regulate heavy metals remains limited, as monitoring trace-level HM concentrations over large spatial scales is technically challenging. This study presents a novel and automated ensemble learning framework for large-scale prediction of arsenic (As) and lead (Pb) concentrations in mangrove soils using multisource Earth observation data. Optical Sentinel-2 (MSI) combined with C-band SAR (Sentinel-1), and L-band SAR (ALOS-2 PALSAR-2) data were integrated with field measurements from 101 soil cores to derive complementary spectral, vegetation and soil indices, textural, and backscatter features. Multiple machine learning models were trained and systematically optimized using stacking and five-fold cross-validation within the AutoGluon framework. The weighted Level-2 ensemble consistently outperformed seven Level-1 base learners, achieving high predictive accuracy for both metals (R2 > 0.75). The hybrid genetic algorithm-particle swarm optimization (GA-PSO) approach for optimal feature selection further improved performance, increasing R2 to 0.816 (As) and 0.886 (Pb), while reducing RMSE to 4.266 and 6.293 mg kg−1, respectively. The proposed workflow demonstrates the added value of multisensor data fusion, feature selection, and automated ensemble learning for mapping trace-level soil contaminants in complex coastal environments. This is the first systematic, large-scale assessment of heavy metal accumulation in Vietnamese mangroves that integrates optical and multi-frequency SAR data with advanced ensemble modelling. The framework is computationally efficient, scalable, and transferable, offering a practical solution for regional to national-scale environmental monitoring and ecological risk assessment from space.