Multi-indicator water-quality prediction in mining areas using a feature-tokenizer transformer with spatiotemporal...

Mining activities can produce persistent and spatially heterogeneous impacts on water quality, but stable multi-indicator prediction across large-scale mining areas remains challenging. We developed a Feature-Tokenizer Transformer (FT-Transformer) multi-task framework for nationwide mining-area water-quality prediction by integrating 55,744 monitoring records with spatial, mining-related, temporal, categorical, and interpolated monthly mean air temperature. The model jointly predicted pH, dissolved oxygen (DO), ammonium nitrogen (NH₄-N), and permanganate index (COD_Mn), and was evaluated using 5 × 3 nested cross-validation against Elastic Net, Random Forest, and XGBoost. The FT-Transformer achieved a mean coefficient of determination (mean R²) of 0.790 ± 0.008, with R² values of 0.756, 0.756, 0.826, and 0.820 for pH, DO, NH₄-N, and COD_Mn, respectively. It achieved the best overall performance and outperformed the benchmark models for pH, NH₄-N, and COD_Mn, whereas Random Forest showed a slight advantage for DO. Multi-task ablation indicated that shared representations improved joint prediction, especially for NH₄-N and COD_Mn. Response, attribution, subgroup, and scenario analyses showed indicator-specific and context-dependent model behavior: temperature was most consistently associated with DO, while mining-stage and regional background variables contributed more strongly to NH₄-N and COD_Mn. The proposed framework supports multi-indicator water-quality prediction, model applicability screening, and monitoring-priority assessment in heterogeneous mining areas.