Machine and deep learning-driven geospatial prediction modelling of soil cadmium and nickel with pollution risk mapping underlaid by black shale in Qianjiang County, China

Study region: Qianjiang County, China Study focus: The spatial distributions of soil Cd and Ni underlaid by black shale were predicted using inverse distance weighting (IDW), back propagation neural network (BPNN), random forest (RF), and particle swarm optimization (PSO) coupled with BPNN and RF to construct PSO-BPNN and PSO-RF and long short-term memory (LSTM) models. New hydrological insights for the region: The results indicated that the LSTM model displayed the highest R2 value (Cd: 0.841, Ni: 0.852) and lowest root mean square error (RMSE), mean absolute error (MAE), and mean bias error (MBE). The R2 values were followed as the PSO-RF (Cd: 0.827, Ni: 0.827), PSO-BPNN (Cd: 0.813, Ni: 0.833), RF (Cd: 0.771, Ni: 0.791), BPNN (Cd: 0.766, Ni: 0.781), and IDW (Cd: 0.549, Ni: 0.592) models. The prediction accuracy was ranked as follows: deep learning (LSTM) > optimization machine learning (PSO-RF, PSO-BPNN) > single machine learning (RF, BPNN) > traditional interpolation (IDW), demonstrating that the LSTM model achieved the best estimation performance. The areas with high and moderate Cd pollution risk were clustered in the central-western part of the study area. Such areas of risk were not observed for Ni. This study provides different prediction models for soil Cd and Ni underlaid by black shale and offers a theoretical basis for hydrological or rockwater interaction research in this region.