PSO-Transformer for mapping soil heavy metals using UAV hyperspectral data with spectral calibration and indices

In this study, UAV hyperspectral imagery and 52 topsoil samples were collected from a representative industrial legacy site in Tieling City, Liaoning Province. UAV spectra were calibrated to lab spectra using piecewise direct standardization (PDS). Feature bands were selected using the competitive adaptive reweighted sampling (CARS) method; feature band combinations were created using the dual-band spectral indices (DBSIs) and three-band spectral indices (TBSIs). A Transformer algorithm was optimized by particle swarm optimization (PSO) to predict soil copper (Cu) and arsenic (As) and to generate distribution maps. The results indicated that the PDS substantially reduced environmental effects in the UAV data. Spectral indices improved prediction accuracy, and TBSIs consistently outperformed DBSIs. The PSO-Transformer with TBSI-3 achieved the best performance, with validation R 2 of 0.83 for Cu and 0.88 for As. We propose a "sky-ground" hyperspectral inversion model. It enables high-accuracy prediction of soil heavy metal concentrations and provides a robust tool for monitoring contamination in industrial legacy sites.