Soil heavy metal prediction using GF-5 hyperspectral image and environmental covariates with data augmentation

Hyperspectral images offer an efficient and reliable approach for monitoring soil heavy metal (SHM) contents over extensive regions. However, neglecting environmental covariates and limited soil samples can reduce the accuracy of SHM prediction models. Therefore, this study proposes a prediction method integrating environmental covariates and applying data augmentation to improve the reliability and accuracy of Pb, Zn, and Ni prediction. Firstly, geographic detector and Shapley additive explanations were employed to identify key environmental covariates, which were characterized by relevant spectral indices for modeling. Subsequently, a spectral smoothing-constrained Wasserstein generative adversarial network with gradient penalty (SS-WGAN-GP) was proposed to increase the sample size and enhance the reliability and realism of generated samples. The TabNet model was finally used to construct the SHM prediction models and generate SHM spatial distribution maps. The proposed method was validated using 105 topsoil samples collected from the Dongsheng coalfield in Inner Mongolia and a corresponding GF-5 hyperspectral image. The results indicate that the environmental covariates significantly improved model accuracy. Further enhancement of model performance was achieved by augmenting samples using the SS-WGAN-GP, yielding R² values of 0.70, 0.70, and 0.72 for Pb, Zn, and Ni, respectively. Spatial mapping of SHM contents and pollution levels based on the Nemerow Pollution Index suggested that elevated SHM contents were mainly distributed around mining areas and along major roads. This study provides a feasible framework for SHM prediction and pollution risk assessment under limited sample conditions.