Retrieving soil heavy metal concentrations using hyperspectral remote sensing images based on three-band indices at a coal mine in China

The soil heavy metal contamination in open-pit mining areas, receiving widespread attention globally, poses a severe threat to the eco-environment and public health. Currently, spectral indices derived from visible and near-infrared (VNIR), as well as short-wave infrared (SWIR) regions, have been introduced to estimate heavy metal contents. However, traditional ergodic spectral indices lack consideration of the spectral response features of heavy metals, resulting in low inversion efficiency and accuracy. Here, we collected 110 in situ surface soil samples and the Gaofen-5 (GF-5) hyperspectral satellite image from an open-pit coal mine in northern China. The three-band spectral indices under consideration of spectral response (CTBSIs) were constructed referring to existing heavy metal spectral responses. The Continuous Wavelet Transform (CWT) was employed for feature band selection. Subsequently, machine learning algorithms were adopted to infer heavy metal contents, and zinc (Zn) contents were mapped. Finally, the possible reasons for the Zn distribution were explored. The results indicated that the CWT significantly enhanced spectral responses, and the Zn inversion accuracy was obviously improved via the CTBSIs, referring to existing heavy metal spectral responses. The spectral activity bands for Zn are concentrated in 400–600 nm, 900–1000 nm, and 2200 nm. The Random Forest (RF) demonstrated higher accuracy in estimating Zn contents ((Formula presented)  = 0.82) than AdaBoost ((Formula presented)  = 0.78). Zn contents ranged from 43.36 to 128.34 mg/kg, and 98.5 % of the study areas exceeded the local background value of 48.60 mg/kg. Zn is mainly accumulated in the southwestern and northeastern parts of the study area. The topography and distance to mining sites may be the reasons for Zn distribution. This scheme can be a reference for other soil heavy metals inversion using hyperspectral remote sensing data. © 2025 COSPAR. Published by Elsevier B.V. All rights are reserved, including those for text and data mining, AI training, and similar technologies.