Abstract: Aiming at the problems of weak spectral response between remote sensing image reflectivity and heavy metal elements, and low accuracy of soil heavy metal classical inversion models.In this study, Sentinel-2 remote sensing image is used as the data source, and the dimidiate pixel model is used for image spectral unmixing. The characteristic spectra with high correlation are selected as spectral parameters. PLS model and GMDH model of soil reflectance and heavy metal Cr content under different spectral transformations and pixel linear unmixing are constructed.The results show that there is a significant correlation between the spectra and the content of Cr, and the number of bands is increased and the correlation is enhanced.The GMDH model constructed based on the unmixed soil spectrum and the heavy metal Cr content has better model stability, stronger predictive ability and better accuracy.This research method expands the traditional method of remote sensing image inversion, and can provide a useful reference for monitoring soil heavy metal pollution in a large scale.

Key words: spectral unmixing, image soil reflectance, PLS model, GMDH model, heavy metal Cr