Improving rock unit mapping in vegetation-covered areas using multi-scale hyperspectral wavelet features

The obscuring effect of vegetation on rock spectra has become a major challenge for remote sensing-based rock unit classification in vegetated areas. In this study, the Hapke nonlinear mixing model was used to simulate vegetation-rock mixed spectra, and the separability of different rock types was evaluated using GaoFen-5 Advanced Hyperspectral Imager (AHSI) data under varying vegetation cover conditions. To address spectral mixing caused by vegetation, one-dimensional (1D) discrete wavelet transform (DWT) high-frequency features were employed, which significantly improved rock separability in the spectral domain. In the spatial-spectral dimension, multi-scale texture features extracted using three-dimensional discrete wavelet transform (3D-DWT) achieved higher overall accuracy (OA) than traditional texture extraction methods including the grey-level co-occurrence matrix (GLCM) and local binary patterns (LBP). Simulation experiments demonstrated that 1D-DWT method significantly improved lithological separability under vegetation interference. Under category-balanced conditions, 3D-DWT improved OA by 28.42% (from 50.71% to 79.13%) in rock unit classification accuracy. Rock unit mapping using only 1% training samples achieved an OA of 75.26% outperforming results obtained with original features and 1D-DWT features. Experimental analysis also showed that the haar wavelet function and larger spatial windows further improved classification performance. By constructing localized time-frequency frameworks within the spatial-spectral domain, 3D-DWT better captures the spatial texture and spectral variation characteristics of rock units within vegetation-covered areas, demonstrating strong potential for rock classification under limited training samples.