Lithology identification through hyperspectral remote sensing based on the SCNP-1DCNN-DRS network model

Hyperspectral remote sensing enables the simultaneous capture of spatial and spectral information. Based on this advantage and in combination with the significant learning capacity of deep learning technology, this study proposed a deep learning-based hyperspectral remote sensing approach for lithology identification, aiming to further enhance the quality and efficiency of lithologic mapping. Through a comprehensive analysis of the merits and demerits of existing lithology classification methods based on hyperspectral remote sensing and taking into account the geological attributes in lithologic mapping, this study proposed a higher-precision lithologic mapping model tailored to satellite hyperspectral remote sensing-SCNP-1DCNN-DRS. While inheriting the outstanding efficiency and feasibility possessed by the one-dimensional convolutional neural network (1DCNN) in lithologic mapping, the proposed SCNP-1DCNN-DRS network model introduced the learning strategy of spectral correlation of neighboring pixels (SCNP) to overcome the limitations of 1DCNN in lithologic mapping, such as the presence of numerous isolated noise pixels and imprecise boundary delineation. Additionally, a deep residual shrinkage (DRS) denoising module was integrated into the network to further mitigate the impacts of noise and redundant information in satellite hyperspectral images. Experimental validation demonstrates that the SCNP-1DCNN-DRS network model yielded an overall sample testing accuracy of 99.89% and an overall full-image testing accuracy of 83.41%. Compared to other models, the SCNP-1DCNN-DRS network model exhibits the highest accuracy in full-image testing, indicating its great application potential in lithologic mapping practices.