An Evaluation of Convolutional Neural Networks for Lithological Mapping Based on Hyperspectral Images

Hyperspectral remote sensing images are characterized by nanoscale spectral resolution and hundreds of continuous spectral bands, dominating significantly in geological applications ranging from lithological mapping to mineral exploration. A major challenge lies in how to incorporate spectral and spatial information, therefore promoting classification performance for detecting closely resembling and mixed minerals and lithologies. Recent advances in deep learning techniques have facilitated the application of hyperspectral images in geological studies, especially experts at handling high-dimensional data with strong neighboring correlation. As a result, this study focuses on the evaluation of deep learning algorithms for lithological mapping based on hyperspectral images and further provides guidance on mineral exploration. Four deep convolutional neural networks (CNNs), including 1-D CNN, 2-D CNN, 3-D CNN, and a hybrid of 1-D and 2-D CNN, were constructed for spectral, spatial, and spatial–spectral feature extraction. The proposed frameworks were verified through the case studies of lithological mapping to aid in prospecting rare metal deposits using Gaofen-5 hyperspectral images in the Cuonadong dome, Tibet, China. Lithological classification maps indicated that the dual-branch 1D–2D CNN yields better performance than others in both visual and quantitative comparisons, owing to the support of joint spatial–spectral feature learning. An overall classification accuracy of up to 97.4% further illustrates the feasibility of CNN models for lithological mapping based on hyperspectral images, which provides a realizable and promising approach for mineral exploration by mapping specific lithologies.