Interpretable Dual-Channel Convolutional Neural Networks for Lithology Identification Based on Multisource Remote Sensing Data

Lithology identification provides a crucial foundation for various geological tasks, such as mineral exploration and geological mapping. Traditionally, lithology identification requires geologists to interpret geological data collected from the field. However, the acquisition of geological data requires a substantial amount of time and becomes more challenging under harsh natural conditions. The development of remote sensing technology has effectively mitigated the limitations of traditional lithology identification. In this study, an interpretable dual-channel convolutional neural network (DC-CNN) with the Shapley additive explanations (SHAP) interpretability method is proposed for lithology identification; this approach combines the spectral and spatial features of the remote sensing data. The model adopts a parallel dual-channel structure to extract spectral and spatial features simultaneously, thus implementing lithology identification in remote sensing images. A case study from the Tuolugou mining area of East Kunlun (China) demonstrates the performance of the DC-CNN model in lithology identification on the basis of GF5B hyperspectral data and Landsat-8 multispectral data. The results show that the overall accuracy (OA) of the DC-CNN model is 93.51%, with an average accuracy (AA) of 89.77% and a kappa coefficient of 0.8988; these metrics exceed those of the traditional machine learning models (i.e., Random Forest and CNN), demonstrating its efficacy and potential utility in geological surveys. SHAP, as an interpretable method, was subsequently used to visualize the value and tendency of feature contribution. By utilizing SHAP feature-importance bar charts and SHAP force plots, the significance and direction of each feature’s contribution can be understood, which highlights the necessity and advantage of the new features introduced in the dataset.