Abstract: Lithology identification is a critical task in geology and geotechnical engineering. However, traditional methods heavily depend on professionals with geological expertise, making the process subjective, time-consuming, and labor-intensive, which is unsuitable for large-scale or rapid deployment. To enhance identification performance in complex backgrounds and with ambiguous rock features, the authors propose an intelligent lithology identification method based on adaptive fusion of visual features. The method integrates color, texture, and shape features for fast and accurate classification of rock images. First, a Panoptic FPN approach based on the feature pyramid network is employed to automatically generate pixel-level annotations, considerably minimizing the need for manual labeling and reducing time costs. Next, a Grad-CAM-based segmentation model is utilized to isolate rocks from the background using class activation heatmaps, which reduces background interference and improves feature extraction accuracy. During feature extraction, color, texture, and shape features are extracted, and their contributions to classification are quantified using the integrated gradients method. Based on these evaluations, a weighting mechanism is introduced to adaptively fuse the three feature types. The resulting fused representation is then combined with shallow features from the original image to facilitate intelligent lithology identification. Experimental results show that: (1) the proposed method achieves a classification accuracy of up to 99.1% on the test set, representing a 3.5% improvement over models trained on original images. This confirms the effectiveness of the multifeature fusion and weight-guided mechanism in enhancing identification performance. (2) The feature contribution analysis based on integrated gradients shows strong alignment with manual interpretations, suggesting that the method not only improves model accuracy but also enhances interpretability, making it a valuable tool for geological surveys and geotechnical engineering.

Key words: lithology identification, deep learning, image classification, feature visualization, feature fusion, weighting mechanism