Remote Sensing Lithology Intelligent Segmentation Based on Multi-source Data

Due to complex geological and physicochemical processes, rocks indicate spectral variability and diversity, especially in areas with vegetation development and mountain shadows. In addition, there are differences in the preservation information of multi-source remote sensing data. This article focuses on the problem of traditional models and single remote sensing data which are difficult to effectively extract geological features. A remote sensing lithology semantic segmentation method based on multimodal data adaptive fusion is proposed. To address the issue of redundant information interference caused by direct fusion of multimodal data, utilizing the advantage of high resolution of optical data, a step-by-step fusion method was adopted, which combined SAR data and DEM data separately. The channel attention mechanism was used to learn the eights of optical data to other types of data, and the obtained weights are weighted on each type of data. In addition, in order to distinguish the importance of various features, multiple attention was used to explore the connections between space and channels to enhance the model’s ability to extract key feature information. A remote sensing lithology semantic segmentation method based on prior knowledge embedding was also established. In addition, the existing small-scale geologic map was taken as a priori knowledge, which was used as a label to add additional semantic segmentation tasks, further mining the deep information hidden between lithology, and enhancing the lithologic feature extraction and generalization capabilities of the model. Experimental comparisons were conducted with various popular models on the lithology segmentation dataset constructed in book, and the results proved the superiority of the method in this article.