Geological Remote Sensing Interpretation via Multilevel Feature Integration Network

Geological remote sensing interpretation enables the extraction of task-specific geological information from imagery, playing a vital role in geological investigation and analysis. Nevertheless, deep learning-based methods still lag behind expert visual interpretation in terms of accuracy, owing to the high spectral similarity, fragmented spatial distributions, and class imbalance of geological elements. This study proposes a multilevel feature integration network (MLFIN) to address the above challenges in multiclass geological element interpretation. Specifically, a class-wise random sampling strategy is designed to alleviate class distribution imbalance, improving the capability of MLFIN to learn features from diverse geological elements. The encode–decode modules can enhance feature representation and improve the utilization of element information. The former integrates convolution operations with attention mechanisms to extract multiscale spatial–spectral features, while the latter enables the adaptive fusion of multilevel features. To mitigate the boundary ambiguity segmentation, the global attention module models long-range contextual dependencies, while the segmentation module refines spatial features, enriching the spatial details of geological elements and strengthening boundary feature discrimination. Experimental results on two geological datasets indicate that MLFIN outperforms current deep learning-based segmentation methods, achieving Overall Accuracy value of 72.83% and 67.36%, and mean Intersection-over-Union values of 55.6% and 46.16%, respectively. It validates the strong generalization capacity across diverse geological environments and proves the effectiveness of its submodule in multiclass geological interpretation tasks.