Answers are grounded in this paper and cite their sources.
Lithological interpretation, a fundamental task in geological remote sensing, aims to discriminate different rock and soil types, playing a crucial role in geological mapping, resource exploration, and environmental monitoring. Although deep learning techniques have shown superior capabilities in feature extraction and generalization, their application to lithological interpretation remains challenging. These challenges primarily arise from inconsistent responses in multimodal data, which lead to redundant information that degrades model performance. In addition, conventional convolutions show limited capability in mining lithological features during multimodal fusion. To overcome these challenges, we propose a response-optimized structure-aware multimodal fusion network (ROSANet) for lithological interpretation. First, we propose a responsiveness-aware cross-modal channel substitution module to identify and replace low-response channels between modalities based on activation and normalization statistics, thereby suppressing redundant information and enhancing salient features. Second, we propose a dynamic structure-aware fusion module that fuses multimodal features through compression and reconstruction, integrating a learnable nonuniform receptive mechanism to capture lithological directional features. Furthermore, Sentinel-1 SAR data are incorporated as a complementary modality, providing scattering-based structural cues that compensate for the limitations in optical channels. Comprehensive experiments conducted on the Yingjisha dataset and the Djibouti dataset demonstrate that ROSANet achieves improvements of 2.7% in overall pixel accuracy and 4.2% in mean intersection over union compared to eight state-of-the-art models. These results verify the effectiveness and robustness of ROSANet for lithological interpretation in complex geological environments.
No takes yet. Share an insight, caveat, or question.
Han et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: