Lithologic Unit Classification Attention-Based and Multiscale Geology Knowledge-Guided Framework in Vegetated Areas

Lithologic unit classification is essential for resource surveys and infrastructure planning. Remote sensing provides a rapid and scalable alternative to traditional field surveys but faces challenges in vegetation-covered areas due to limited lithological information. To address this, we propose an attention-based multiscale a priori knowledge-guided Network (AMSNet), which integrates multiscale-enhanced cross-spatial attention (MSECA) and wavelet-enhanced crisscross attention (WCCA). MSECA extracts lithologic features by aggregating multiscale neighborhood information and introducing cross-spatial learning, while WCCA enhances long-range contextual understanding using wavelet transforms to mitigate information loss during key-query generation. Experiments on Qichun and Tieshan datasets demonstrate that AMSNet achieves an overall accuracy (OA) of 69.03% and 83.77%, a mean intersection over union (mIoU) of 48.41% and 48.18%, and a Macro- $F1$ score of 63.65% and 60.68%, outperforming baseline models. Ablation studies further confirm the effectiveness of MSECA and WCCA in improving lithologic feature extraction. These findings demonstrate the potential of AMSNet to improve lithologic classification in complex terrains, contributing to advancements in geoscience applications.