Multimodal loess cave extraction using LMUNet-RCAM: Integrating hillshade and RGB features with local attention mechanisms

Loess caves (LCs), a typical type of micro-landform widely distributed across the Chinese Loess Plateau, constitute a significant geological hazard. Accurate identification of LCs is essential for mitigating risks to human life, infrastructure, and environment. However, the complex morphology and subtle topographic expression of LCs present substantial challenges for reliable detection. This study proposes LMUNet-RCAM, an improved global–local feature enhancement network that integrates hillshade information derived from Digital Elevation Models (DEMs) to improve representation robustness. To address the irregular spatial patterns of LCs, the network incorporates novel Local Transformer mMoBA(LM) to refine feature representations prior to their input into the U-Net backbone, thereby expanding the effective receptive field in both global and local contexts. Furthermore, a Recurrent Convolutional Attention Module (RCAM) is introduced to enhance spatial and channel-wise semantic discrimination. Within RCAM, Gated-Swin-RNN (G-S-R) is employed to reconstruct local feature sequences, improving network stability under complex terrain conditions. Extensive experiments on test area demonstrate that LMUNet-RCAM consistently outperforms existing state-of-the-art semantic segmentation methods in LCs detection.