Automatic segmentation of landslides from remote sensing imagery is challenging because traditional machine learning and early CNN-based models often fail to generalize across heterogeneous landscapes, where segmentation maps contain sparse and fragmented landslide regions under diverse geographical conditions. To address these issues, we propose a lightweight dual-stream siamese deep learning framework that integrates optical and topographical data fusion with an adaptive decoder, guided multimodal fusion, and deep supervision. The framework is built upon the synergistic combination of cross-attention, gated fusion, and sub-pixel upsampling within a unified dual-stream architecture specifically optimized for landslide segmentation, enabling efficient context modeling and robust feature exchange between modalities. The decoder captures long-range context at deeper levels using lightweight cross-attention and refines spatial details at shallower levels through attention-gated skip fusion, enabling precise boundary delineation and fewer false positives. The gated fusion further enhances multimodal integration of optical and topographical cues, and the deep supervision stabilizes training and improves generalization. Moreover, to mitigate checkerboard artifacts, a learnable sub-pixel upsampling is devised to replace the traditional transposed convolution. Despite its compact design with fewer parameters, the model consistently outperforms state-of-the-art baselines. Experiments on two benchmark datasets, Landslide4Sense and Bijie, confirm the effectiveness of the framework. On the Bijie dataset, it achieves an F1-score of 0.9110 and an intersection over union (IoU) of 0.8839. These results highlight its potential for accurate large-scale landslide inventory mapping and real-time disaster response. The implementation is publicly available at https://github.com/mishaown/DiGATe-UNet-LandSlide-Segmentation (accessed on 3 November 2025).

Autonomous vehicles rely heavily on accurate and efficient scene segmentation for safe navigation and efficient operations. Traditional Bird’s Eye View (BEV) methods on semantic scene segmentation, which leverage multimodal sensor fusion, often struggle with noisy data and demand high-performance GPUs, leading to sensor misalignment and performance degradation. This paper introduces an Enhanced Channel Attention BEV (ECABEV), a novel approach designed to address the challenges under insufficient GPU memory conditions. ECABEV integrates camera and radar data through a de-noise enhanced channel attention mechanism, which utilizes global average and max pooling to effectively filter out noise while preserving discriminative features. Furthermore, an improved fusion approach is proposed to efficiently merge categorical data across modalities. To reduce computational overhead, a bilinear interpolation layer normalization method is devised to ensure spatial feature fidelity. Moreover, a scalable cross-entropy loss function is further designed to handle the imbalanced classes with less computational efficiency sacrifice. Extensive experiments on the nuScenes dataset demonstrate that ECABEV achieves state-of-the-art performance with an IoU of 39.961, using a lightweight ViT-B/14 backbone and lower resolution (224 × 224). Our approach highlights its cost-effectiveness and practical applicability, even on low-end devices. The code is publicly available at: https://github.com/YYF-CQU/ECABEV.git.