CoastVisionNet: transformer with integrated spatial-channel attention for coastal land cover classification

Introduction The rapid advancement of satellite sensing technologies and the growing need for high-resolution environmental intelligence have highlighted coastal land cover classification as a vital yet challenging task in remote sensing. Coastal zones, being highly dynamic and spatially heterogeneous, require sophisticated semantic modeling strategies that account for both spectral variability and spatial morphology. While traditional convolutional neural networks and fixed-resolution transformer models have made notable strides, they often struggle to generalize across varying topographies and spectral distributions. These limitations stem from rigid spatial encoding schemes, insufficient spectral differentiation, and a lack of dynamic reasoning capabilities. Methods To overcome these challenges, we introduce CoastVisionNet, a transformer-based framework with integrated spatial-channel attention tailored for coastal land cover classification. The system builds on a robust theoretical foundation and is structured around three components: a novel Spectral-Topographic Encoding Network (STEN) for dual-path spectral and morphological representation, a geometry-aware self-attention for cross-modal feature fusion, and a Spectrum-Guided Semantic Modulation (SGSM) strategy for adaptive inference. STEN captures fine-grained spectral gradients and terrain-aware vector fields, enabling the model to preserve topological and spectral consistency across heterogeneous coastal scenes. SGSM enhances generalization by incorporating spectrum-conditioned priors, uncertainty-aware regularization, and curriculum-based spectral reweighting. Results Extensive experiments on diverse coastal satellite datasets demonstrate that CoastVisionNet significantly outperforms existing baselines in classification accuracy, especially in out-of-distribution regions and under varying imaging conditions. Discussion Furthermore, the model exhibits high transferability across different sensors and temporal snapshots, making it well-suited for the complex, evolving nature of coastal environments. This work aligns strongly with emerging priorities in intelligent remote sensing, offering a scalable, interpretable, and physically grounded framework for next-generation coastal monitoring.