Entropy-guided multi-scale spatial attention for discriminative coastal scene classification in high-resolution remote sensing images

Accurate coastal scene classification from high-resolution remote sensing imagery remains challenging due to strong inter-class visual similarity and the presence of multi-scale spatial patterns. To address these challenges, this paper introduces a Multi-Scale Spatial Entropy Squeeze-and-Excitation (MSSE) attention block designed to enhance discriminative feature learning within convolutional neural networks. MSSE integrates multi-scale spatial pooling with entropy-driven channel and spatial attention, enabling adaptive emphasis on informative features across different spatial extents. The proposed method is evaluated using a unified CNN backbone on a coastal-oriented subset of the MLRSNet dataset, comprising five representative classes: wetland, river, lake, beach, and island. Extensive experiments demonstrate that MSSE consistently outperforms established attention mechanisms, including SE, CBAM, and ChannelEntropySE. Specifically, MSSE achieves a macro F1-score of 95.1% and an IoU of 89.4%, representing improvements of +3.2% (F1) and +4.1% (IoU) over the standard SE module and +1.4% (F1) and +1.5% (IoU) over ChannelEntropySE. In addition, repeated runs with different random seeds show that MSSE exhibits lower performance variance, confirming its stability and robustness. Further analyses across varying train–validation splits and computational cost assessments reveal that MSSE provides a favorable balance between accuracy and efficiency, introducing only a modest increase in parameters and FLOPs compared to baseline attention modules. These results highlight the effectiveness and practicality of entropy-aware multi-scale attention for robust coastal scene classification and related remote sensing applications.