Systematic evaluation of deep learning backbones and attention mechanisms for semantic segmentation of SAR-based satellite imagery

Semantic segmentation of Synthetic Aperture Radar (SAR) imagery remains challenging due to speckle noise, low signal-to-noise ratio, and complex backscatter characteristics. Although recent studies increasingly adopt transformer-based backbones, their effectiveness for SAR imagery remains insufficiently understood. This study presents a systematic evaluation of backbone–attention interactions across four segmentation architectures (BiSeNetV1, DeepLabv3+, UNet, and UNetFormer) using two datasets with distinct characteristics: Gaofen-3 PolSAR and Sentinel-1. The results show that the effectiveness of the backbone is strongly influenced by data characteristics. On the more complex Gaofen-3 dataset, cardinality-driven convolutional backbone ResNext consistently achieves the highest segmentation accuracy. In contrast, on the simpler Sentinel-1 dataset, performance differences across evaluated backbones become much smaller, and EfficientNet-V2-S provides the best accuracy–efficiency trade-off. Transformer-based backbones consistently underperform in both datasets, suggesting limited compatibility with the evaluated SAR data characteristics. Among the attention mechanisms, SE-Block consistently provides the largest and most stable performance gains, while CBAM offers additional improvements primarily in more complex scenes. Transformer-style window attention yields limited or negative benefits. These findings challenge the widespread preference for transformer-based designs and indicate that aligning architectural inductive biases with SAR data characteristics is critical for developing accurate, efficient, and scalable SAR semantic segmentation models.