Introduction
Semantic segmentation of Polarimetric SAR (PolSAR) images is notoriously difficult due to intense speckle noise and a historical lack of large training datasets. This paper introduces MP-ResNet, a Multi-path Residual Network designed to capture broader contextual information and mitigate noise through parallel multi-scale branches. By effectively enlarging the network's receptive field without a massive computational penalty, MP-ResNet achieves state-of-the-art performance on high-resolution Gaofen benchmarks.
The Challenge of PolSAR Segmentation
Synthetic Aperture Radar (SAR) is invaluable for Earth observation due to its all-weather, day-and-night imaging capabilities. However, segmenting high-resolution PolSAR images into land-cover classes faces two major hurdles:
- Speckle Noise: The coherent imaging mechanism of SAR systems creates intense, granular noise, making pixel-level signals highly erratic.
- Data Scarcity: Deep learning models require massive amounts of labeled data, which has historically been scarce for SAR. This often leads researchers to use shallow networks that fail to capture the high-level semantic context needed to look past noise.
Methodology: The MP-ResNet Architecture
The authors propose the Multi-path Residual Network (MP-ResNet), which builds upon a Fully Convolutional Network (FCN) with a ResNet backbone.
Key innovations include:
- Quick Embedding of Shallow Features: A large 7x7 initial convolutional kernel aggregates features from a wider neighborhood to smooth speckle noise early. The feature scale is quickly reduced to 1/8, identified as the "sweet spot" for regional modeling.
- Multi-path Semantic Embedding: Inspired by HRNet, MP-ResNet splits into three parallel encoding branches. Unlike HRNet, it focuses purely on high-level features to enlarge the Valid Receptive Field (VRF), allowing the network to use broad spatial context to identify noisy pixels.
- Coarse-to-Fine Decoder Fusion: Instead of standard skip connections that might reintroduce low-level noise, MP-ResNet uses a channel-wise "Bottleneck" spatial deconvolution module to fuse only rich, high-level semantic features from its parallel encoder branches.
Experimental Results
MP-ResNet was evaluated on the Gaofen dataset (Gaofen-3 satellite, 1m-3m resolution). The model achieved an Overall Accuracy (OA) of 93.95% and a mean F1 score of 92.47% on validation sets.
As shown in the comparison table, MP-ResNet outperforms standard optical segmentation models like DeepLabV3+ and PSPNet, as well as SAR-specific models like MS-FCN. Notably, it achieved a 1.23% fwIoU boost on the hidden contest test set, securing 2nd place in the Gaofen Challenge.
In terms of efficiency, MP-ResNet requires 115.93 GFLOPS. While slightly higher than the baseline FCN (90.97 GFLOPS), it remains significantly more efficient than architectures like UNet (221.68 GFLOPS) or SegNet (341.06 GFLOPS).
Conclusion
MP-ResNet demonstrates that context is the ultimate noise filter for PolSAR imagery. By prioritizing broad spatial context over local pixel intensities and avoiding noise-prone low-level skip connections, the architecture sets a new benchmark for high-resolution SAR land-cover classification. Future work will explore integrating attention mechanisms and dilated convolutions to further refine performance.
Summary updated