Road Segmentation in SAR Satellite Images With Deep Fully Convolutional Neural Networks

Henry, Corentin; Azimi, Seyed Majid; Merkle, Nina · 2018 · Crossref

DOI: 10.1109/lgrs.2018.2864342

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of automatically extracting road networks from Synthetic Aperture Radar (SAR) satellite images, a task critical for updating maps and supporting disaster relief. While SAR data offers robustness against weather and illumination changes, roads are difficult to identify due to visual similarities with other linear features like rivers and railways. Although deep learning has succeeded in optical imagery, its application to SAR road segmentation remains underexplored. The authors evaluate the potential of Fully-Convolutional Neural Networks (FCNNs) for this task, specifically investigating how to adapt these architectures to handle the thin, sparse nature of roads in noisy SAR data. The study employs three FCNN architectures: FCN-8s with a VGG-19 backbone, Deep Residual U-Net, and DeepLabv3+. To address the class imbalance and the thinness of road objects, the authors introduce specific training adjustments. They replace the standard cross-entropy loss with a class-weighted Mean Squared Error (MSE) loss, treating the task as binary regression rather than classification. Additionally, they implement a spatial tolerance parameter ($t_{max}$) that smooths the ground truth labels, allowing the model to tolerate small spatial deviations. The models were trained from scratch on a custom dataset of high-resolution TerraSAR-X images from Lincoln, England, Kalisz, Poland, and Bonn, Germany. Urban areas were masked out to focus on rural road networks, and manual labeling was performed using optical references. The authors also tested the impact of Non-Local filtering and Fully-connected Conditional Random Fields (FCRFs) as pre- and post-processing steps. The results demonstrate that FCNNs can effectively extract road candidates from SAR images. DeepLabv3+ achieved the best overall performance, averaging an Intersection over Union (IoU) of approximately 44% across test sets, with smoother and less noisy predictions than FCN-8s. FCN-8s achieved comparable IoU scores (e.g., 45.46% on the Lincoln test set) but produced noisier outputs. Deep Residual U-Net performed significantly worse, likely due to insufficient depth to abstract SAR speckle noise. The spatial tolerance parameter improved recall significantly, while a loss weighting coefficient of 2 optimized the IoU. Conversely, Non-Local filtering degraded performance by discarding meaningful information, and FCRFs failed to improve connectivity, acting instead as an erosion operation. The study concludes that while off-the-shelf FCNNs are not natively efficient for road segmentation, they can yield promising results when properly tuned with spatial tolerance and loss adjustments. The findings suggest that segmentation quality does not scale linearly with network depth, indicating a need for specialized architectures designed for thin object extraction. The authors propose using these FCNNs as reliable road candidate extractors, which can then be refined into coherent road networks through graph-based reconstruction methods.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• distraction detection algorithms