Source-linked AI summary
Road Segmentation in SAR Satellite Images with Deep Fully-Convolutional Neural Networks
Corentin Henry, Seyed Majid Azimi, Nina Merkle
TL;DR
Road extraction in SAR imagery is difficult because roads resemble other objects and prior methods commonly depend on classical segmentation. This paper evaluates FCNNs adapted for thin roads, finding promising extraction results while identifying generalization and annotation limits.
Problem
Few studies evaluate deep learning for SAR road segmentation, although roads are difficult to distinguish from similar targets and existing methods commonly rely on classical candidate extraction.
Method
The paper evaluates three FCNNs for high-resolution SAR road segmentation, adapting training with spatial tolerance and class-weighted MSE loss.
Results
44.98% IoU was achieved at tmax = 4px, while DeepLabv3+ converged 2.4 times faster than FCN-8s and produced smoother, less noisy predictions.
Takeaways & Limitations
Well-fitted FCNNs can serve as effective road candidate extractors in high-resolution SAR images.
Takeaways & Limitations
The models struggled to generalize across varied patterns, missed many less-visible roads, and were affected by limitations in the annotation scheme.
Abstract
from arXiv · showhide
Remote sensing is extensively used in cartography. As transportation networks grow and change, extracting roads automatically from satellite images is crucial to keep maps up-to-date. Synthetic Aperture Radar satellites can provide high resolution topographical maps. However roads are difficult to identify in these data as they look visually similar to targets such as rivers and railways. Most road extraction methods on Synthetic Aperture Radar images still rely on a prior segmentation performed by classical computer vision algorithms. Few works study the potential of deep learning techniques, despite their successful applications to optical imagery. This letter presents an evaluation of Fully-Convolutional Neural Networks for road segmentation in SAR images. We study the relative performance of early and state-of-the-art networks after carefully enhancing their sensitivity towards thin objects by adding spatial tolerance rules. Our models shows promising results, successfully extracting most of the roads in our test dataset. This shows that, although Fully-Convolutional Neural Networks natively lack efficiency for road segmentation, they are capable of good results if properly tuned. As the segmentation quality does not scale well with the increasing depth of the networks, the design of specialized architectures for roads extraction should yield better performances.
c⃝2018 IEEE
The document is a pre-print accepted for publication by IEEE and includes restrictions governing reuse and redistribution.
- The paper is identified as an IEEE Geoscience and Remote Sensing Letters pre-print accepted in 2018.
- The document provides a DOI for the accepted paper.
- IEEE permits personal use of the material.
- Other uses, including reprinting, redistribution, and reuse in other works, require permission from IEEE.
I. INTRODUCTION
Road extraction from SAR imagery is important for updating transportation maps, but roads resemble other objects and existing methods often depend on classical candidate segmentation. This letter evaluates FCNNs adapted for thin-object road segmentation.
- Urban growth and changing transportation networks require frequent road-map updates for monitoring, geolocation, and disaster relief.
- SAR imagery supports high-resolution topographical mapping, but roads can resemble railway tracks, tree hedges, and rivers.
- Conventional SAR road extraction typically detects candidates, constructs a topological graph, and optimizes a coherent network using contextual information.
- Most existing methods depend on efficient road-candidate extraction, while a CRF approach jointly extracts candidates and applies topological constraints.
- The letter evaluates FCN8s, Deep Residual U-Net, and DeepLabv3+ on high-resolution SAR datasets using class-weighted MSE and spatial-tolerance controls.
A. Segmentation with Fully-Convolutional Neural Networks
FCNNs enable pixel-wise segmentation of large images because convolutional layers remove fixed input-size constraints and adaptive bottlenecks preserve broad context. The section introduces FCN-8s as one implemented architecture.
- FCNNs are suited to large-scale segmentation because they process images of any size and incorporate wider contextual information.
- Adaptive bottleneck layers connect a DCNN encoder with the up-sampling component of the segmentation network.
- FCNNs replace classification networks’ fixed-size fully connected layers with convolutional layers, removing the upstream maximum input-size constraint.
- The implemented FCN-8s uses a VGG-19 backbone and skip connections that fuse early-layer resolution into up-sampling to improve spatial accuracy.
B. Adjusting the FCNNs for road segmentation
Because roads are thin and easily outweighted by clutter, the FCNNs are adapted through tolerant labels, regression-style outputs, and class-weighted loss. These changes emphasize road pixels during training.
- Road labels precisely cover visible road outlines and embankments to address class imbalance caused by roads’ thin appearance and surrounding clutter.
- The task becomes binary regression, with each pixel assigned a continuous estimate of how likely it is to be a road.
- The adapted FCNNs use sigmoid logits and replace cross-entropy with Mean Squared Error loss.
- Background pixels receive weighting coefficient 1, while several road-class weighting coefficients are tested to address the two-class imbalance.
- The loss uses tolerant-ground-truth values and sigmoid predictions, with per-pixel weights determined by whether binary ground truth marks road or background.
C. Applying pre- and post-processing
The study examines non-local filtering and fully-connected conditional random field post-processing to improve SAR image homogeneity and segmentation-map consistency.
- Non-local filtering improves feature homogeneity in SAR images and can mitigate speckle noise.
- Fully-connected conditional random fields refine boundaries between object regions and improve the consistency of FCNN segmentation maps.
- The study evaluates both pre-processing with non-local filtering and post-processing with fully-connected conditional random fields.
A. Experimental procedure
The experiments use custom TerraSAR-X data, train FCNNs from scratch, and evaluate thresholded binary road masks with segmentation-focused metrics.
- Dataset: The custom dataset uses high-resolution TerraSAR-X spotlight images with manually labeled major roads, country roads, and dirt paths merged into binary ground truth.Road labels were smoothed before evaluation.
- Dataset: Urban areas were excluded because roads and building edges were difficult to distinguish, limiting the dataset to mostly non-urban regions.
- Training: All networks were trained from scratch because ImageNet-pretrained weights caused a considerable performance drop on SAR images.The authors attribute this to the different nature of SAR and optical imagery.
- Evaluation metrics: Predictions were thresholded at 0.5 and compared with binary ground truth using IoU, precision, and recall.
- Evaluation metrics: Accuracy was considered unsuitable because roads occupy around 5% of ground-truth pixels, allowing background-only predictions to reach 95% accuracy.
B. Discussion
The experiments compare tolerance, loss weighting, post-processing, and three FCNN architectures across test areas. DeepLabv3+ offers smoother predictions and faster convergence, while generalization and annotation quality remain limiting factors.
- B. Discussion: 44.98% IoU is achieved with tmax = 4px, balancing an 8% recall gain against less than 10% precision loss versus tmax = 0px.Tolerance values of 0, 1, 2, 4, and 8px were tested.
- B. Discussion: 45.46% IoU is reached with loss weighting coefficient 2, yielding a maximum 0.48% IoU gain.Approximately 5% of ground-truth pixels are roads, motivating coefficients of 1, 2, 4, and 8.
- B. Discussion: NL-filtering decreases IoU by 8%, plausibly because FCNN filtering already preserves spatial information through down-sampling and convolution.The authors suggest that NL-filtering discards meaningful information.
- B. Discussion: FCRF post-processing fails to improve connectivity and behaves close to erosion, removing both spurious and valid predictions.Already smooth and regular segmentations limit the benefits of FCRFs.
- B. Discussion: DeepLabv3+ and FCN-8s outperform Deep Residual U-Net, while DeepLabv3+ converges 2.4 times faster and produces smoother, less noisy predictions.Performance decreases by at most 4.73% IoU when models are applied outside the training region.
- B. Discussion: The models generalize across unrelated areas and produce mostly continuous, low-noise predictions, supporting subsequent road-graph construction without candidate pre-processing.The authors identify annotation thickness, irregular borders, missed faint roads, false detections, and disconnected intersections as remaining issues.
IV. CONCLUSION
The conclusion finds that tuned FCNNs can extract roads effectively from SAR imagery, despite native inefficiency for thin objects. It also identifies specialized road-segmentation architectures as a direction for improving performance.
- IV. CONCLUSION: FCNNs are effective for road extraction from SAR images when enhanced with tolerance to spatially small mistakes.The approach uses a spatial tolerance rule to improve sensitivity to thin roads.
- IV. CONCLUSION: 44% average IoU is achieved by modified DeepLabv3+ using class-rebalanced Mean Squared Error regression loss across the test sets.The loss is rebalanced toward the road class.
- IV. CONCLUSION: FCN-8s reaches scores close to DeepLabv3+ despite being substantially shallower, but its predictions are noisier and less smooth.DeepLabv3+ is therefore described as the more robust road candidate extractor.
- IV. CONCLUSION: The narrow performance gap between architectures indicates a need for FCNN designs specialized for road segmentation.Such architectures are proposed as a route toward better performance.