Source-linked AI summary
Fully Convolutional Network for Automatic Road Extraction from Satellite Imagery
Alexander V. Buslaev, Selim S. Seferbekov, Vladimir I. Iglovikov, Alexey A. Shvets
TL;DR
The paper addresses accurate and efficient automatic road extraction from high-resolution satellite imagery, where manual segmentation is costly and road-like features complicate extraction. It proposes a U-Net-family fully convolutional network with a pre-trained ResNet-34 encoder, reports a best public leaderboard score of 0.64, and identifies label quality as an important boundary for further improvement.
Problem
Automatic road extraction from satellite imagery requires accurate pixel-level segmentation, while manual segmentation is time-consuming and error-prone.
Method
The method uses a U-Net-like encoder-decoder with a pre-trained ResNet-34 encoder, skip connections, a combined binary cross-entropy and IoU loss, and test-time augmentation.
Results
0.64 best public score on the public leaderboard for the road-extraction challenge.
Takeaways & Limitations
The proposed model provides a practical road-segmentation approach using moderate memory and test-time augmentation for improved predictions.
Takeaways & Limitations
Further robustness and precision depend on preparing higher-quality labeled masks, especially given limitations in existing annotations.
Abstract
from arXiv · showhide
Analysis of high-resolution satellite images has been an important research topic for traffic management, city planning, and road monitoring. One of the problems here is automatic and precise road extraction. From an original image, it is difficult and computationally expensive to extract roads due to presences of other road-like features with straight edges. In this paper, we propose an approach for automatic road extraction based on a fully convolutional neural network of U-net family. This network consists of ResNet-34 pre-trained on ImageNet and decoder adapted from vanilla U-Net. Based on validation results, leaderboard and our own experience this network shows superior results for the DEEPGLOBE - CVPR 2018 road extraction sub-challenge. Moreover, this network uses moderate memory that allows using just one GTX 1080 or 1080ti video cards to perform whole training and makes pretty fast predictions.
1. Introduction
High-resolution satellite imagery supports broad road mapping, but reliable pixel-level segmentation requires accurate automation. Fully convolutional and U-Net-family networks address this by producing dense segmentation maps with skip connections for localization.
- Data opportunity: High-resolution satellite imagery at 0.5–1.0 meters is sufficient for road extraction and broadly available worldwide.Its availability can support topographic databases across many regions.
- Motivation: Manual pixel-level segmentation is time-consuming and prone to mistakes and inconsistencies, motivating accurate automatic methods.Automation can segment imagery as soon as it is acquired.
- Related work: Fully convolutional networks replace fully connected layers with convolutions and upsample spatial feature maps into dense pixel-wise outputs.This approach improved segmentation accuracy over common methods on datasets such as PASCAL VOC.
- Related work: U-Net combines low-level and high-level feature maps through skip connections, enabling precise pixel-level localization.Its upsampling path propagates contextual information to higher-resolution layers.
2. Dataset
The road-extraction dataset contains high-resolution RGB satellite images paired with grayscale road masks, but annotation quality is imperfect, especially for small rural roads.
- Dataset composition: 6,226 RGB satellite images form the training data, each measuring 1024x1024 pixels at 50 cm resolution.The imagery was collected by DigitalGlobe satellites.
- Labels: Each image is paired with a grayscale road mask in which white denotes road pixels and black denotes background.A threshold of 128 is recommended for mask binarization because values may not be exactly 0 or 255.
- Label quality: The labels are imperfect because segmentation-mask annotation is costly, and small farmland roads are sometimes intentionally left unannotated.These limitations affect the reliability of the training labels.
3. Model
The model is a U-Net-like encoder-decoder for pixel-wise road segmentation, using a pre-trained ResNet-34 encoder and skip-connected decoder to combine context with precise localization.
- Architecture: The U-Net-like network uses a contracting path to capture context and an expanding path to enable precise localization.The contracting path downsamples feature maps while increasing channel counts; the expanding path restores resolution.
- Decoder: The expansive path upsamples feature maps and combines them with high-resolution contracting-path features through skip connections.The model outputs a pixel-by-pixel mask assigning a class to every pixel.
- Encoder: A pre-trained ResNet-34 serves as the encoder, while decoder blocks connect to corresponding encoder blocks in the U-Net-like architecture.The encoder begins with a 7 × 7 stride-2 convolution, max-pooling, and residual downsampling blocks.
4. Training
Training evaluates segmentation with Jaccard similarity and combines it with classification loss, while augmentation and test-time averaging improve robustness during model development and prediction.
- Evaluation: Jaccard index, or Intersection Over Union, measures similarity between predicted and reference pixel sets.It is defined for two sets A and B using their intersection and union.
- Loss: Binary segmentation uses binary cross entropy as the classification loss, while multi-class segmentation uses categorical cross entropy.The loss formulation combines this classification term with the Jaccard-based term.
- Loss: The combined loss simultaneously increases probabilities for correctly predicted pixels and the intersection between masks and predictions.The weighting parameter α is set to 0.7 from hold-out evaluation.
- Augmentation: Training applies geometric and color augmentations, including scaling, rotation, random 448x448 crops, and color transformations.These augmentations artificially increase the effective training set size.
- Optimization: Training uses Adam with learning rate 1e-4 and decay 1e-4, runs for 20k iterations, and applies spatial dropout with p = 0.3.Weights from several best iterations are saved because validation IoU varies substantially.
- Inference: Predictions use whole 1024x1024 images without padding and average four 90-degree-rotation predictions through test-time augmentation.The image side is divisible by 32, enabling the stated whole-image procedure.
5. Conclusions
The paper presents a binary road-segmentation model combining a pre-trained ResNet-34 encoder with a U-Net-like decoder and skip connections. It achieves a 0.64 public leaderboard score, while the authors identify labeling quality and further validation or augmentation as improvement paths.
- The model uses a pre-trained ResNet-34 encoder, U-Net-like decoder, skip connections, and a loss combining binary cross entropy with IoU.
- 0.64 was the model’s best public leaderboard score.
- Five-fold cross-validation, stronger image augmentation, and additional test-time transformations are proposed improvements.
- High-quality labeled masks are identified as important for robust and precise predictions.