Convolutional neural networks for road surface classification on aerial imagery.

Area of Science:

• Remote Sensing
• Computer Vision
• Geospatial Analysis

Background:

• Roads are ubiquitous modifications by humans, crucial for transportation and societal functions.
• Road surface type significantly impacts their functionality, yet automated classification from remote sensing data is underexplored.
• Understanding road surface composition is vital for infrastructure management and urban planning.

Purpose of the Study:

• To investigate the effectiveness of convolutional neural networks (CNNs) for automated road surface classification using aerial imagery.
• To compare the performance of different CNN architectures (FCN, U-Net, SegNet, DeepLabv3+) for semantic segmentation of road surfaces.
• To evaluate the impact of near-infrared bands and overfitting strategies on classification accuracy.

Main Methods:

• Utilized aerial imagery with 10 cm spatial resolution for road surface classification.
• Applied and compared several CNN models: Fully Convolutional Network (FCN), U-Net, SegNet, and DeepLabv3+.
• Assessed the influence of adding a near-infrared band and employed overfitting strategies like dropout and data augmentation.
• Compared CNN performance against single-pixel based random forests.

Main Results:

• CNNs successfully distinguished between compact (asphalt, concrete) and modular (paving stones, tiles) road and sidewalk surfaces.
• U-Net emerged as the top-performing model, achieving an overall accuracy of nearly 92%.
• CNNs demonstrated a significant advantage over random forests due to their context-aware feature extraction, with U-Net achieving ~25% higher accuracy.

Conclusions:

• Convolutional neural networks, particularly U-Net, are highly suitable for automated semantic segmentation of road surfaces from aerial imagery.
• The context-aware nature of CNNs provides superior performance compared to traditional single-pixel methods.
• Careful consideration of near-infrared band inclusion and overfitting strategies is necessary for optimal results in road surface classification.