Fractional derivative based weighted skip connections for satellite image road segmentation

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Road segmentation from satellite imagery is difficult due to complex backgrounds, occlusions, and optical artifacts.
• Accurate road network extraction requires integrating both local and global contextual information.

Purpose of the Study:

• To propose a novel deep learning model for road segmentation in satellite images.
• To enhance road network extraction by effectively combining local and global features using fractional calculus.

Main Methods:

• A densely connected convolutional neural network (CNN) architecture was employed.
• Fractional derivative-based weighted skip connections, utilizing the Grunwald-Letnikov definition, were integrated.
• Weights for skip connections were determined via fractional derivatives to capture non-local dependencies.

Main Results:

• The proposed model achieved superior performance on the Massachusetts Road database (MRD) and Ottawa Road database (ORD).
• On MRD, an F1-score of 0.748 and mean Intersection over Union (mIoU) of 0.787 were obtained at fractional order 0.4.
• On ORD, a mIoU of 0.9062 was achieved at fractional order 0.5, outperforming state-of-the-art methods.

Conclusions:

• Fractional derivative-based weighted skip connections effectively enhance road segmentation by incorporating memory and combining local/global features.
• The proposed method offers a robust solution for accurate and continuous road network extraction from challenging satellite imagery.
• The model demonstrates significant improvements over existing techniques on diverse road datasets.