Road Network Detection

Introduction

Satellite imagery as deep learning research regarding vision is experiencing a huge rate of improvement starts gaining attraction meaning that a variety of satellite imaging based applications that were impossible to realize a few years ago are now feasible. Machine learning algorithms are more reliable and capable of completing a variety of tasks ,given satellite data, such as area segmentation , construction monitoring , vessel detection etc..

The Problem

The problem was to predict new roads on areas of China from high resolution satellite images on China urban areas.

Solution

The proposed system is based on a U-net CNN architecture for image segmentation. The algorithm for image segmentation has two phases: the learning phase and the operating phase. The trained model operates on satellite images and predicts with high accuracy all the visible roads and removes the already existing roads from the predicted ones. That way it highlights only the new uncharted roads.

Process Followed

• Data collection, annotation creation and organization
• Data preprocessing and augmentation application.
• Model training and selection of the best.
• Detections are saved and corrected for further improvement.

Technologies

• Programming languages:
  • Python
• Packages:
  • Pytorch
  • OpenCv
  • Numpy
  • Scikit-Image
  • Pandas

Benefits

• High image segmentation accuracy.
  • more than 90%.
• Automatic digital map data update from satellite imaging.