Road Network Identification in Satellite Imagery using Deep Learning

In the pursuit of enhancing road network detection from satellite imagery, this paper implements semantic segmentation techniques on the DeepGlobe Road Extraction dataset. Our approach entailed preprocessing the data with one-hot encoding for masks, followed by a training and validation split, and augmenting the training set to improve model robustness. Two prominent image segmentation architectures, namely DeepLabV3+ and U-Net, are compared and analyzed. DeepLabV3+ with a ResNet50 encoder is utilized for its advanced encoder-decoder structure, adept at capturing significant features and delineating road boundaries. A U-Net model is constructed, emphasizing a symmetrical encoderdecoder design. Both models underwent training with suitable parameters, loss functions, and learning rate schedules. Model performance is evaluated on the validation set, leveraging metrics such as the Intersection over Union score and Dice loss. The results, including normal accuracy metrics and binary cross entropy loss, are visualized to compare the original satellite imagery, ground truth masks, and predicted segmentation. Performance analysis of both the deep learning models show promising results, thus highlighting the effectiveness of the applied methodologies in road network detection and segmentation.