Optimizing Road Extraction with Residual U-Net: Enhanced Training

Abstract

Computer vision and remote sensing depend heavily on extracting roads from satellite or aerial photos. This study presents a novel approach to road extraction employing a Residual U-Net architecture with integrated data augmentation techniques. The proposed method utilizes a deep learning model with residual blocks for improved feature extraction and semantic segmentation. The dataset is preprocessed, and data augmentation is applied during training to enhance model robustness. The augmentation includes random, non-critical actions and horizontal flips. The Residual U-Net architecture consists of an encoder-decoder structure with skip connections, facilitating the learning of intricate spatial dependencies. The model is trained to optimize road segmentation using a customized loss function based on the Dice Coefficient. Additionally, the code incorporates batch normalization and activation functions for improved convergence and generalization. The experimental findings show how effective the suggested strategy is for road excavation jobs. Training and validation sets are generated using a custom data generator class. The model is trained over several epochs, and its performance is evaluated on a validation set. Ground truth versus predicted value visualizations showcases the model's ability to delineate road networks accurately. This study contributes to road extraction by introducing a Residual U-Net architecture with data augmentation, providing a robust and accurate solution for road segmentation in satellite imagery.