Corresponding Author: Palvi Sharma
Road Extraction (RE) from Remote Sensing Images (RSI) has become essential in various areas, including automatic vehicle navigation, urban planning, traffic management, and disaster management. RE from RSI remains a challenging process owing to the complex nature of road structure and occlusions from nearby trees, vehicles, and buildings. A modified Dilated Inception Deep Residual U-Net Channel Attention Model (DIDRU-CAN) is proposed, combining the Dilated Convolutional Module, Dilated Inception Module, and Channel Attention Mechanism (CAM) to capture multi-scale features and improve feature selection. Trained on the Massachusetts Road dataset, the model achieved recall of 82.51%, precision of 85.97%, overall accuracy of 97.61%, and a Dice coefficient of 76.84%, with comparative studies showing DIDRU-CAN captures complex road networks obscured or shadowed by buildings, cars, and trees.