DIDRU-CAN: Dilated Inception Deep Residual U-Net with Channel Attention for Road Extraction using High-Resolution Remote Sensing Images

Road Extraction (RE) from Remote Sensing Images (RSI) has become essential in many areas, such as automatic vehicle navigation, urban planning, traffic management, and disaster management. RE from RSI is still a challenge process due to the complicated nature of the road structure and occlusions or shadows represented by nearby trees, vehicles, and buildings. A modified Dilated Inception Deep Residual U-Net Channel Attention Model (DIDRU-CAN) is embedded, combining the advantages of the Dilated Convolutional Module, the Dilated Inception Module, and the Channel Attention Mechanism (CAM). The Dilated Inception Module works effectively in capturing the multi-scale features, so that the model can extract road networks effectively. In addition, CAM provides further improves the feature selection by showing the road-related feature and suppressing irrelevant ones. CAM improves the accuracy of road extraction by utilizing skip connections that provide detailed spatial information from the decoder and high-level context information from the encoder. The Channel Attention over the skip connection provides high-level context information from the encoder and specific spatial information from the decoder to improve road extraction accuracy. In order to evaluate the performance of the proposed model, we trained the model using Massachusetts Road dataset. After the training of the model, the results shows superior performance in terms of overall accuracy, dice coefficient, and better precision and recall measures, where recall was 82.51%, precision 85.97%, overall accuracy (OA) of 97.61%, and Dice coefficient of 76.84%. Comparative studies carried out using traditional methods show that DIDRU-CAN captures complex road networks and roads that are obscured or shadowed from buildings, cars, and trees.