Change Detection of Urban GIS Maps Using Multi-scale U-Net-Based Attention Neural Network Architecture

The unplanned expansion of urban settlements has caused adverse effects on living environments. High quality of remote sensing data is available and monitoring building changes is a key aspect in the development of urban settlements. Existing methods for change detection fails to capture time-dependent features from satellite images, face accuracy and generalization change detection difficulties. This paper proposes change detection in urban landscapes especially buildings using multi-scale attention neural network architecture (MAN-Net). It uses enhanced U-Net for multi-scale segmentation, applies multi-scale modules to extract features at different scales at various intervals of time to generate feature maps with different scales. The obtained feature maps are passed on to attention module to enhance and suppress the feature maps depending on the significance of the feature map. This task is done automatically and the mechanism allows the network to focus on time-dependent features by computing interrelations between pixels, reducing the influence of ineffective features. Finally, the obtained weighted attention maps are sent to distance metric which yields the building change. MAN-Net model reached an F1 score of 0.89, accuracy is around 97%, which signifies that by accompanying feature extraction with an attention mechanism, yields better results compared to single models.