Hierarchical Siamese-Swin Transformer Network with Deep Difference Discrimination Model for Aerial Image Change Detection

In recent years, aerial imagery has become an indispensable tool in urban planning, environmental monitoring, disaster management, and surveillance. The ability to detect and analyze the changes in aerial images is crucial for effective decision-making processes and can aid in timely decision-making and resource allocation. The change detection algorithms aim to identify changes between bi-temporal images, providing valuable insights into dynamic phenomena and human activities. With the advent of deep learning, the datadriven approaches leverage the representation learning capabilities of neural networks to extract relevant features from raw data automatically. In this research work, a framework called the hierarchical Siamese-Shifted windows (Swin) transformer network with a deep difference discrimination model for effective aerial image change detection is proposed. This framework combines the hierarchical feature extraction, Siamese architecture for paired image comparison, Swin Transformer for handling spatial dependencies, and a deep difference discrimination model for capturing nuanced changes between images. The Siamese architecture facilitates the comparison of pairs of images, enabling the model to focus on detecting changes rather than absolute pixel values. The Swin Transformer leverages self-attention mechanisms to model long-range spatial dependencies effectively, enhancing the framework’s ability to discern meaningful changes amidst complex backgrounds. The deep difference model also enhances the framework’s sensitivity to subtle variations between images, further improving its change detection performance. The proposed change detection framework significantly improves performance in terms of Precision, Recall, and Overall Accuracy (OA) metrics, achieving 93.79%, 91.35%, and 99.15% on the LEVIR-CD dataset and 91.12%, 90.47%, and 97.89% on the DSIFN-CD dataset, respectively.