Swin Transformer with Convolution Neural Network for City Change Detection from Satellite Images

Satellite images of high resolution are more accessible for semantic understanding of urban multi-temporal. Current approaches failed to fine-grained features and possess poor ability to acquire long-range space time information, and therefore the micro changes are missed and the edges of change types are smoothed. Therefore, in this research Swin Transformer with Convolution Neural Network (ST-CNN) is proposed for city change detection from satellite images. The ST -CNN utilizes hierarchical representations with self-attention mechanisms that improve feature extraction and spatial aspects of the understanding by changes in urban form. Besides capturing whole local-global context images, ST architectures also considered long-distance dependencies in a way that does not limit them to traditional CNN architectures. The preprocessing uses three methods such as augmentation, Bilateral Filtering (BF) and normalization. The data quality is improved through image augmentation which included horizontal flipping, rotation and brightness adjustment procedures. BF uses smooth images while keeping edges intact by carrying out non-linear pixel value combinations from nearby spots. Normalization process transforms RGB values through a scale adjustment to establish a standard range which reduces illumination effects while improving feature extraction quality. The ST-CNN achieves better results of 99.58% accuracy, 98.27% recall, 98.62% precision, 98.56% f1-score and 95.13% Intersection over Union (IoU) for Hi-UCD dataset which is better than baseline models.