Adaptive Swin Transformer with Multiscale Attention for Urban and Rural Remote-Sensing Image Classification

Urban and rural planning faces challenges in processing complex image data. Traditional methods, relying on manual feature extraction and rule-based models, struggle with large-scale remote-sensing data. Although deep learning methods have improved local feature extraction, they still face difficulties in handling complex backgrounds and dynamic changes. This paper proposes an adaptive Swin Transformer model that enhances the Swin Transformer’s feature extraction capabilities by incorporating multiscale attention and adaptive feature fusion. The multiscale attention module captures both fine-grained and large-scale features, while the adaptive feature fusion strategy dynamically adjusts feature weights based on scene complexity, ensuring optimal feature extraction across various planning scenarios. Experiments across multiple data sets demonstrate that the proposed model significantly improves feature extraction, classification accuracy, and adaptability in diverse urban and rural contexts, particularly in complex and changing environments. Our enhancements empower urban and rural planners and policymakers to conduct more precise evaluations of remote-sensing imagery. This facilitates the creation of well-informed planning strategies and the efficient allocation of resources in complex, real-world scenarios.