A Remote Sensing Object Classification and Extraction Method Based on Improved U-Net

This paper addresses the problems of low accuracy and poor adaptability to complex scenes in traditional remote sensing object classification methods, and proposes a remote sensing object classification extraction method based on an improved U-Net. We retained the U-shaped symmetrical structure of U-Net and introduced the Swin Transformer block with self-attention mechanism to replace the original convolutional operation, enhancing the model's ability to extract global features. We also constructed a Bottleneck structure between the encoder and decoder to learn deep feature representations and form a new U-shaped network based on Transformer, improving the network's image semantic segmentation ability in scenarios with small targets and complexity. Experimental results show that the method proposed in this paper has an improvement of more than 9% in evaluation indicators such as recognition accuracy and recall rate compared with existing U-Net and Res-UNet methods, has high semantic segmentation accuracy and robustness, and provides new ideas and technical support for remote sensing image object classification extraction.