Research on SAR Image and Optical Image Fusion Classification Based on Deep Learning

In order to better integrate and classify multi-source image features, we constructed a multi-source remote sensing data feature extraction and fusion classification framework U-DSDNet based on a combination of dual channel convolutional neural networks and self attention. Build two independent channel networks to extract feature information. Introducing self attention mechanism to fuse multi-source image feature information, improving classification accuracy and stability. Meanwhile, we use Sentinel series satellite data to demonstrate the classification accuracy of U-DSDNet. Through comparative analysis with UNet, DeeplabV3+, and PSPNet, the experimental results show that in the SEN-OPT-SAR dataset, the overall accuracy of the U-DSDNet model is 3.8%, 0.2%, and 1.6% higher than the three models, respectively, verifying the superiority and effectiveness of the U-DSDNet model in land classification tasks.