Enhancing Spectral Quality through Multisource Fusion of SAR and Optical Imagery Using Deep Learning

This study evaluates the effectiveness of a multi-source satellite image fusion method using deep learning to enhance spectral feature quality in urban environments. The input data consist of synthetic aperture radar (SAR) images from Sentinel 1 and optical images from Landsat 9. Two deep learning models were implemented: patch-wise shallow convolutional neural networks (CNN) and convolutional autoencoders (CAE). Each of Red-Green-Blue optical band was fused separately with the VV/VH ratio image derived from radar data. The fusion results were assessed using RMSE, SSIM, UIQI indices, and the Pearson correlation coefficient. The CAE demonstrated better spectral reconstruction capability with lower RMSE and higher SSIM across all three bands. Conversely, the CNN model achieved higher UIQI on some bands and produced images with visually superior sharpness. Visual assessment indicated that CNN better preserves edge details and fine structures, while CAE generated smoother images but with some blurring of objects. These results suggest that deep learning-based fusion methods hold great potential for improving input image quality for urban analysis in areas affected by cloud cover.