Satellite Image Fusion for Obtaining High Resolution Images Using Deep Neural Network

Abstract

Due to its critical function in a wide range of applications, scene categorization of high-resolution remote sensing (RS) photos has drawn increasing attention. A technique for spatiotemporal fusion using deep neural networks (DNNs) with a large amount of remote sensing data as the application background. An innovative multispectral image fusion architecture is proposed in this paper. The proposed method for fusing satellite images entails two phases, each using two neural networks. In the first stage, an adaptively weighted injection-based joints detailed approach to remotely sensed image fusion is discussed. Multispectral (MS) and panchromatic (PAN) images are used to extract spatial features using a wavelet transform. In contrast to the conventional detail injection technique, dictionary learning from the sub-images themselves is used to construct the primary joint details by sparsely representing the extracted features. To minimize spectrum distortions in the fused images while keeping spatial information, we implemented a unique loss function for this DNN. This network is known as the ’Spectral Reimbursement Network (SRN).’ Finally, using three datasets, full-reference, and limited-reference criterion, the proposed strategy is compared against several state-of-the-art methods. Experiment findings demonstrate that the suggested technique can compete in both spatial and spectral parameters.