Multibranch Cnn-Based Pansharpening With Skip Connection

Recent research on multispectral (MS) and panchromatic (PN) images fusion that known as pansharpening has progressed with the development of Convolutional Neural Networks (CNN). However, the states-of-the-arts methods are principally based on simple networks with shallow architectures that may limit their performance. Recently, residual learning (ResNet) exhibit improved performance in many application domains. At the same time, numerous upsampling methods were developed, from the classical interpolation to learning based methods. In this paper, ResNet is employed to make the full exploitation of the high nonlinearity of CNN. Moreover, an ensemble of upsampling methods were joined in the developed Multibranch Pansharpening Network (MPN) that prove good performance to reconstruct high-resolution MS images. The proposed approach is applied to QuickBird data, its efficiency is assessed with universally used performance criteria in spatial and spectral domains. Experimental results of the proposed approach show better spatial performance than classical methods and competitive spectral performance against the state-of-the-art approaches.