Pansharpening via Subpixel Convolutional Residual Network

Abstract

In this article, we propose a new pansharpening architecture called subpixel convolutional residual network to obtain high-resolution multispectral (MS) images. Different from previous works, we extract features from MS images in a low-resolution space and pay more attention to the balance of spectral and spatial information. Our architecture consists of two branches: the feature extraction branch and the residual branch. The former adopts a four-layer convolutional network to extract features, and then upsamples the feature maps using a subpixel convolution layer. For the latter, we combine the nearest neighbor interpolation and guided filter to yield a preliminary image with fundamental spectral and spatial information. With the outputs of the two branches, we can merge them and yield a pansharpened image. The proposed method was compared with several representative methods. The experimental results demonstrate that our method achieves high fusion accuracy while maintaining a good balance between the spectral and the spatial resolution.