Multiscale spatial–spectral transformer network for hyperspectral and multispectral image fusion

Fusing hyperspectral images (HSIs) and multispectral images (MSIs) is an economic and feasible way to obtain images with both high spectral resolution and spatial resolution. Due to the limited receptive field of convolution kernels, fusion methods based on convolutional neural networks (CNNs) fail to take advantage of the global relationship in a feature map. In this paper, to exploit the powerful capability of Transformer to extract global information from the whole feature map for fusion, we propose a novel Multiscale Spatial–spectral Transformer Network (MSST-Net). The proposed network is a two-branch network that integrates the self-attention mechanism of the Transformer to extract spectral features from HSI and spatial features from MSI, respectively. Before feature extraction, cross-modality concatenations are performed to achieve cross-modality information interaction between the two branches. Then, we propose a spectral Transformer (SpeT) to extract spectral features and introduce multiscale band/patch embeddings to obtain multiscale features through SpeTs and spatial Transformers (SpaTs). To further improve the network’s performance and generalization, we proposed a self-supervised pre-training strategy, in which a masked bands autoencoder (MBAE) and a masked patches autoencoder (MPAE) are specially designed for self-supervised pre-training of the SpeTs and SpaTs. Extensive experiments on simulated and real datasets illustrate that the proposed network can achieve better performance when compared to other state-of-the-art fusion methods. The code of MSST-Net will be available at http://www.jiasen.tech/papers/ for the sake of reproducibility.