MSSDIBNet: Multiple Spatial–Spectral Dual-Injection Balance Network for Pansharpening

Pansharpening aims to generate high-resolution multispectral (HRMS) images from paired panchromatic (PAN) images and low-resolution multispectral (LRMS) images. Some deep learning models employ end-to-end skip connection to learn the differences between HRMS and LRMS images. Although these models achieve satisfactory pansharpening effects, their spectral information processing methods are inadequate, and the end-to-end residual connection may lead to inaccurate propagation of spectral information. Due to the differences in spectral range and resolution between PAN and multispectral (MS) images, direct injection of spatial information can introduce spectral distortion. To enhance spectral information fidelity and improve the injection of spatial-detail information, we propose a multiple spatial–spectral dual-injection balance network. Leveraging iterative refinement, the network performs a cascade of dual-injection stages. Each stage consists of a spatial injection subnetwork followed by its spectral counterpart. Within every stage, the spatial subnetwork first enriches spatial details; immediately afterward, the spectral subnetwork serially corrects spectral deviation. This “enhance-then-correct” synergy alternately refines sharpness and fidelity without mutual interference, ensuring balanced optimization and substantial performance gains. The spatial injection subnetwork comprises a global processing module and a local processing module, each designed to process global and local spatial information, respectively. The global processing module is effective for tasks involving small datasets. The spectral injection network emulates the structure of the spatial injection network to learn spectral details. To optimize the integration of these two distinct types of information, we developed an adaptive spatial attention module and adaptive channel attention module, and further designed spatial fusion module and channel fusion module based on them to enable different feature integration. Extensive experiments on three datasets demonstrate the superior performance and effectiveness of the proposed model.