Mutually Reinforced Attention Aggregation Network for Pansharpening

Remote sensing enables urban planners to gain macroscopic insights into urban structures and issues. Pan-sharpening plays a key role by merging multispectral and panchromatic images, combining high spatial resolution with rich spectral information. Yet, existing methods still face challenges in jointly integrating global–local context and spectral–spatial information. To address these issues, we propose the Mutually Reinforced Attention Aggregation Network (MRAAN), whose core is the Dual-Aggregation Transformer Block (DATB) embedded within a multiscale U-shaped architecture. In DATB, the Directional Spatial Attention within the Dual-Aggregation Spatial Attention captures the anisotropic spatial structures in remote sensing imagery, while the Channel-Wise Attention within the Dual-Aggregation Channel Attention models the reflectance-driven inter-band correlations. In addition, depthwise convolutions are incorporated to supplement the attention mechanisms and extract richer local information. Furthermore, the Adaptive Interaction Gate with Spatial-to-Channel and Channel-to-Spatial branches enables bidirectional information flow between the spatial and channel dimensions. Unlike previous methods that rely solely on convolutional neural networks or lightly modified Transformer variants, MRAAN achieves a superior balance between preserving spatial details and maintaining spectral consistency. Benefiting from its dual-level spatial–spectral aggregation mechanism, which consists of intrablock feature aggregation and interblock progressive aggregation, the network can effectively capture long-range dependencies while retaining essential local textures, thereby producing fused images that exhibit both high spatial clarity and strong spectral fidelity.