PAMNet: Heterogeneous Progressive Feature Modeling for Pansharpening

Pansharpening aims to fuse a high-spatial-resolution panchromatic (PAN) image with a low-spatial-resolution multispectral (LRMS) image to generate a high-resolution multispectral (HRMS) image with both high spectral fidelity and rich spatial details. Existing deep pansharpening methods still struggle to jointly handle long-range structural dependency modeling, spatial–spectral feature transformation, and local detail enhancement within a unified framework. To address this issue, we propose PAMNet, a progressive pansharpening network based on a heterogeneous feature modeling strategy. Specifically, PAMNet employs a feature modeling block composed of an adaptive scanning Mamba block (ASMB), a frequency-domain transformation block (FTB), and a PAN-guided detail enhancement block (PDEB) to cooperatively capture long-range dependencies, refine fused spatial–spectral representations, and compensate local details from PAN priors. These complementary components are organized in a coarse-to-fine reconstruction framework to progressively improve structural consistency and detail reconstruction. Experiments on the WorldView-3 (WV3), GaoFen-2 (GF2), and QuickBird (QB) datasets demonstrate that PAMNet achieves favorable and consistent performance in both reduced- and full-resolution pansharpening tasks.