An Adaptive Progressive Recursive Network for Pansharpening

Pansharpening faces inherent limitations in reconciling spatial enhancement with spectral fidelity due to the fixed interpolation-based upsampling of low-resolution multispectral (LRMS) images. To overcome this, we propose a novel Dense Adaptive Upsampling Module (DAUM) that learns pixel-wise refinement weights to optimize bilinear interpolation, thereby adaptively recovering spatial details without artificial information injection. Building upon this, we introduce ProRecFusionNet (PRFN) with dual core components: (1) a Gradual Fusion Module (GFM) enabling coarse-to-fine alignment of panchromatic (PAN) and upsampled LRMS features via hybrid CNN-Transformer interactions, and (2) a Recursive Fusion Module (RFM) iteratively suppressing artifacts through multi-scale residual optimization. This holistic framework achieves synergistic complementarity of high frequency spatial details (from PAN) and spectral signatures (from LRMS). Extensive experiments on GaoFen-2, QuickBird, and WorldView-II show that PRFN surpasses nine leading methods, with average gains of −0.916 in SAM, −0.849 in ERGAS, and +0.042 in HQNR (vs. the per-dataset mean of nine baselines). This cross-sensor robustness supports large-scale land-cover monitoring. Code: https://github.com/Zhipeng Zhao99/ProRecFusionNet-main.