Content-Aware Degradation Modeling for Supervised Remote Sensing Pansharpening
Abstract
Pansharpening aims to reconstruct a high-resolution multispectral (HRMS) image by fusing a low-resolution multispectral image with a high-resolution panchromatic image. Most supervised pansharpening methods rely on Wald’s protocol to construct training pairs, where fixed modulation transfer function (MTF)-based degradation is commonly used to simulate reduced-resolution observations. However, such predefined degradation is content-agnostic and may not reflect effective degradation variations caused by scene-dependent structures, imaging conditions, and preprocessing operations, leading to a discrepancy between synthetic training pairs and real full-resolution observations. To address this issue, we propose a content-aware degradation modeling framework for supervised remote sensing pansharpening. Instead of using a fixed degradation kernel, the proposed framework learns a sample-adaptive degradation surrogate from paired panchromatic (PAN) and low-resolution multispectral (LRMS) observations. Specifically, a wavelet-based scene encoder (WSE) extracts frequency-aware scene representations from the aligned PAN image, and a mixture-of-experts (MoE) blur module generates valid degradation kernels by softly aggregating multiple input-conditioned expert kernels. A learnable spectral response mapping is further introduced to establish consistency between the degraded PAN image and the single-band response of multispectral observations. The learned degradation model serves as a training-data generation strategy applicable to different downstream pansharpening networks. We further instantiate the complete pipeline with a PAN-guided spatial–spectral fusion network, and experiments on four benchmark satellite datasets demonstrate that the learned degradation strategy improves full-resolution generalization across representative backbones while the complete pipeline achieves competitive full-resolution performance. The source code will be released at: https://github.com/19suede/CAD