Skip to content
Open access

Leveraging Pretrained Natural Image Priors for Parameter-Efficient Remote Sensing Superresolution

2026 · IEEE Geoscience and Remote Sensing Letters · Vol 23, pp. 6021705-6021705 · 0 citations · 22 references

Abstract

Remote sensing image superresolution (RSISR) is pivotal for enhancing the utility of satellite imagery in applications such as surveillance and urban planning. However, existing RSISR methods typically train lightweight models from scratch, failing to fully exploit the rich generative priors encapsulated in large-scale natural image superresolution (SR) models. Conversely, fully fine-tuning these large models for the remote sensing domain is computationally prohibitive and prone to overfitting under limited training data. To bridge this gap, we propose PET-SR, a parameter-efficient adapter-based transfer learning framework for RSISR. PET-SR freezes the pretrained backbone of a state-of-the-art natural image SR model and injects task-adaptive Transformer blocks (TABs) into the last Transformer layer of each residual stage to enable efficient domain adaptation. Each TAB is equipped with a hierarchical convolutional experts (HCEs) adapter, which adopts a mixture-of-experts (MoE)–style dynamic gating mechanism to aggregate multiscale convolutional experts, effectively handling the extreme scale variations inherent in remote sensing scenes. Extensive experiments on the UCMerced and AID datasets demonstrate that PET-SR achieves competitive performance under the evaluated bicubic degradation protocol and introduces only a small number of additional trainable parameters.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.