CTF-SRNet: A Cross-Modal Text and Feature Modulation Network for Remote Sensing Image Super-Resolution

Abstract

The remote sensing image super-resolution (RSISR) has achieved significant progress in recent years. However, effectively recovering fine-grained structural details while maintaining high-level semantic consistency remains a challenging problem. Most existing RSISR methods rely solely on visual information from low-resolution (LR) inputs, ignoring auxiliary semantic cues that are crucial for reconstructing complex scenes in remote sensing imagery. To address these limitations, we propose a novel cross-modal text and feature modulation network, termed CTF-SRNet, for RSISR. Unlike conventional approaches, the CTF-SRNet explicitly incorporates textual semantic information to guide the super-resolution (SR) process. The proposed framework consists of a multiscale channel-modulated image encoder for robust feature extraction, an image–text feature interaction block (ITFIB) that injects text-guided semantic priors into visual features, and a feature modulation block (FMB) that adaptively refines cross-modal representations to enhance structural detail recovery. In addition, a multilevel loss supervision strategy is employed to progressively constrain intermediate outputs, encouraging consistent texture reconstruction across different scales. Extensive experiments on multiple remote sensing benchmarks, including UC Merced, RSSCN7, Aerial Image Dataset (AID), Dataset for Object Detection in Aerial Images (DOTA), and DIOR datasets, demonstrate that CTF-SRNet consistently outperforms the state-of-the-art RSISR methods in terms of both quantitative metrics and visual quality, while maintaining a lightweight model design. Furthermore, the downstream classification experiments validate that the proposed cross-modal SR framework effectively enhances discriminative structural and semantic information, highlighting its practical value for real-world remote sensing applications.