AI-based 2-m super-resolution of Sentinel-2 imagery for high-resolution SDG monitoring

作者：Decai Jin, Yaozhong Pan, Martin Brandt, Lei Fan, Yu Zhu, Wendong Qi, Huaguo Huang, Qiao Wang, Peijun Shi, Jianbo Qi · 发表于：GIScience & Remote Sensing · 年份：2026 · DOI：10.1080/15481603.2026.2640266 · 被引用次数：2 · 研究领域：Advanced Image Fusion Techniques、Advanced Image Processing Techniques、Satellite Image Processing and Photogrammetry

High-resolution Earth observation (EO) is critical for tracking spatially heterogeneous sustainable development goals (SDGs), such as cropland dynamics and urban expansion. However, persistent limitations in spatiotemporal continuity (satellite revisit gaps) and cost-efficiency (prohibitive pricing of commercial <2 m data) hinder its scalability. While deep learning-based single image super-resolution (SR) techniques offer a potential solution, their quantitative equivalence to native high-resolution data and the generalizability across geographies remain unproven. Here, we demonstrate that AI-powered SR can systematically transform freely available 10-m Sentinel-2 imagery with visible (RGB) and near infrared (NIR) bands into 2m-resolution images with RGB-NIR bands while preserving spectral-temporal fidelity. Specifically, we trained a geospatially constrained transformer-based SR framework (GeoSR) with 3.15 million km² of co-registered Gaofen-1/6 and Sentinel-2 pairs, achieving near-native performance with <1% F1-score loss in critical applications: cultivated land parcels mapping (F1-score = 0.84 vs. 0.85 for native Gaofen-1/6), urban footprint extraction (0.82 vs. 0.81), and fine-grained land use and land cover classification (0.43 vs. 0.43). Notably, the geospatial module enables large-scale generalization, retrospectively reconstructing decade-long environmental dynamics from historical archives – an unprecedented capability unattainable solely through launching new sate...