Super-resolution for harmonizing UAS and satellite imagery

Authors: MASRUR, ARIF - Esri; OLSEN, PEDER - Microsoft Research Lab; JACKSON, CARLAN - Alabama A & M University; Adler, Paul

Submitted to: IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 2/18/2026
Publication Date: 2/24/2026
Citation: Masrur, A., Olsen, P.A., Jackson, C., Adler, P.R. 2026. Super-resolution for harmonizing UAS and satellite imagery. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing. 19:8740-8755. https://doi.org/10.1109/JSTARS.2026.3667863.
DOI: https://doi.org/10.1109/JSTARS.2026.3667863

Interpretive Summary: An important problem in using remote sensing tools for precision crop management decisions is that freely available satellite imagery is too coarse for detailed field-level decisions, while high-resolution multispectral drone imagery is expensive and limited in coverage. Researchers developed an AI-driven data fusion framework that effectively combines low-resolution satellite data (Sentinel-2) with limited, targeted drone imagery, achieving a very high 80x improvement in image resolution. This innovation matters because it allows farmers to get high-resolution, spectrally-rich data for precision agriculture using only low-cost RGB drones and widely available satellite images, significantly reducing the cost and complexity of advanced field monitoring. This approach promises a scalable, affordable early warning system for identifying problems like nutrient stress or pest outbreaks across large farming operations.

Technical Abstract: This paper presents an end-to-end remote sensing data fusion framework that combines spectral harmonization with a super-resolution workflow to fuse low-resolution satellite (Sentinel-2) and UAS hyperspectral imagery across spatial and spectral domains. Targeted UAS data are used to align sensor responses, enabling removal of atmospheric artifacts and producing high-resolution VNIR bands. These data provide high-resolution RGB inputs (alongside Sentinel-2 VNIR) and ground truth for training super-resolution models. The method enables spectral extension of UAS RGB to the VNIR range, along with extreme upscaling of Sentinel-2 10m and 20m spectral bands to 12.5cm while preserving the spectral fidelity with PSNR values above 30dB on images of unseen crops. The upscaled data outperform both Sentinel-2 and UAS RGB baselines on cover cropping datasets from farms in Maryland and Pennsylvania. The super-resolution model generalizes across crops and seasons and supports flexible configurations, including RGB-only training and satellite-only inference. This sensor harmonization super-resolution approach advances multimodal data fusion enabling scalable, very high resolution, low-cost remote sensing applications.