In August 2023 the group was invited by Olds College of Agriculture & Technology to present at AgSmart, the college’s agriculture-technology expo, held 1-2 August 2023. Joshua Billson (MSc) and Karansinh Padhiar (MSc) delivered the talk, “Enhanced satellite data brings value to agriculture.” The wider project team included Fei Yang (PhD), supervised by Professor Irene Cheng, with collaborator Dr. Alex Melnitchouck, a former CTO at Olds College.
The trade-off farmers live with
The presentation began with the basics of remote sensing for agriculture: how the spectral signature of a crop can be read from above, and how that information helps analyse plant health across a field. It then introduced the central constraint. A single sensor generally trades spatial, temporal and spectral resolution against one another. Freely available satellites such as Landsat 8 (9 bands at 15, 30 and 100 m) and Sentinel-2 (13 bands at 10, 20 and 60 m) give broad, frequent coverage but at coarser detail, while hyperspectral instruments capture hundreds of narrow bands at the cost of other resolutions. The group’s work aims to relax that trade-off with AI, getting more out of the satellite data producers already have.
Reconstructing the hyperspectral view
The first result, hyperspectral reconstruction, co-registers a multispectral image with a hyperspectral image, then trains an AI model to synthesise hyperspectral bands from new multispectral input. Once trained, the model can add hyperspectral-like spectral richness to imagery that started out multispectral.
The reconstruction pipeline: co-registered multispectral and hyperspectral image pairs train an AI model that then synthesises hyperspectral data from new multispectral input. Figure by the ROSS group.
The point of the exercise is the spectral signature: the full reflectance curve across wavelengths that distinguishes one surface from another. Starting from the handful of broad bands a multispectral satellite records, the model recovers a continuous curve close to what a true hyperspectral sensor like DESIS would have measured.
Reconstructing the spectral signature: from the coarse Sentinel and Landsat bands (top), the model recovers a full reflectance curve (red, bottom) that closely tracks the DESIS hyperspectral reference (purple, middle). Figure by the ROSS group.
How close does it get?
The team also showed the error, band by band. Across most of the 400 to 1000 nm range the mean absolute error in reflectance stays low, rising only at the far near-infrared edge where the input satellites carry little information. The pattern held for both Landsat 8 and Sentinel-2 inputs.
Reconstruction error stays low across most bands for both Landsat 8 and Sentinel-2 inputs, climbing only at the far near-infrared edge. Figure by the ROSS group.
Sharper images, not just richer spectra
The second result, spatial enhancement, works in the other direction: it recovers finer spatial detail. The cover image illustrates this, sharpening DESIS hyperspectral imagery from 30 m to an effective 10 m resolution, so field boundaries and watercourses that were blurred in the original become legible. DESIS is a spaceborne hyperspectral imager carrying roughly 235 bands at about 30 m ground sampling.
Both methods were demonstrated over a study area centred on the town of Olds, Alberta, using imagery captured in August 2020 and 2021. The scene mixed farmland, trees, water bodies and built-up land, so the models had to generalise across very different surfaces rather than a single uniform crop. Presented under the group’s Agriculture and Environment theme, the talk connected advanced remote sensing to practical value for farms: getting more spectral and spatial information out of existing satellite data, without waiting for costlier instruments.