This study tested how accurate satellite measurements of snow brightness (reflectance) are in different snowy environments—mountains (alpine), open plains (prairie), and forests (taiga). Researchers used UAVs (drones) equipped with advanced hyperspectral cameras to collect very detailed surface reflectance data in Montana (USA) and Finland. They compared these data with reflectance values from Landsat 8 and 9 satellites, focusing on six key spectral bands. The study found that Landsat often overestimated or underestimated snow reflectance, with some errors as high as 17%. A machine learning model (CNN) was used to isolate snow-only areas and showed that snow reflectance can vary a lot within a scene—something satellites often miss.

The team also built and tested a custom drone-mounted sensor designed to match the wavelengths used by Landsat Band 7 and Sentinel-2 Band 12. This sensor was carefully calibrated in the lab and filed tested to assess the accuracy of Landsat Band 7 measurements. Differences between drone and satellite readings ranged from -27% to +47%.

Overall, the study shows that current satellite data may not always be accurate over snow and recommends using UAVs and machine learning to better validate and improve satellite snow reflectance products.