MODIS sea ice leads detections using a U-Net
Cite this dataset
Hoffman, Jay et al. (2024). MODIS sea ice leads detections using a U-Net [Dataset]. Dryad. https://doi.org/10.5061/dryad.79cnp5hz2
Abstract
Sea ice leads are long and narrow sea ice fractures. Despite accounting for a small fraction of the Arctic surface area, leads play a critical role in the energy flux between the ocean and atmosphere. As the volume of sea ice in the Arctic has declined over recent decades, it is increasingly important to monitor the corresponding changes in sea ice leads. An approach described in Hoffman et al. 2021 uses artificial intelligence (AI) to detect sea ice leads using satellite thermal infrared window data from the Moderate Resolution Imaging Spectroradiometer (MODIS). The AI used to detect sea ice leads in satellite imagery is a particular kind of convolutional neural network, a U-Net. The originally published dataset included only a small case study of results. Here, the dataset is expanded to include the daily detection of leads since 2002 for the season between November through April.
README: MODIS Sea ice leads detections using a U-Net
https://doi.org/10.5061/dryad.79cnp5hz2
Description of the data and file structure
Seasonal files are tar files compressed with gzip compression.
Untar seasonal file to access daily hdf5 leads detection files from November through April.
Naming convention: YYYY-YYYY.SATELITE_NAME.tar.gz, where the first year is the year at the start of November and the second year is the end of April.
Daily naming convention: YYYDDDSATELITE_NAME.result.h5.
Arrays are 7025 x 7025, 1km resolution EASE-Grid 2.0
- coverage_count: The number of satellite overpasses in the day over water pixels
- lead_count: The number of satellite overpasses in which the U-Net detection score suggests a lead. A lead detection is generally considered a positive detection where the lead count is greater or equal to 3 repeat observations within a day. Lead counts of 1 or 2 are more likely false positives.
- mask: 0=non-lead, 1=lead, 2=no coverage (land and low latitude)
- Max_result: maximum lead detection score at each location
latlon.nc is a netcdf file that contains the latitude and longitude for each EASE-Grid 2.0 pixel location
Version changes
01-aug-2024: Added results for 2022-2023 and 2023-2024 seasons
Methods
AI is used to identify sea ice leads in thermal imagery from the 11 µm from MODIS (band 31, AQUA, and TERRA imagery). A U-Net detection model is run for each satellite overpass and reported as daily aggregated results. The lead detection results are projected into a standard 1 km resolution EASE-Grid 2.0 projection. The included data arrays are the daily number of satellite overpasses, the number of overpasses a lead is identified, the maximum lead detection score from the U-Net, and a lead mask for each EASE-Grid 2.0 pixel. Daily files are compressed inside November through April seasonal tar files. The daily results are recorded as hdf5 format files. For each season, the daily results from November through April for each season are combined into a new tar file with gzip compression.
Funding
National Aeronautics and Space Administration, Award: 80NSSC18K0786