Add AI4ArcticSeaIce dataset.
This PR adds the AI4ArcticSeaIce dataset. Rehosted to HF for faster download times (couple minutes vs ~8hours). In the original repo there is some additional useful information.
The Sea Ice Challenge Dataset contains Sentinel-1 SAR imagery, passive microwave radiometer observations
from AMSR2, and numerical weather prediction data from the ECMWF Reanalysis v5 (ERA5) dataset - all
gridded to match the Sentinel-1 SAR scenes geometrically. As label data, the dataset contains ice charts
manually produced by the ice analysts at the Greenland Ice Service and the Canadian Ice Service.
Dataset features:
* Dual-polarization SAR (HH, HV) imagery for each patch.
* Sea Ice Concentration (SIC): the percentage ratio of sea ice to open water for an area,
discretized into 11 10% bins ranging from 0% to 100%.
* Stage Of Development (SOD): type of sea ice, as proxy for ice thickness and
ease of traversing with 6 classes
* Floe size (FLOE): Classifying or segmenting distinct ice floes based on size, shape,
or other geometric properties.
Dataset format:
* each sample scene is stored in a separate .nc file
* pixel dimension of varying sizes up to ~5000pxx5000px
* 80m resolution
TODOS:
- [ ] check plotting again and improve category plotting across the different targets
- [ ] think about resizing since individual tiles can be really large
Thanks @astokholm for the creation and open-sourcing dataset. This dataset has some complexities due to all the different data modalities, so if you have any comments/corrections, it would be much appreciated.
This is the v2 dataset, right? (v1 is still around: https://data.dtu.dk/collections/AI4Arctic_Sea_Ice_Challenge_Dataset/6244065/1). Might be helpful to document this in the dataset class.
This is the v2 dataset, right? (v1 is still around: https://data.dtu.dk/collections/AI4Arctic_Sea_Ice_Challenge_Dataset/6244065/1). Might be helpful to document this in the dataset class.
Good catch, it is actually Version 3, I corrected the link above.