Methodology:

The IceNet model used in this dataset is an ensemble of 10 individual U-Net deep learning models, whose forecasts are averaged to compute the ensemble mean. IceNet's daily inputs comprise sea ice concentration (SIC), 11 climate variables, statistical SIC forecasts, and metadata. IceNet is trained to forecast the next 93 days of daily SIC maps at 25 km resolution. IceNet's training data comprises climate simulations covering 1850-2100 and observational (reanalysis and satellite) data from 1979-2015. Observational data from 2015-2019 was used to validate the model during training, and 2020-2022 was used as the test set. The IceNet forecasts for the 2020-2022 test date ranges are provided in netCDF format. The final model weights used to generate the test forecasts are provided as Tensorflow neural network files (HDF5). We use held-out data from 2015-2022 to compare IceNet's forecasting skill to a dynamical model (ECMWF SEAS5). These results are provided in a summary CSV file.

Satellite telemetry records from collared caribou were analysed alongside SIC satellite data products to generate caribou/SIC relationships. Satellite telemetry data is held by the Government of Nunavut and is not openly available. Here we present an overview of the analysis and primary results generated, removing specifics relating to the telemetry dataset where required. Caribou sea ice crossing-start points were defined by creating an outline of Victoria Island and assessing when and where each caribou left the island and began travelling over sea ice. This Victoria Island coastline is given as a Shapefile, and was derived from OpenStreetMap data. SIC values at crossing-start points were extracted from the Ocean and Sea Ice Satellite Application Facility (OSI-SAF) and the Advanced Microwave Scanning Radiometer 2 (AMSR2) passive microwave satellite records, provided as CSV files. These OSI-SAF and AMSR2 SIC records were used to produce a mapping between a SIC threshold and the percent of collared caribou which had migrated before that SIC threshold. These sic/percent-migrated mappings are provided as CSV files, and can be used to convert IceNet forecasts to expected caribou sea ice crossing-start dates. The full process for generating these sets of results are described in detail in the Methods section and supplementary material of the associated paper. The GitHub repository also contains code associated with the analysis.

Data collection:

IceNet forecasts were generated using the Python library icenet v0.2.6 and can be recreated with icenet-pipeline and icenet@v0.2.x. Results in the analysis were generated in Python v3.9 using the libraries listed in https://github.com/EllieBowler/icenet-caribou-paper.

Data quality:

IceNet makes predictions based on ERA5 reanalysis data and OSI-SAF SIC data - for information on their errors see their associated documentation. IceNet's SIC values were set to zero over a land mask.