Harmonizing Antarctic Sea-Ice Datasets
A reproducible R tutorial for cross-sensor harmonization of passive-microwave sea-ice concentration
Overview
This tutorial shows how to combine two satellite sea-ice products that disagree on spatial resolution and time span into a single, internally consistent record you can analyze as one dataset. The worked example rebuilds the harmonized sea-ice dataset behind Wethington and Lynch (Wethington and Lynch 2026), but the method is general. It applies to any two gridded sea-ice concentration products that overlap in time, in either hemisphere.
The problem this solves
Satellite passive microwave sensors have measured sea-ice concentration since 1979. Sea-ice concentration (hereafter SIC) is the fraction of a grid cell covered by ice, reported from 0 to 100 percent. Two facts about the satellite record are in tension. The long record is coarse: the U.S. National Snow and Ice Data Center (NSIDC) Sea Ice Index runs from 1979 to the present at 25 km resolution. The fine record is short: the Advanced Microwave Scanning Radiometer (AMSR) products resolve 12.5 km but only begin in the 2000s. Many analyses want both at once, the spatial detail of the fine product and the temporal depth of the coarse one.
Harmonization bridges the two. Rather than splicing the records and accepting a discontinuity where the sensors change, we learn the relationship between the products during the years they overlap, then use that relationship to express the entire coarse record in terms of the fine one. The result is one continuous dataset whose values are consistent across the full period.
What you will build
By the end you will have produced a daily harmonized SIC dataset for the western Antarctic Peninsula spanning 1979 to 2023, expressed on the 12.5 km AMSR grid, together with the figures that validate it against the real AMSR observations over the overlap years. The final chapter then uses that dataset to compute the standard sea-ice metrics (extent, concentration, persistence, and duration) and to estimate a long-term winter trend.
The method in brief
The core idea is a cross-sensor transfer function. We pair the coarse and fine products pixel by pixel and day by day over their overlapping years, train a model that predicts the fine-product SIC from the coarse-product SIC, and then apply that model to the full coarse record. Following the paper, the model is a small artificial neural network (ANN) fit with the nnet package. Because a one-input ANN invites the question of whether something simpler would do, we also fit a linear and a quantile-mapping baseline and compare them, so the choice of model is something you can see rather than take on faith.
Who this is for
You should be comfortable reading and running R and manipulating data frames. Prior geospatial experience helps but is not assumed; we explain each raster operation as it appears, using the terra package throughout. The statistics stay at the level of a regression and a small neural network, both of which we build up step by step.
How the data works
The repository ships a small bundled sample under data/sample/, drawn from the real NSIDC products and clipped to the study area, so that every chapter runs end to end in a few minutes with no account and no large download. For real work you fetch the full products from NSIDC, which requires a free Earthdata login; 2 Getting the data from NSIDC walks through that. A single environment variable, SIH_MODE, switches the code between the bundled sample and your own full download, so the same chapters serve both the quick read-through and the production run.
A few steps in the full pipeline are genuinely heavy, including the multi-decade download and the deployment of the trained model across the entire 1979 to 2023 record. Where that is the case, the chapter shows the code and the result it produced but does not re-run it during the site build. Each such block is labeled so you know what executes on the sample and what is shown frozen.
Scope
This tutorial covers the harmonization itself and the sea-ice analysis that immediately follows from it. It stops at sea-ice metrics and a trend example. The penguin population modeling that motivated the original paper is deliberately out of scope, so that the method stands on its own and transfers to other questions.
Citation and reuse
If you use this tutorial or the harmonization method, please cite the paper:
Wethington, M. and Lynch, H. J. (2026). Winter sea-ice decline drives gentoo penguin population expansion along the western Antarctic Peninsula. Marine Ecology Progress Series 787, 1-19. https://doi.org/10.3354/meps15149
The code is released under the MIT License and the written content under CC BY 4.0. The source repository, including the original research code this tutorial is adapted from, is linked in the sidebar.