Greenland ice

An undergraduate research position at Dalhousie, supervised by Zahra Sadeghi. My part of the group’s work is the patch datasets and the U-Net models. The archive is Landsat-8 Collection 2 surface reflectance over West Greenland, 30 m, median composites, cut into 256x256 patches on a 64 px stride. Four other researchers in the group train against those patch sets.

Most of my time on it went into auditing. An export named files without the season in them, so summer and winter 2020 wrote identical filenames and collided on Drive: the winter TIFFs were gone and two of four tiles had never arrived, which is why one season had 3,322 patches against summer’s 10,827. Two 2015 sets turned out to be August imagery carrying winter weather, and 4,009 of 4,039 patches in the old winter folder were byte-identical to summer. Re-exporting both 2015 seasons and all four winter tiles left 40,836 patches and 132 GB.

The U-Nets are scored against NSIDC-0714, an independently published ice mask that already sits on the same grid, so validation is a lookup rather than a judgement call. That matters, because a model trained on the spectral labels reproduces them faithfully, errors and all: it scores IoU 0.2499 against the mask where the labels themselves score 0.2492, while the U-Net reaches 0.766, about 89% of what two expert products manage between themselves. Winter is the open problem: its surface reflectance saturates over bright snow, about 16% of winter pixels carry a value above 1.0, and the mask labels ice sheet where a snow index labels any snow at all, so no winter number can be checked against it. Summer is the row to read.