The latitudes where the glacier map is weakest

The number itself

IceBoost v2.0 is a decision-tree scheme trained on 7 million ice thickness measurements, and it computes the thickness of every glacier in the Randolph Glacier Inventory: 215,547 outlines in version 6.0 and 274,531 in version 7.0. The global volume that comes out is 150,000 cubic kilometres with an uncertainty of 38,000 cubic kilometres, and the sea-level equivalent is 323 millimetres with an uncertainty of 91 millimetres. The figure has a concrete counterpart: glaciers make up roughly 25 per cent to 30 per cent of today's sea-level rise and have lost about 5 per cent of their total mass over the past two decades.^[1]

The total agrees with two earlier global estimates, which gave 141,000 and 158,000 cubic kilometres. Yet the paper writes that reconstructed thickness distributions can vary substantially from model to model for individual glaciers, ice caps and even large glacier complexes. Three models can therefore meet on the same total while disagreeing about where the ice sits. Water planning for a basin uses the thickness in that basin, and a global sum leaves that question open.^[1]

The geography of confidence

Against measurements, the model's root mean square error is 20 per cent to 45 per cent lower than that of other models in the high Arctic and comparable elsewhere. The authors say directly where the confidence is high: at high latitudes, where training data are abundant. Over steep mountainous terrain, on small glaciers and in lower-latitude regions with limited training data, confidence falls. The Geikie Plateau in East Greenland shows the scale of what that means; there the model finds nearly twice as much ice as previously reported.^[1]

One reading of that pattern is that no physical law is imposed during training. Because the model learns from data, it works well where measurements are plentiful and poorly where they are sparse. Another reading is available: steep, thin, crevassed mountain glaciers are harder targets for any method, so the gap in confidence may come from the terrain itself independently of how much data there is. The paper's Jensen Gap analysis carries a separate warning: the model is strongly concave over low-slope, thick-ice regions, which can mean a tendency to underpredict thickness under input uncertainty.^[1]

Whose problem is the missing observation?

The places where the data gap sits are the Himalaya, the Karakoram and the Patagonian ice fields. These are also the basins where glacier water carries agriculture, drinking supply and electricity. The map is therefore most reliable where ice touches human life least and most uncertain where it touches it most. That imbalance follows from where measurement campaigns have gone so far.^[1]

One of the uses the authors list for the dataset is exactly this: informing the design of field campaigns. That yields something to watch for. If new thickness measurements are collected in the lower-latitude mountain basins, the expectation is that in the next release the global total will stay inside its uncertainty band of 38,000 cubic kilometres while regional volumes move noticeably. For sea level the number that matters is the total; for water it is the thickness of the ice above one valley.^[1]