TerraNostra-Bench: Reconstructing the land use of the planet at the nearest kilometre, year after year

Researchers at the Barcelona Supercomputing Center offer a test bench that judges models on the rare pixels where land use changes, where the uncertainty of the carbon cycle is involved.

The imperfect representation of land surfaces in Earth system models remains a major source of uncertainty on the carbon cycle, and therefore on climate projections. A team led by Amirpasha Mozaffari proposes to reconstruct land use at 1 km resolution, year by year, by treating the problem as an image segmentation: low-resolution scenarios, climate classification and fixed geophysical characteristics serve as inputs to produce annual maps.

Judging models where things are moving

The TerraNostra-Bench protocol separates the spatial downscaling from the projection over time and evaluates the patterns on the pixels that actually change, rather than on the stable background that dominates the maps. The authors compare U-Net and Swin architectures, with or without accounting for the previous year. Using the map from the previous year is the most useful information, but the errors accumulate: these models lose 5.7 to 7.2 points of accuracy over 13 years, compared to 2.2 to 2.8 for the others.

This trade-off depends on this design choice, not the architecture, and can be adjusted without retraining thanks to a confidence threshold applied to the inference, which allows up to 10 points to be regained. Tested outside their training window, from 1955-1959 to 2015-2019, the models retain most of their performance. This work prepares a future continuous land use product.