AI and the World's Hard Problems · Earth & climate science · arXiv · 2026-10-21
TerraNostra-Bench: Benchmarking Spatial Downscaling and Temporal Rollout for Global 1-km Land-Use Reconstruction
Uncertainty in the terrestrial carbon cycle constrains climate projections, in part because land surface processes are imperfectly represented in Earth system models. The authors propose a data-driven framework that treats high-resolution land-use reconstruction as semantic segmentation, combining coarse-resolution scenario data and climate classification with static geophysical features to produce annual land-use maps at 1 km resolution, and introduce TerraNostra-Bench, which separates spatial downscaling from autoregressive temporal rollout and scores models on rare land-use-change pixels. Comparing U-Net and Swin backbones against persistence, linear, and shallow convolutional baselines shows the autoregressive prior is the most informative input yet accumulates error, with autoregressive models losing 5.7–7.2 accuracy points over 13 years versus 2.2–2.8 for non-autoregressive ones; the trade-off stems from the autoregressive choice rather than the backbone, can be tuned without retraining via an inference-time prior-trust threshold and training-time objective weighting, and persists when extrapolating to 1955–1959 and 2015–2019, with calibration recovering up to 10 accuracy points.