Climate AI is learning to work where the data runs out

Shared training across ecosystems

Earth system models usually come in specialist flavours, a separate model per variable: rivers, wildfire, carbon. A preprint introducing a unified transformer-based framework for time series prediction of river discharge, wildfires, and carbon fluxes argues that sharing parameters across ecosystems beats keeping them apart. Unified transformer models outperform domain-specific baselines for land-surface predictions across diverse ecosystems, and multi-domain training significantly improves predictive skill in data-scarce regions, specifically the Amazon basin, where observations are thin and a specialist model has little to learn from. m:…0608

The most surprising detail is which training partner helps most. Arctic process-model simulations are the most effective training partner for improving generalization in other ecosystems, which suggests that simulated data from a very different environment can teach a model patterns that transfer. The implication for anyone building earth models is that the answer to sparse data may not be more sensors in the same place, but training that borrows strength from elsewhere, including from simulations. m:…0608

Lightning risk, mapped with uncertainty

A second preprint applies deep learning to a concrete hazard: lightning in Bangladesh. A 12-member deep-learning ensemble reduces mean squared error by 20% compared to monthly climatology for lightning stroke density mapping, and convective available potential energy is the primary predictor for lightning hazard in Bangladesh across all 12 ensemble runs. The ensemble matters because it carries uncertainty with the prediction rather than a best guess with no uncertainty attached, which is what a disaster manager needs when deciding where to warn people. m:…4694

The map itself points to a clear priority. The north-eastern haor basin contains 50 upazilas with a probability of 0.9 or higher of being high or very high lightning risk, impacting 15.5 million people. District-level lightning deaths in Bangladesh correlate more strongly with pre-monsoon risk, at 0.40, than with annual risk, at 0.19, so the danger is seasonal as well as geographic. A risk map that averages the year smooths away the season when people actually die, and this preprint argues for planning around the pre-monsoon peak instead. m:…4694

Filling the gaps clouds leave

The third preprint tackles a problem that quietly limits every satellite-based climate product: clouds. KD-Assim, an incremental Kalman-Delta assimilation method, rebuilds optical satellite time series through cloud contamination by bringing in radar data that sees through cloud. It achieves 28.52 dB PSNR, 0.880 SSIM, and 9.86 degrees SAM on the SEN12MS-CR-TS dataset, which are the reported reconstruction quality scores on the benchmark the authors used, and it uses 52% of the FLOPs required by the strongest single-model baseline. m:…5558

Efficiency is the point as much as accuracy. The model predicted spread is better calibrated than a five-member deep ensemble at one-fifth of the inference cost, so it gives a more honest sense of its own uncertainty while costing far less to run. That combination matters for drought assessment and flood mapping in places where compute budgets are tight and cloud cover is frequent, because a method that only works on a supercomputer does not help the teams that need it most. m:…5558

Read together, these preprints describe a shift in how climate AI is judged. The question is no longer only whether a model beats a baseline on a well-observed benchmark, but whether it holds up where observations are scarce, whether it says how uncertain it is, and whether it runs cheaply enough to be used. A unified model that borrows strength across ecosystems, a lightning map that names its riskiest season and districts, and an assimilation method that works through cloud while using 52% of the FLOPs of the strongest single-model baseline all point in that direction. These are preprints, so the results await peer review and independent testing on other regions and datasets. m:…0608 m:…4694 m:…5558