Researchers from Niklas Boers’ team reduce both the spatial and temporal scales of temperature and precipitation without having to retrain the model for each task.
Climate simulations are often archived in the form of coarse-resolution monthly averages, for reasons of storage and computational cost. However, the assessment of impacts, on agriculture or floods for example, requires detailed daily data, capable of reproducing extremes. Most downscaling methods deal with space and leave aside the temporal dimension.
From monthly at 1° to daily at 0.25°
Yu Huang, Niklas Boers and colleagues use a diffusion model to jointly produce daily temperature and precipitation fields at 0.25° from monthly inputs at 1°. Using a conditional guidance technique, a pre-trained model can be oriented by the monthly data without additional adjustment or retraining. According to the authors, the generated fields exhibit consistent temporal dynamics and physically plausible spatial structures. The probabilistic approach aims at impact studies and the construction of digital twins of the Earth.
Read the full article on the original source
Go to the source →