Downscaling Air Temperature Forecasts along Intricate Coastlines

Output from numerical weather prediction (NWP) models are often too coarse to resolve the fine-scale air temperature variability along intricate coastlines. For example, the operational NWP model used by MET Norway has a grid spacing of 2.5 km, which can result in forecast errors as high as 10°C for locations where the land/sea mask in the model is not representative of the point location and when land and ocean temperatures are very different.

We present an approach that alleviates this issue by attempting to resolve the subgrid-scale variability statistically. The NWP model operates on a land area fraction (LAF) that is representative for each 2.5 x 2.5 km grid cell, however our approach uses a high-resolution LAF at 1 km resolution. The approach uses the NWP model’s local land-to-ocean temperature gradient and corrects the nearest neighbour’s temperature based on the deviation of the model’s LAF and the high-resolution LAF. The gradient is computed from a regression between temperature and LAF of multiple grid cells within a small neighbourhood surrounding the specific location. This causes land temperatures to extend out on unresolved peninsulas and ocean temperatures to extend into unresolved bays.

We tested the method using temperature observations from 80 weather stations located along the coast of Norway. The method lowers the root mean squared error for most stations and generally creates forecasts that have smaller diurnal and seasonal biases.