Handout (2.2 MB)
First, the coarse resolution forecasts are interpolated onto the fine resolution grid using a nearest neighbour approach. Then post-processing is done gridpoint by gridpoint by training a calibration based on linear regression using historical forecasts from both models and treating the fine resolution forecasts as the truth. This corrects systematic temperature biases in cases where the land area fraction of the coarse resolution gridpoint is not representative of the higher resolution gridpoint.
Standard application of this procedure yields good result when evaluated using independent measurements from observing stations and has been operational at MET Norway for several years. However, in cases where the ocean temperature is very different from the land temperature, the procedure can give unstable calibrations especially when for example an ocean gridpoint from the coarse model is attempted to be calibrated towards a pure land point in the fine resolution grid. In this case the success of the calibration relies on the ocean temperatures being very similar in the training an evaluation period. We show that a better approach is to not only use the nearest neighbour in the coarse domain but also the nearest pure sea and nearest pure land point in the regression. The calibration will then find the optimal weights of sea and land points.
The method is evaluated in using observations at stations throughout Norway.
![](images/flag.png)
![](images/blue_ribbon.png)