Analog_practice

PRACTICE: Downscaling 2-month ahead System4 hindcast predictions for maximum summer temperature in a set of GSN station in the Iberian Peninsula

Learning goals:

The goal of this practice is to become familiar with the classical data handling steps involved in typical downscaling exercises, in particular:

• Loading data of various types from local and remote locations using the downscaleR and ecomsUDG.Raccess tools to this aim
• Application of a downscaling method (analogs)
• Manipulation of climate data classes in downscaleR: creation of multifields, pre-processing of predictors, EOF analysis, re-gridding...
• Visualization of the downscaled forecast skill and basic assessment of the quality of the predictions

Aim of the practice

During the practice we will perform analog downscaling of the System4 seasonal forecast of summer (JJA) maximum surface temperature on six weather stations of the GSN network in Spain. NCEP reanalysis data will be used for training the model and then we will perform the test on the S4 data. To this aim, we will access the ECOMS User Data Gateway for loading both predictors and test data, while the predictand (weather station data) will be obtained from the built-in dataset included in the downscaleR package. To undertake the practice, you will need to obtain some data from the paper used in the previous demo.

Data needed and methods:

• Predictands: Max daily surface temperature. GSN station data of Spanish stations (buit-in data in downscaleR library).
• Predictors: Atmospheric pattern P5 in the reference paper (see Tables 1 and 2). Load data from ECOMS-UDG. Use the 10 first Principal Components
• Spatial domain: Domain Z6 in the reference article (approximated)
• Test data: All members from System4 seasonal forecast (15 members). Forecast lead month = 2 (i.e., April run)
• Target season: Boreal summer (JJA)
• Training period: 1991-2000
• Test period: 2001-2010
• Downscaling method: M1c (Table 3)
• Validation: multimember skill assessment by means of a plot of terciles (type help("tercileValidation"))
• Extra: If the practice is successfully accomplished and you have still some extra time, repeat the exercise using as predictand the gridded dataset WFDEI instead of the GSN stations, for the same spatial domain.

Practice steps:

1- Load all the GSN stations within the domain of analysis for the calibration period and the target season using the appropriate arguments of the loadStationData function. Explore the structure of the object returned and retrieve some key information like the station names and their codes, their altitude... Also, plot the annual time series for both stations (Note: you don't need to replicate exactly the example plot below).

2- Load the predictors from the ECOMS User Data Gateway, considering again the domain of analysis, the target season and the calibration period. Explore the returned objects and make a visualization of each field.

3- Join the predictors loaded to create a multifield. Note that their grids have different resolution, so prior to multifield creation you will need to perform a re-gridding...

4- Perform the PCA on the calibration data (remember: retain just the first 10 PCs).

6- Perform the downscaling (remember: method M1c).

7- Load the validation data.

8- Visualize the downscaled forecast skill using the tercileValidation function.

---

Extended practice The next steps are optional once the previous steps have been accomplished.

9- Load the gridded observational dataset WFDEI and visualize it.

10- Do the downscaling using the same predictors as before, and applying the method M1a of the reference paper (Table 3).

11- Visualize the skill of the downscaled forecast.

13- Calculate the mean bias of the downscaled output with regard to the validation data. Do it for each member sepparately and then plot a map of the mean multimember bias.