Hi Farzad,

I found that using precipitation trends are useful to make a downscaling. Here I send you some papers that might help you. 

Hi Farzad,

I found that using precipitation trends are useful to make a downscaling. Here I send you some papers that might help you. 

If you are looking to project changes in to future (e.g., by 2100), then I highly recommend using ARRM method (https://www.researchgate.net/publication/263223774_An_asynchronous_regional_regression_model_for_statistical_downscaling_of_daily_climate_variables) since it can capture extreme events (which are important for hydrological impacts assessment) and also it captures the changes in probability distribution of climatic input (e.g., precipitation) as well as day to day variations. It is statistical method and can be trained based on point measurement (e.g., weather station). Although keep in mind if you are trying to model a large basin and aggregating daily output to use them as monthly input to your model, then most statistical models should work for you (e.g., BCSD). Just stay away from Delta method since it is very unreliable! 

Article An Asynchronous Regional Regression Model for Statistical Do...

Dear Farzad,

have a look at the data access and downsclaing portal from the university of Santander. 

http://www.meteo.unican.es/en/portal/downscaling

But please be aware that the choice of downscaling procedure heavily depends on the scientific question to be answered; e.g. do you focus on extremes or medians etc....

This article is a very handy guideline on how to select proper downscaling technique:

http://onlinelibrary.wiley.com/enhanced/doi/10.1002/2014EF000259/

http://onlinelibrary.wiley.com/enhanced/doi/10.1002/2014EF000259/

Dear Farzad,

I had the same problem in my PhD. Please find my solution in the attached link .

The choice of the method depends on your research question, the available data, and scale of the analysis. If you do a large scale analysis and you don't want to spend much time with the method, you may also try the readily available CMIP3 data, which is available from the SWAT website. However, I did not use this myself so far...

http://link.springer.com/article/10.1007/s10113-013-0481-z/fulltext.html

I think LARS WG is a good option if you know the predictors and predictands already. We have used that method for some years now with good results. You should be aware of the correlation matrix between these two data to decide if this would be a good method for your purposes.

A very simple method for empirical downscaling is the analog-method. Although I used it for upscaling (Schenk and Zorita, 2012), it was originally used mainly for downscaling precipitation. Technically, the method works the same by just replacing predictant and predictor. You find relevant references for downscaling also in my paper, i.e. look for Matulla et al., 2005 for scenarios and Zorita and von Storch (1999) for a general comparison of analog-method vs. other methods for downscaing.

Article Reconstruction of high resolution atmospheric fields for Nor...

The reliability of the GCM precip. outputs may be a bigger problem, but beyond your control.  What you are worrying about is like putting a modern-day telescope on a flintlock rifle to improve your accuracy.  Good luck.

@John: Yes, GCM precipitation might be not good. If you are lucky, at least the change in statistics of GCM precipitation is useful. In this case, you might link poor precipitation GCM output to reasonable observational or reanalysis driven regional model fields or time series using the above mentioned analog. The advantage of the analog-method is that it is non-linear and does not assume any type of probability distribution for the predictor or predictand.

One of the downscaling methods that have obtained better results for precipitation is the two-step analogue/regression technique, developed by Ribalaygua et al. (2012) and validated for Europe (Linden and Mitchell 2009, Goodess et al. 2011). Another option is to use a quantile mapping as the parametric approach of Benestad (2010) or the proposed in Monjo et al. (2014).

Article Description and validation of a two-step analogue/regression...

Article Probabilistic correction of RCM precipitation in the Basque ...

Dear Farzad

Definitely now a days your question is the big problem that many PHD student have challenge with it. But I think it strongly depends on your case study.If it is big you don't need to downscale GCMs outputs. But if you want to use finer scale (such as: meteorology station) you should downscale your data. As you told there are different ways & methods for this aim. But uncertainty of these methods are high. I think Lars-wg is not sufficient for you. It is a stochastic method so the uncertainty of it is so high. I think SDSM and ASD are not sufficient  for you too.

I think as others said, the best way for you is using regression method that there are many R package of them in web.

Best wishes,

Nasrin

@Nasrin: The question is which kind of regression should that be? Precipitation has a clearly non-normal distribution. Most likely, the distribution of the coarse GCM (predictor) will be very different from the distribution of your location (predictant). I would recommend to try it with the analog-method. There are several papers about that method for precipitation.

Dear Farzad Emami,

you might be interested in the results of the VALUE network. We have intercompared some 50 statistical downscaling and bias correction methods. The first results are published in Maraun et al., Int. J. Climatol., 2017. The key point is really to look for a method that is structurally suitable, and to find suitable predictors (so if time is a problem you may want to use a change-factor weather generator or bias correct an RCM ensemble). If you want to use "classical" statistical downscaling, you should not use any method based on linear regression. Maybe the analog method might work (although it might screw up your temperature signal), otherwise I would recommend generalised linear models with a stochastic noise component (see the publication above).

An in depth discussion of assumptions, lots of pitfalls, performance and so on can be found in a forthcoming book: Maraun and Widmann, Statistical Downscaling and Bias Correction for Climate Research, Cambridge University Press, 2017.

Best,

Douglas

Hi Douglas,

long time no see ;-)

Yes, good point that the analog-method might not capture the temperature signal. It might be tested however to use multivariate predictors, so precipitation and temperature to predict precipitation (maybe with some kind of weighting). Many things are still to be tested there.

Cheers,

Frederik

Hi

You can find some descriptions about it by searching "statistical downscaling methods in AgriMetSoft"

Good Luck

Check SDSM. SDSM produces point information equivalent to station data from a GCM that can be used in impact assessment model. The downscaling process is based on the predictor predcitand relationship.

Dear all

I am not recommending you and any other researcher for using statistical downscaling methods regarding the projection of precipitation. Since we should regard all effective factors before occurring rainfalls, thus dynamical methods are strongly suggested in the case of precipitation. Therefore, it is better to use dynamical downscaling for projection of precipitation. These methods regard at least the factor of elevation changes as an input in any study area. So, the latter ones tend to express more reliable results in the outputs compared to the statistical methods. Regards

@mohsen, it's right that RCMs add topographic information, but still face large bias in representing the local climate. Hence, Rcms can only be used for regional assessment rather than local scale. To make sure which model works for your area you can check for the reference period.

Well sorry I am coming late.

The point is to down scale. Not to show sophistication. I would rather suggest you stick with statistical downscaling in the environment of SDSM.

Reason being time..... Like you said is of essence here

Regards

There is no reliable way to downscale precipitation over any but the flattest of topography. The GCMs themselves don't do a very good job on precipitation, with even the direction of change unreliable in much of the tropics. Elevation predicts temperature pretty well, but for rainfall you need a whole lot more information on winds etc. which is usually simply not available in most places.