Dear Farzad,

Do you absolutely need to use observed precipitations? I not, you may want to consider using any satellite rainfall estimate to fill your gaps, thus avoiding errors due to interpolation of other's locations observations.

The possibility of doing this will depend on your time period and on the spatial resolution of your model.

Cheers

Dear Adrien,

Thank you for your comment.

I am working on river  watershed scale which is totally 22000 km2.

 I have few simulated rainfall data such as CRU precipitation dataset. and I am using the observed data together with this simulated dataset.

I think the observed data are more reliable based on th ephysical characteristics of the watershed and the simulated rainfall (satelliote based data) aren't enough for assessing a watershed rainfall variations.

As you mentioned some satellite rainfall databases are available but the question is how can I validate them to be sure that they are fitted to the real conditions of the watershed.

So I prefer to use both of them together to improve the quality of my input rainfall data and I am looking for an efficient method to estimate the missing values for the observed ones.

Best Wishes,

Farzad

Dear Jean-Luc,

Thank you for your comment.

I have seen your article before and I found it interesting.

But the problem is my objective in this step of my PhD project is calibration of runoff and if I make the precipitation based on the  runoff, it seems like a deception of model in my case.

Best,

Farzad

try to use several artificial intelligence based methods like SVM, M5 Regression tree etc

have you tried the multiple linear regression method?

You could have a look at TRMM data. If you are not expecting too much precipitation as snowfall it might provide you with good estimates. You can bin your in situ data to, e.g. monthly amounts, to test against the satellite derived estimates for validation. For the relatively large catchment you are investigating, the spatial resolution might be sufficient in order to capture regional differences. Interpolating precipitation in time is difficult because storm events are obviously discrete. If your investigation period is the last year, you might want to test GPM data, the successor of TRMM, with much better spatial resolution.

good luck,

eric

Also you can create an artificial statistic weather station in order to use it in SWAT model by modifying SWAT database

Dear Farzad,

My homogenization package 'Climatol' at http://www.climatol.eu/index.html can be a good alternative to filling your missing data. If you do not want to spend time in learning how to use it, you can send me   your series (and coordinates X, Y, Z of the stations) and I will return them without missing data. Then, if you find those results useful, I can give you the details on how to proceed.

Best regards,

Jose

Dear Farzad,

Good morning.  I think you can use artificial neural network self-organizing maps (ANN-SOM) to infill the missing gaps in your time series. It is really efficient but you need matlab and also to download the SOM toolbox from the web. I used it too earlier in my research and I had acceptable results. In my case I had missing gaps for up to two years.

I have attached 2 journal papers that can give you a general idea of what it is all about.

hope you find it useful

Dear Farzad,

the data gaps at some of your stations are very large. First, you may think about using only those stations that do have a certain amount of data (the threshold is up to you).

For stations that do not have large gaps, you may identify those stations that best match the missing stations using a linear regression approach. I would recommend to do this based on the available data at both stations. In doing so you derive a linear regression equation that can be used to fill data gaps. However, you should validate this method for your available data. 

If you have larger data gaps, you may go for spatial interpolation methods and you definiftly need a good covariate (e.g. often elevation could be used). I would also recommend to use rainfall patterns from satellite data (e.g. TRMM data) as that worked very well in my research.

I am describing these methods in more detail in one of my papers (see link below). Best,

Paul

https://www.researchgate.net/publication/242344250_Comparison_and_evaluation_of_spatial_interpolation_schemes_for_daily_rainfall_in_data_scarce_regions

Article Comparison and evaluation of spatial interpolation schemes f...

Dear Researchers,

Thank you all for your helpful comments.

Best Regards

If you have  a good reanalysis precipitation data in this region or can make it with a suitable model, try to directly interpolate the gridded precipitation data to the station site as a reasonable estimate of missing precipitation.  For the daily precipitation mostly shows a irregular variation at daily scale, the statistical method could not work well to meet your aim.

Dear Paul,

Thank you for your helpful comment.

I will test these methods in my case.

Best Wishes,

Dear Sasi,

As you mentioned I know that the statistical methods can't generate the real values of a random based phenomena like percipitation and only simulate the missing values using a mathemathical model based on available values. In addition, as paul mentioned the statistical methods should be validated using available values. So these methods are reliable if the model can be calibrated and validated well.

As you mentioned the SWAT model like all models have assumptions and uncertainities. So I will investigate the percipitation validation before using in SWAT model.

Thank you for your comment.

 Best Wishes

Hey Farzad, Was your time series in daily time scale? Did you use the Climatol package suggested by Jose? Can you please follow up how you solved the issue since I also have the similar situation of missing precipitation time series data.

Thanks