Charles, I believe model validation is about verifying predicted data with experimental/observed data. Therefore, I'm in doubt that this will be possible. 

@ Charles,

Calibration means: "adjust model parameters to match observations". Without measured data, it is not logical to calibrate a model.

I mean it is not possible.

May you can find or ALLOCATE the model parameters based on reports/literature around your study area according the feasible range of model structure/parameters.

Using empirical methods can be another way to find the parameters of the used model.

Another possible option may be use of sediment data from adjacent watersheds (the dissimilarity of the watershed conditions from different points of views is the main issue: topography, climate, geology, land-use, etc.).

The provided discussion is about validation of a model without calibration (Link)

https://www.researchgate.net/post/In_modeling_applications_is_only_calibration_of_the_model_sufficient_In_what_instances_is_only_calibration_of_a_model_applicable

Raoof is correct - calibration and validation requires data. Without data, you are entering the real of the ungauged basin. This requires some means of inferring the parameter values. The common approach is to use parameters from nearby gauged sites (i.e. sites with data) using a similarity or regionalisation based approach. The similarity approach uses measures of similarity in catchment characteristics to decide which catchment is the donor (i.e. the catchment where you are going to get the parameter values from).See Thorsten et al 2007 (10.1111/j.1749-8198.2007.00039.x).

The regionalisation approach can come in several forms:

1. using relationships between catchment attributes and parameter values to estimate the parameter values for the site in interest based on it's attributes.

e.g. Post et al (1999) doi:10.1016/S0304-3800(99)00125-8

2. regional calibration, where the model is calibrated simultaneously at a number of gauged sites, using catchment attributes as an input into the model.

e.g. https://www.researchgate.net/publication/222688476_Regional_calibration_of_the_Pitman_model_for_the_Okavango_River

3. Constraining model parameters using a regression between catchment attributes and response signatures.

e..g.

https://www.researchgate.net/publication/2944379_Regionalisation_of_Rainfall-Runoff_Models

https://www.researchgate.net/publication/2943857_A_Proposed_Rule-Discovery_Scheme_for_Regionalisation

Yadav et al (2007) 10.1016/j.advwatres.2007.01.005

The model can be validated in terms of comparing the model predictions to those at the gauged sites - typically through a cross validation approach, where one or more gauged sites are left out of the calibration, and the similarity/regionalisation approach used to estimate the parameter values at the sites left out, and the model output compared with the data. You cannot validate the model at the ungauged site - only infer the confidence in the result based on the cross validation.

If you have no data for any site, then you need to take a literature based approach to infer the parameter values. However in this case, you have no way of validating the model.

This question was the focus of the Prediction in Ungauged Basins iniitiative (2003-2013) and the current Panta Rhei initiative (which focuses more on modelling the impact of change) - see http://distart119.ing.unibo.it/pantarhei/.

Article Regional Calibration of the Pitman Model for the Okavango River

Conference Paper Regionalisation of Rainfall-Runoff Models

Conference Paper A Proposed Rule-Discovery Scheme for Regionalisation

Further to other colleagues discussion, I would like to add that you need at least a proxy model based on measured data.

Thank you all for the insightful comments.

I agree with Raoof and most of the other responders that it is not feasible to calibrate or validate a model without 'observed' data.  It's better to clearly state at the outset that the model is uncalibrated and unvalidated, so users of the results know that the output values have very large uncertainties.

   The devil in these details is in the interpretation of what constitutes 'observed' data.  I firmly believe that this should be rigorously interpreted as field measurements within the watershed.  It is possible to constrain model parameters in ungauged basins by supplying parameter ranges based on indirect methods such as regional relationships (values in neighboring watersheds), empirical functional relationships, or proxy data (e.g., using temperature and solar radiation data to compute ET and soil moisture).  I would argue that this is essentially trying to calibrate the model with another model. Unless the secondary model (e.g., soil moisture = f(ET, radiation)) is calibrated with field observations, you may not be any better off with regard to uncertainty.

  When I was a graduate student there was a joke that if the model doesn't fit the data you can fix everything very easily by changing the data.  This is, of course, fallacious reasoning and completely inappropriate if not unethical (it was a joke after all).  In reality, models can give widely varying output depending on small adjustments to parameters.  The recommended practice - with or without observed data - is to run sensitivity tests to see which parameters are likely to be governing the output.  When the model is highly sensitive to a certain parameter, you should be very wary of using values for that parameter that are not tightly constrained by field observations.

When you say data... is it the data in stream/river or overland? If you have just overland  sediment yield (literature or field based data) is good enough for making sure that the model is simulating properly. If not, there are studies that have transfered parameters from other watershed which was calibrated for seidment yield, however, the watershed should have similar hydrological conditions.

What if some model outputs is not simply possible  to measure? What is the sensors measuring those outputs are exhaustive? So why we are modelling and creating functions if model are useless without calibration? Models are based on the common sense, physics and mathematical relationships between masses. Can we trust them?

This is a nice thread. Thanks everyone.