Just to clarify my ...

Just to clarify my assumptions here: where you say "no in situ measurements for hydro-meteorological data" I assume that means "no precipitation measurements" because then you say "I have 112 stations" which I assume are stream discharge data. To narrow down the pool of evaluations, you could: 1. select 52 stations at random from the population (somewhat akin to a jackknife procedure, 52 being approximately the number of CFS Reanalysis grid points in your watershed based on stated grid resolution) 2. do a cross-correlation analysis of the discharge time series for all 112 stations (a simple step, even in MS Excel), rank the correlations (keeping the information on which stations produced each correlation value) from least to most, and then select those station pairs with the lowest correlations until you have a manageable collection of stations (however many you have time to analyze) for your evaluation step. The idea here is to maximize your accounting for the overall variability of stream discharge observations across the watershed. Two stations that are in series on the same stream will likely be highly correlated, so you don't necessarily need both of those stations, as you'd be repeating information in your analysis. Two stations on different streams will be less correlated, so having both will be useful to gauge how well you are representing the spatial variability of precipitation-runoff processes in the watershed. Two stations on opposite sides of the watershed will (likely) be quite different and have a low correlation. Throwing all of the stations into that mix will mean that you get station relationships like that, but also some internal to the watershed, in different sub-watersheds, some at headwaters and some at outlets, etc. with the likely outcome that you end up with a pool of station locations that are spread out all over the watershed, representing both the modeled area and its internal variability.