Validating the validation?
So, I’ve been pondering an interesting scientific dilemma recently: how do you validate a validation technique? That is, if you’re using a certain procedure to validate some data (that is, check how correct/accurate it is), how can you validate the validation procedure itself?
This has come up in my work recently in relation to validating Aerosol Optical Thickness (AOT) data from satellites. Validation is normally performed by comparing satellite-derived data to data from the Aerosol Robotic Network (AERONET). However, this is challenging for a number of reasons: the main one being that AERONET measurements are point measurements of the AOT above the measurement site, whereas the satellite will measure AOT averaged over a large pixel area (10 x 10km for MODIS). A number of validation methods have been proposed to deal with this, and they are all based on spatial and temporal averaging of the data, to bring it into a comparable state.
The big question then becomes:
If there are a number of different methods for validating this data, then how do we choose which one is best?
This is a big problem because the results of the validation will be used to infer the accuracy and uncertainty of the dataset – and if we use a validation that gives misleading results then we will have a misleading opinion of the dataset. This gets even more difficult when you need to choose parameter values for use in these validation procedures. For example, if the validation involves temporal or spatial averaging then what period/distance should the averaging be done over? 30 minutes? 1 hour? 5 hours?
So, how should we deal with this?
Unfortunately, I don’t know. I know that I have a problem with some of the current methods (for example, trying a wide range of parameter values and choosing the ones that give you the best validation results – that is, the lowest error) as some of them really do seem to be ‘cheating the system’. This is something that I’m hoping to come back to over time – I already have some preliminary ideas on how to ‘fix’ the AOT validation system – but I think it is a problem which won’t go away.