joewheaton
commented
6 years ago Dear Fu-Chun Wu, Thanks for your inquiry and we are glad you found the GCD 6 and DoD3. We're no longer supporting earlier versions of GCD, as we've just released GCD 7 and have worked hard to move everything over to that new platform. However, there are a number of plausible answers to your questions:
I perform probabilistic (Bayesian) thresholding with DoD3 and GCD6, their results are different even when I use the same propagated error (sigma_DoD = 0.25) and CI (threshold probability = 0.95). The first plot attached below is from DoD3, and the second is from GCD6. Does this difference arise from different Bayesian treatments used in DoD3 and GCD6?
- It is difficult to say for sure if the visual differences in the two rasters you provide images of are actually different, or those are simply differences between the color ramps and display properties. The best way to check is to actually difference the two DoDs (or in GCD 7, there is now a hand Intercomparison Tool). It also helps to display both with the exact same symbology by importing the symbology from another layer to the current layer.
- There are a number of ways in which the precise answers between different versions of GCD and DoD software could be wrong. DoD is an old Matlab code, GCD 5 and 6 were a C++ code and 7 is C# and all had different ways of storing rasters in memory, different significant figures, different degrees of Arc dependencies (DoD3 and GCD 7 have none) and these can all result in real minor differences between calculated values in any of the many steps along the way.
- @MattReimer did find some bugs in our Spatial Coherence filter algorithm used in GCD 6 that was not always present (sometimes) when refactoring from GCD6 to GCD 7. I can't remember exactly what these were, but its possible this could be adding to the problem.
- To properly identify the exact place along the way where there is a discrepancy (e.g. is it in rawDoD.tif, properror.tif, apriorProb.tif, nbrEros.tif, nbrDep.tif, condProb.tif, or postProb.tif, or threhDoD.tif) is possible, but a lot of work. From what you've shown, I suspect these are just differences in rounding precision, as opposed to fundamental conceptual differences in the algorithms behind each output.
More fundamentally, and as an aside...
- Be careful with using the spatial coherence filter via Bayesian updating anyway. If you are really confident that your survey does not suffer from any systematic bias, then the spatial coherence filter can be a nice way of recovering information. However, it has a tendency to recover some of the stuff you want and a lot of the stuff you don't want.
- A confidence interval of 95% is a little on the conservative side (good when using a spatial coherence filter, maybe), but you might consider using an 80% confidence interval instead.
- As a general rule of thumb, I'd recommend a 'default' of doing probabilistic at 80% confidence interval as a starting point. Bayesian Updating is good, but I don't like using our spatial coherence filter to derive a conditional probability everywhere for most surveys (I'd like it for a masked area of low magnitude changes the user says they trust)...
The bottom line is whether or not it is a fundamental change so important that its worth worrying about, or just one of these many minor differences described above. I'd suggest its probably not worth worrying about, and I would worry more about whether or not using the spatial coherence filter is the best method here. We've tried to add the ability to use multi-method masking in error modelling as a more practical work around to get through these issues, as well as AOI implementation in thresholding of results.
I hope that helps...
This message was recieved from a DoD3 and GCD 6 user: