Ocean Color Forum - Not logged in
I'm still trying to find an answer to the question I posed in http://oceancolor.gsfc.nasa.gov/forum/oceancolor/topic_show.pl?tid=4479.
Maybe I should ask it in a different way.
The OC algorithm coefficients were updated using NOMAD version 2 on 18 March 2010. I have two questions: First, were the coefficients determined by using the same "tuning" process described in O'Reilly et al. 1998 and in Tech. Memo Vol 11 in which the slope and intercept of the log (modeled C) vs log (observed C) regression were fixed to 1 and 0? If the answer to this is 'yes,' then exactly how was it done (IDL code available)? If the answer is 'no', were they determined by standard linear regression?
The process of deriving coefficients was modeled after that described in O'Reilly et al. 2000 (T/M Vol 11). We developed analogous routines to those described by Jay. An original fit was made to the data, then this fit was refined to force the regression slope and intercept to 1 and 0. SVD was used for the polynomial fitting and a Type II linear regression was used within the slope/intercept adjustment process. Your question is timely ... an ATBD is currently being composed to fully describe this algorithm development process (which will perhaps include software). Hope this helps.
Well, that sort of helps. I hesitate to get into much detail in a public forum because this is work in progress but suffice it to say that I'd like to use the same methods for deriving the coefficients that you do on a different set of data. I've done the initial fitting step (using a robust fitting routine based on poly_fit rather than svd) and have the Type II regression results; what I don't understand (or know how to implement) is the iterative step of the slope/adjustment process. It would really help just to have an idea (even in general terms) about how you approach that (I could not find anything specific in T/M Vol. 11 - if its there, I've been missing it). I can sort of imagine developing a cost function based on the difference between 1 and the regression slope combined with the difference between 0 and the intercept, but are they weighted in some way? combined? Maybe the best I can do is offer to be an external reviewer of the forthcoming ATBD
. Anyway, thanks for the reply. After not hearing any response to my initial query I was worried that the answer to my question was so obvious that I shouldn't have asked.