Error in my.RSS.table[as.character(i), 3:4] <- c(pot.index[opt], break.RSS[opt]) : replacement has length zero

[GreatEmerald]

This is a very odd edge case. With a specific h value and specifically when using both trend and harmon components of bfastpp, for specific data, I get the error:

Error in my.RSS.table[as.character(i), 3:4] <- c(pot.index[opt], break.RSS[opt]) :
    replacement has length zero

Also a lot of warnings about sqrt() producing NaN values. Looks like that line fails because opt is an integer(0), and that is because break.RSS is a single NaN (which probably comes from the sqrt problem, which in turn comes from fr in recresid being a negative value, which comes from X1 having negative values, and I don't know what X1 stands for).

To reproduce, see attached CSV, and then run the code:

breakpoints(formula(response~trend+harmon), data=
    bfastpp(bfastts(CSV$evi, CSV$time), "irregular"), order=3), h=46)

bug.txt

[angeliquelv]

Have you figured out this issue already? I am running in to exactly the same problem.

[GreatEmerald]

It seems to be a rare issue and probably related to singularities in the regression... If you change the components, it works, so you can use response~trend+season or so. In practice, I just do a tryCatch() to capture these pixels. Of course, it would be nice if we could do that automatically in the package...

[angeliquelv]

Thank you for your quick reply.
Is that also the explanation for the warning that sqrt() prodcues NaN ? Because that doesn't seem to be so rare in my dataset, it happens with a lot of pixels. In contrast to the error, that indeed only happens in a few pixels and for specific parameter settings. Interestingly, it never happens when I set the order to 1.

[GreatEmerald]

It's probably a separate issue, I was talking about the error specifically. Indeed, any change in parameters seems to fix it, maybe it's just that the second order is impossible to calculate in this case...