Auto-embedding for AIS varies greatly between Ragwitz and maximum bias-corrected AIS

[Lukas Paulun]

Hi there,

I am currently trying to determine the optimal embedding parameters for computation of AIS in our locust data (see my previous post for description of the data).

First thing I did was to run the auto-embedding both with RAGWITZ and MAX_CORR_AIS on different variables (acceleration, velocity, heading, ...) for each tracked locust individually and look at the resulting distributions of k_HISTORY and TAU across individuals.

First, I get quite different results depending on which auto-embedding method I use. For example, in the case of acceleration I get k=1, tau=1 for almost all individuals when I use maximum bias-corrected AIS and k=4, tau=3 when I use the Ragwitz criterion. Obviously, this would make a huge difference in the interpretation of the embedding.

I can't really find out what the pros and cons of the two criteria are. In some papers, Ragwitz is used, in others it's the maximum bias-corrected AIS and there are no comments on the reasons for a particular choice.

Secondly, for some variables (e.g. velocity and heading) the optimal parameters vary greatly across individuals. For some locusts I get k=1, tau=1, for others it is k=50, tau=5. It seems like in such a case it is not justified to assume homogeneity of individuals and pool everything together as done in Crosato et al., 2018 and Hansen et al., 2021.

But even if I wanted to perform that type of pooling, I don't really understand how this is achieved in the mentioned papers. Obviously, the time series cannot just be appended to each other, because they are separate observations of overlapping time windows. Do I somehow need to use "startAddObservations(), addObservation(), finaliseAddObservations()" instead of "setObservations()", which I am using now?