DSM with camera trap distance sampling data
Benjamin Debetencourt
Hello everyone,
I am working with camera trap distance sampling, and I want to use those distance data from the cameras to make a DSM. As I have not seen yet any studies using CTDS data to do some DSM, my first question is actually
1) Is it ok to use the data from CT, which are highly non independent in a GAM and in the function dsm that is built under R in the package dsm right now?
I still dived into DSM with my CT data without being sure I was not outrageously violating some modelling requirements, I manage I think to control well for the effort in it, and I obtained sensible results. I use as predictors a couple of spatial predictors, like distance to the river, a human population density that is weighted with the distance cost from the village to my points and a couple of other. I have strong concurvity between some spatial predictors and my latitude and longitude that are in my model. I first removed all the predictors that were having the “worst” concurvity check above 0.8 with my latitude and longitude and kept my s(X, Y) term. But then I do not explain why I have those spatial variations in abundance biologically… So my second question would be
2) Can I instead, remove s(X, Y) from my model ? In all the examples I read and worked through, the lat and long are always in the model. If I remove s(X, Y) in the model, do I have to add a specific term to take into account what I think s(X,Y) does which is if I have a lot of individuals at my specific point, I will more likely have more individuals close to this points ?
Finally, I am struggling with something a bit more specific, linked to the area we are sampling. I am working in a highly mosaic habitat region, mostly bushy and woody savanna, with a bit of gallery forest. We are studying chimpanzees, and they mostly use the gallery forest. In the design, we stratified by habitat, placing more cameras in gallery forest, a bit less in woody savanna, and less in bushy savanna. For the DSM I combined all my distance data together to have this one detection function. The issue I am encountering is that my number of capture or distances I have of chimpanzees locally in each camera is mostly driven by which habitat the camera was placed in. Yet, I did not integrate in my model this local habitat where my camera is placed. Because all of my predictors are on larger scale, like percentage of gallery forest in a 2000m buffer around my point (roughly small chimp territory size). I am not capturing this very local variation between my cameras. But if my camera is set locally in a savanna habitat, but has high forest gallery around it, I am going to have points with zero capture with high value of forest gallery and vice versa. I ended up with the forest having no effect on my count of chimp which I know is not true. My predictors as they are now do not explain why I have more capture at one point and less at the other point. So here is my third question:
3) How could I consider the habitat my cameras are set in in the model formulation for dsm? My problem being that I do not see how I could possibly integrate this very specific local habitat into a predictor value for my predictor cells (the habitat map I have is 10m x10 m resolution VS predictor cell I am using so far 1km² (I could make it finer, but I do not think it makes sense to do a prediction on 10x10m grid cell?)
I hope my questions are clear enough, and I would greatly appreciate any input on any of my three questions!
Best regards,
Benjamin Debetencourt
David Lawrence Miller
I can try to answer some parts of your questions and perhaps those who
know more about camera trapping can correct me where I'm wrong.
1. I think it's totally fine to put CT data into a DSM, in theory. The
issue I can see at the moment is that I haven't written a very
sophisticated way to deal with temporal availability of animals. You can
currently make a per-segment correction via the availability= argument,
but this doesn't account for the uncertainty in the data (cf how this is
handled in bootdht in the Distance package,
https://rdrr.io/cran/Distance/man/bootdht.html). It seems to me like
that aspect probably needs more investigation.
2. You don't have to have any particular covariates in the model. The
reason usually to have s(x,y) in a model is that it usually provides
good results within the surveyed area (though can be problematic when
extrapolating). You should certainly try any options that make sense to you.
3. One option for this kind of situation I have seen is using
"proportion of habitat type X" for a given segment or prediction cell
(e.g.,
https://esajournals.onlinelibrary.wiley.com/doi/full/10.1890/11-1400.1).
That would allow you to use larger cells while retaining some
information and also give you a continuous variable to potentially
smooth over.
Hope that helps,
--dave
On 16/09/2021 15:25, Benjamin Debetencourt wrote:
>
> Hello everyone,
>
> I am working with camera trap distance sampling, and I want to use those
> distance data from the cameras to make a DSM. As I have not seen yet any
> studies using CTDS data to do some DSM, my first question is actually
>
> right now?
>
> I still dived into DSM with my CT data without being sure I was not
> outrageously violating some modelling requirements, I manage I think to
> control well for the effort in it, and I obtained sensible results. I
> use as predictors a couple of spatial predictors, like distance to the
> river, a human population density that is weighted with the distance
> cost from the village to my points and a couple of other. I have strong
> concurvity between some spatial predictors and my latitude and longitude
> that are in my model. I first removed all the predictors that were
> having the “worst” concurvity check above 0.8 with my latitude and
> longitude and kept my s(X, Y) term. But then I do not explain why I have
> those spatial variations in abundance biologically… So my second
> question would be
>
> possibly integrate this very specific local habitat into a predictor
> value for my predictor cells (the habitat map I have is 10m x10 m
> resolution VS predictor cell I am using so far 1km² (I could make it
> finer, but I do not think it makes sense to do a prediction on 10x10m
> grid cell?)
>
> I hope my questions are clear enough, and I would greatly appreciate any
> input on any of my three questions!
>
> Best regards,
>
> Benjamin Debetencourt
>
> You received this message because you are subscribed to the Google
> Groups "distance-sampling" group.
> To unsubscribe from this group and stop receiving emails from it, send
> an email to distance-sampl...@googlegroups.com
> <mailto:distance-sampl...@googlegroups.com>.
> To view this discussion on the web visit
> https://groups.google.com/d/msgid/distance-sampling/5d3de899-fe52-488e-a482-5cb546583cddn%40googlegroups.com
> <https://groups.google.com/d/msgid/distance-sampling/5d3de899-fe52-488e-a482-5cb546583cddn%40googlegroups.com?utm_medium=email&utm_source=footer>.
Benjamin Debetencourt
David Lawrence Miller
Re: 1), I think that sounds right to me. If you're calculating an
overall temporal availability (i.e., it does not vary in space) then I
think that you can just add the squared CV from the availability to the
squared CV from the abundance estimate to get a total squared CV (and
root that to get the total CV). I think that's reasonable, as there's
nothing there to be correlated with space (since the estimate is
"flat"), so the independence assumption is okay. If you're doing
something more complicated (estimating temporal availability per
stratum?) you might be in for a headache. As I said before, perhaps
other folks who know more about camera trapping will hopefully wade-in
here if I'm talking nonsense.
Hope that helps,
--dave
> Hi Benjamin, hi listfolk,
>
> I can try to answer some parts of your questions and perhaps those who
> know more about camera trapping can correct me where I'm wrong.
>
> 1. I think it's totally fine to put CT data into a DSM, in theory. The
> issue I can see at the moment is that I haven't written a very
> sophisticated way to deal with temporal availability of animals. You
> can
> currently make a per-segment correction via the availability= argument,
> but this doesn't account for the uncertainty in the data (cf how
> this is
> handled in bootdht in the Distance package,
> https://rdrr.io/cran/Distance/man/bootdht.html
>
> 2. You don't have to have any particular covariates in the model. The
> reason usually to have s(x,y) in a model is that it usually provides
> good results within the surveyed area (though can be problematic when
> extrapolating). You should certainly try any options that make sense
> to you.
>
> 3. One option for this kind of situation I have seen is using
> "proportion of habitat type X" for a given segment or prediction cell
> (e.g.,
> https://esajournals.onlinelibrary.wiley.com/doi/full/10.1890/11-1400.1
> > <mailto:distance-sampl...@googlegroups.com
> <mailto:distance-sampling%2Bunsu...@googlegroups.com>>.
> >
> https://groups.google.com/d/msgid/distance-sampling/5d3de899-fe52-488e-a482-5cb546583cddn%40googlegroups.com
> <https://groups.google.com/d/msgid/distance-sampling/5d3de899-fe52-488e-a482-5cb546583cddn%40googlegroups.com>
>
> >
> <https://groups.google.com/d/msgid/distance-sampling/5d3de899-fe52-488e-a482-5cb546583cddn%40googlegroups.com?utm_medium=email&utm_source=footer>>.
> --
> You received this message because you are subscribed to the Google
> Groups "distance-sampling" group.
> To unsubscribe from this group and stop receiving emails from it, send
> an email to distance-sampl...@googlegroups.com
> <mailto:distance-sampl...@googlegroups.com>.
> To view this discussion on the web visit
> <https://groups.google.com/d/msgid/distance-sampling/CANq5Wkh8P2S2Mreg4ZRgt%2BMPNBTo7FXK3u2ZfnBbSu30o%3DO3tw%40mail.gmail.com?utm_medium=email&utm_source=footer>.
