trouble using covariates and buffer correctly

[cdevens226]

Hello,

I have a large leopard occurrence data set from camera trap surveys done across multiple regions throughout the Western Cape. I am looking to estimate density across the separate survey areas, as well as overall. While I could keep things simple and just use the most basic of scripts, I am excited about (and frankly overwhelmed by) all the possibilities that secr has to offer. Unfortunately I am very new to R and struggling to understand what and how to incorporate various covariates, behavioral response models, Dsurface options, etc, to ultimately come up with a density estimate that makes sense.

I'm trying to establish density in conjunction with varying land cover and land use. I'm trying to use covariates like human footprint, vegetation status, altitude, some of the Worldclim variables, anthropogenic biomes and GlobCover. Perhaps this is too many/ much?

After much frustration (and likely some dumb luck) I have the script running and providing me with density estimates but I don't have much confidence in what I have gotten. I'm currently running each regional camera trap survey separately (made master trap and capture inputs with all the surveys but haven't been able to get it to run).

This is one of the scripts I'm currently using for one survey region with 48 camera stations and 210 captures over 4 months (2 survey sessions):

PP_robCH_allCov <- read.capthist(captfile = 'robertson_CAP.csv',
trapfile = 'robertson_TRAP.csv',
detector = 'count', fmt = 'trapID', covnames = 'sex',
trapcovnames=c('/Hft','Veg4_terr_eco_status','Veg4_cons_value','Veg4_prot_value',
'Veg4_eco_status','Altitude','Bio_4','Bio_12','Bio_15','Bio_1',
'Anthro_biomes','Globcover_2009'))

secr.fit(PP_robCH_allCov, model = g0 ~ 1, trace = FALSE, buffer=7000)


RESULTS:

Fitted (real) parameters evaluated at base levels of covariates
session = RoberstonPhase1
link estimate SE.estimate lcl ucl
D log 8.800222e-05 1.443208e-05 6.394763e-05 1.211052e-04
g0 log 1.969809e-02 2.577642e-03 1.525845e-02 2.542949e-02
sigma log 4.155779e+03 2.106761e+02 3.762952e+03 4.589614e+03

session = RoberstonPhase2
link estimate SE.estimate lcl ucl
D log 8.800222e-05 1.443208e-05 6.394763e-05 1.211052e-04
g0 log 1.969809e-02 2.577642e-03 1.525845e-02 2.542949e-02
sigma log 4.155779e+03 2.106761e+02 3.762952e+03 4.589614e+03


...WITH NO BUFFER:

secr.fit(PP_robCH_allCov, model = g0 ~ 1, trace = FALSE)

Fitted (real) parameters evaluated at base levels of covariates
session = RoberstonPhase1
link estimate SE.estimate lcl ucl
D log 1.331803e-01 2.177714e-02 9.686607e-02 1.831083e-01
g0 log 6.150240e-04 5.707113e-05 5.129489e-04 7.374117e-04
sigma log 4.548505e+09 6.777801e+01 4.548505e+09 4.548505e+09

session = RoberstonPhase2
link estimate SE.estimate lcl ucl
D log 1.331803e-01 2.177714e-02 9.686607e-02 1.831083e-01
g0 log 6.150240e-04 5.707113e-05 5.129489e-04 7.374117e-04
sigma log 4.548505e+09 6.777801e+01 4.548505e+09 4.548505e+09

Problem is with a buffer 8.8 leopard per hectare makes no sense. Using Capture and other more basic density equations, we have a ballpark idea that leopard density should be maybe 1.3 to 1.6 leopard/100km.... and this density would mean there are 880 leopard/100km2!?! I chose the 7000m (7km) buffer because its about the average distance moved per day by both genders. Though, without the buffer I'm still getting 133 leopard/100km2. Am I using the buffer completely wrong? I tried to use the suggest.buffer command:

suggest.buffer(PP_robCH_allCov, detectfn = NULL, detectpar = NULL,
noccasions = 545, ignoreusage = FALSE, RBtarget = 0.001,
interval = NULL, binomN = NULL)

and got: [1] 19061

I'm not sure how to use this? Apologies for throwing so much into this message, but could you please explain if and how to use the buffer and if you think my script would greatly benefit from adding or omitting any covariates, commands, etc. Am I using some of these commands incorrectly? Am I interpreting this wrong?

I'd eventually love to try to use the mask and Dsurface options, but I am first desperate to get a density estimate that makes sense!

A huuuuge thank you in advance!!! (apologies for the novel!)

[murray.efford]

Hello cdevens226*

Your estimated density is shown as 8.800222e-05 leopards per hectare i.e. 88 leopards per 100 km^2. I don't see the problem. (If that number is too low for you, think about whether you have included non-habitat in the buffered area.)

The habitat 'mask' should include all sites (home-range centres) from which a leopard might reach a camera. Setting a buffer distance is a quick way to specify a mask. With sigma = 4km it is possible leopards centred 12-16 km away will still sometimes turn up at a camera - hence the suggested buffer of 19 km. You get better control by constructing the mask separately with 'make.mask' and providing that as an argument to secr.fit instead of a buffer width.

You have included spatial covariates as detector-level covariates ('trap covariates'). For modelling variation in density you need either mask-level covariates (observed at each mask point) or session- covariates (observed for each spatial block treated as a 'session'). Session covariates should be provided in the secr.fit argument 'sessioncov' (a dataframe). Mask covariates are often added to a mask with addCovariates. Either may be used in the formula for density in the 'model' argument.

My sense is that even quite rich capture-recapture data contains limited information on spatial variation in density (i.e. the intensity of a spatial point process for home-range centres) in relation to covariates, so I would not expect too much.

Hope this helps a little. It's good to start out simple!
Murray

* I'd rather talk to a real person!

[cdevens226]

Hi Murray!

Thank you so much for getting back to me so quickly and clarifying the habitat 'mask' and covariate options! Looks like I will have lots to keep me busy in the coming days! I just want to reply quickly to explain a bit better about what is a realistic density for me to get (to give you a better context)!

As far as the density of 8.800222e-05 leopards per hectare that i got, this is far too high! The smaller 20-40kg leopard that I am studying in the Western Cape of South Africa have much larger ranges compared to their 70-90kg northern counterparts. Males (in the region of the survey I shared here) have an estimated average range of 250 km² (25,000 ha) and females 120 km² (12,000 ha). If I just crudely calculate density by taking number of leopard individuals divided by surveyed area (including 5km outer buffer around camera traps) I get 1.52 leopard per 100km2. This is why I'm so concerned and wondering how I got a secr density so off the realistic mark.. whether I didn't chose appropriate script options or because I've incorporated extras like covariates incorrectly.

I'll get to work on your suggestions ASAP and see if this doesn't get things heading in the right direction! Fingers crossed!

Thanks so much!

Carolyn (oops!! )

[murray.efford]

Hi
Let's get this straight: 8.8e-5 leopards per hectare means 8.8e-3 leopards per sq km (100 ha per sq km), or 0.88 leopards per 100 sq km. We both fumbled that first time round! I think it's working.
Murray