yes camera can as well be used for monitoring of sloth bear(Melursus ursinus) which is of course a nocturnal animal. there are infra-red cameras in market which are very efficient, that can used for the same at night. the same have been to monitor the rare mountain Bongos in Kenya.

Are these species unlikely to make multiple appearances? I would think the same species reappearing would give a higher count? Does this camera get images with enough detail to differentiate? I have no knowledge of this sp. but curious. 

Yes, you can do it. You can use a Spatial Counts, and you can improve your estimation if you have some aditional information, like home range of the target species or maybe some individual recognision.

See the attached file

Yes you can use Royle Nichols model using single season or even multiple season data. The data should be converted in 0 - 1 matrices as you do in capture matrix. Here you need trapping stations and days in rows and columns. Finally use program PRESENCE to analyse the data. You will get a naive and relative abundance (occupancy) value. Thereafter, using the info such as home range of your species from past literatures and the Effective area sampled from your study, you can get abundance value. But this is a relative abundance insex not exactly the absolute abundance.

Non-identified animal densities can be estimated using a generalised Random Encounter Model. See links to papers below for further information:

http://onlinelibrary.wiley.com/doi/10.1111/j.1365-2664.2008.01473.x/full

http://onlinelibrary.wiley.com/doi/10.1111/2041-210X.12346/full

http://onlinelibrary.wiley.com/doi/10.1002/jwmg.902/full

Hope these help!

David

To add to David Jacoby's answer, you can estimate animal densities with the Random Encounter Model, but you will need information about your species of interest. You will need to have independent estimates of animal movement rate (the v in the REM equation) and a good estimate of the detection distance and the sensor area of the cameras.

We used the REM for wild boar, roe deer, red foxes, racoon dogs, brown hares, red deer...

It does work, although it is restricted due to the restrictions. Especially the distance moved per day is quite important to know exactly and per season. If you have high mevement distances (like in roe deer during rutting season) this will influence your estimated density very strong!

And you have test ALL your trailcams.

But id does work!

You don't necesseraly need to use Royle-Nichols models with the data. You can use single or multi season occupancy models. Occupancy can be used as a surregate for abundance when you can't identify individuals. 

For an introduction: http://www.uvm.edu/rsenr/vtcfwru/spreadsheets/occupancy/Occupancy%20Exercises/Exercise3/MacKenzie%20et%20al.%20single-season.pdf

for multi season models: 

http://www.sierraforestlegacy.org/Resources/Conservation/SierraNevadaWildlife/CaliforniaSpottedOwl/CASPO-MacKenzie_etal2003.pdf

In this second paper Mackenzie et al. explain how is it possible to estimate extinction and colonization when species are detected imperfectly. As Subhadeep said, the data must be converted to a binary matrix. 

Dear Cesar. You're right, but "classical" site-occupancy methods are not as precise as capture-recapture (and spatial capture-recapture), because the lack of spatial information. With spatial counts (SC) you can estimate density, and you can easily improve the quality of your estimation if you add some additional information. I never used REM, but with SC, and information about home range, you can reach coefficients of variation below 0.2.

camera trapping is being used in one of the research programme on Malayan sun bear in dampa tiger reserve in India. If you have infra red device to couple with camera trapping that will give you finer results.

There is a parralel that can be made with accoustical detection of cetaceans and abundance estimation. There also there is no access to individual recognition but some nice statistical tools has been developped. Contact Nick Tregenza the designer of the accoustical detection device (C pod) : http://www.chelonia.co.uk/contact_us.htm

The problem with estimating abundance or density when animals are not individually identifiable is that analytical approaches are forced to work with little information and end up relying very strongly on model assumptions. These can sometimes be very strong and not widely applicable to many situations. This is certainly true of the Rowcliffe et al. approach suggested above. Another potential approach is the Chandler & Royle 2013 approach (http://arxiv.org/pdf/1112.3250.pdf), but whichever you choose to use, please do carefully review the (both explicit and implicit) assumptions and see if they hold in your situation.

Sorry, just realized that Jose Jiminez has already suggested the Chandler & Royle approach. Just want to add (after looking at previous suggestions) that the Royle & Nichols model also makes strong assumptions that are unlikely to be met in most situations.

Hi, i am not very experienced like the others here so i will just share my practical knowledge. For determining density of non identifiable species through camera trapping there is a problem of double counting. So you could estimate abundance of a species only by using presence absence data in your sampling plots and analyzing them through Software Presence (using Doyle and Nicole's model of heterogeneity) or by using Software R.

Yes if you can make the judgement that all the unidentifiable species are all of the same species you can use Daryl & Mckenzie occupancy models to estimate the presence and the abundance of the specie using camera traps. However there are some statistical assumptions to take care of into your sampling design. There is also a software PRESENCE whic takes care of all the calculations and you have to feed in the camera trapping history only with the number of individuals caught on each picture. Google "PRESENCE software" and you will find it it is available for free and you can download and use it. 

Just what I needed too. Many thanks to all of you

Useful for me too, thank you!

I would strongly recommend against using the Royle & Nichols (2003) model with camera trap data. As is is, it is problematic to use the basic (i.e. MacKenzie et al. 2002) occupancy model for camera trap data (the size of 'sites' is poorly defined; closure is impossible to assume, but non-closure may be random allowing estimation of Pr(use)). In the case of the RN model, the model assumes that each site has some number N-sub-i, which determine site-specific detection probabiltiy (p=(1-(1-r)**N); that these sites are visited multiple times providing information on detectability and heterogeneity in detectability; that 'abundance' can be described by a model such as the Poisson. In the case of CT data, what is a site...a camera trap? What is abundance IN that site? Does this remain constant? 

Thanks!

Very interesting. Thanks for the links & papers!

Hi all, I have CT data and I want to run some models on it, and this discussion is really helpful and interesting. I would like to ask you if you think it is a good approach to use Royle and Nichols models when I am interested in frequency of site use, in addition to occupancy. I am interested in assessing covariables that affect detection rate (this would be count data with a Poisson distribution) rather than abundance while accounting for detection probability. Do you think it is incorrect to do such analysis and interpretation of the model results? I understand it is difficult to define the site, but I am not interested in the abundance, rather in what factors affect activity assuming a camera with more detections means more animal activity in that area. Sorry to jump into this discussion, but hopefully it can add to it.

Hi Ramiro. Of course you can use N-mixture or RN models, but you need to take in account the key assumptions:

1- Closure assumption

2- No false positive errors

3- Independence of detection (are your CT correlated considering the target species?)

4-Homogeneity of detection among individuals

In RN models you also assume that the abundance is the main source of heterogeneity in detection

I recommend you this paper:

Dénes, F. V., Silveira, L. F., & Beissinger, S. R. (2015). Estimating abundance of unmarked animal populations: accounting for imperfect detection and other sources of zero inflation. Methods in Ecology and Evolution, 6(5), 543–556. http://doi.org/10.1111/2041-210X.12333

Best regards,

Jose

Hi Ramiro, I strongly endorse what Jose said, and would like to elaborate further on a couple of issues that arise when using N-mixture or RN models with CT data:

- Closure: even if you say you are not interested in assessing abundance, both models require that N[i], the true number of individuals in site i remains the same across replicates (which I assume would be one or more calendar days). With CT data, what is the number of individuals in any site? Strictly speaking the site is a very small rhombus defined by the two intersecting arcs of the infrared sensors of the two cameras if you're using paired CTs or the arc defined by the sensor of a single camera up to some detection distance. More liberally, you could define a site as a catchment area around each CT, depending on the movement of the species of interest. In either case, I feel it would be very difficult/impossible to assume that the number of individuals in each 'site' remains the same across replicates, however these replicates are defined. If N changes between replicates, then using, for example, the N-mixture model would underestimate detectability and overestimate abundance, as the model assumes that the variation of n[i] across replicates is just binomial(N[i],p).  

- An implicit assumption is that the Poisson model does indeed closely describe the dispersion of individuals across sites. I think this may be a bigger deal for the RN model rather than the N-mixture model, as information on site-specific abundance only comes from the between-site heterogeneity in the detection--non-detection data, and this abundance is assumed to follow a Poisson distribution (thereby mean(N)==var(N)). Using, for example, the negative binomial distribution instead of Poisson will leave the overdispersion parameter weakly identifiable (as stated in the Royle & Nichols 2003 paper) as there is no information to separately assess variation in abundance from mean abundance. 

For these two reasons, I would be strongly inclined to not use either of these models with CT data. Use of, say, the MacKenzie et al. (2002) occupancy model may be possible if non-closure of occupancy state is random between replicates, but of course, this a) would mean reducing the information in your data from counts to detections, b) would answer a somewhat different question, and c) would need redefining both psi and p (see MacKenzie et al. 2006 book for a brief discussion, also Kendall & White 2009).

Cheers!

This discussion on the appropriateness of the Royle-Nichols model for camera trap data highlights the trade off that anyone analyzing data has to make.....i.e., which imperfect analysis do I apply to my imperfect data? I don't think discarding the R-N model (or any model) because an assumption will be violated is necessary. The question is, what is the consequences of the assumption violation. For example, if you expect the detection probability of animal x (e.g. sloth bears) to be highly variable across environmental gradients in your study, particularly if these are the same environmental gradients that influence patterns in abundance, the consequences of ignoring detection will be high and the R-N model may be a better option. Yes the intercepts of the abundance and detection sub-models will be biased due to extra-binomial variation caused by vague site boundaries and issues of non-independence, but the covariate effect estimates may demonstrate very little bias. This becomes even more true for animals that have fairly low abundances and do not demonstrate strong aggregating behavior. At least this is my experience through analyzing simulated data and applying this analysis to fish. 

Its important to have an idea which assumptions you will violate, but don't stop there. Its also important to consider how those violations will impact your inference. Sometimes they do, but sometimes they don't.

Cheers!

Dan

Early results from controlled tests of camera traps with a real animal as a target show that they have strong biases in their detection;

https://www.researchgate.net/project/A-realistic-reproducible-and-rigorous-test-for-wildlife-camera-trap-performance

I am not sure whether these biases violate any of the assumptions of the various population estimate models.

Absolutely agree. In those cases, to estimate density you need to model the baseline probability of detection with a covariate (e.g. using the model of the camera). Those differences that you describe can cause a serious error in the estimate if they are not considered in the model.

Best regards

This might be of your interest too: 

https://www.researchgate.net/publication/313232713_Precision_and_reliability_of_indirect_population_assessments_for_the_Caspian_red_deer_Cervus_elaphus_maral

Article Precision and reliability of indirect population assessments...

Another recent approach is Distance sampling with camera traps: 

http://onlinelibrary.wiley.com/doi/10.1111/2041-210X.12790/full

Dear colleagues:

I'd like to share my paper, with some topics related about this question. We use spatial counts (and telemetry) to estimate density from Egyptian mongoose. The main problem we found is the lack of identificabilty of some parameters using only unmarked individuals (count data). Telemetry data allow us to improve the estimates. However we can not use some interesting models without individual ID (e.g, behaviour, sex, etc) .

Best,

Jose

https://www.researchgate.net/publication/312882505_Estimating_carnivore_community_structures

Article Estimating carnivore community structures

I recommend Peter's answer

Best Regards