Correcting effective survey duration in camera trap survey using still images

182 views

Skip to first unread message

Emeline Auda

unread,

Nov 27, 2023, 1:54:56 AM11/27/23

to distance-sampling

Dear all,

I have recently read with great interest the following article:

Kühl, H. S., Buckland, S. T., Henrich, M., Howe, E., & Heurich, M. (2023). Estimating effective survey duration in camera trap distance sampling surveys. Ecology and Evolution, 13, e10599. https://doi.org/10.1002/ ece3.10599

First, congratulations for this great article that made me question how I calculated my camera trap survey effort in the case of still images.

We have finalized a big camera trap survey in Cambodia in one of our protected areas and I managed to calculate population estimates for 5 species and I just finished estimating distances for 2 other species:

- 2 squirrels species and one treeshrew

- 1 junglefowl species

- 1 pheasant species

- wild pig (population estimate currently being estimated)

- 1 macaque species (population estimate currently being estimated)

I would like to recalculate the survey effort following the methodology described in the article and see how it affects our population estimates. For the two squirrels and the treeshrew, no behavioural reaction of the animal to cameras were observed. In this case, can I use the experimentally derived camera recovery time or do you still recommend to calculate the effective survey duration for each species based on our dataset given that other factors can also potentionally influence the camera recovery time?

For the other species, reaction to cameras was observed. The time interval data should therefore be manually inspected for each of those 4 species. I would first need to clean a bit the dataset for each species and remove any images where I have other species on the images (we did see sometimes wild pig and pheasant or junglefowl on the same images, so these photos should be removed from the dataset). This is where I am not fully sure about the next step. From what I understood, I need to filter for only those pictures belonging to the same animal to get a ‘clean’ time interval distribution that is not contaminated with time intervals between pictures belonging to different animals. I have in my record table the number of individuals recorded on images, especially for wild pig where we can have images with 7 individuals. Do I then need to first remove those images from the dataset (images with number of individuals higher than 1), and then look through the rest of the images to be sure it is the same animal all along? For example for wild pig, sometimes, I have one individual entering the FOV then going out, then maybe 5 seconds later a second individual entering the FOV. And then the two individuals again enter the FOV at the same time. Since I filtered for only images with 1 individual, I won't get that images with 2 individuals. But I do know in my case it is 2 different events because later on I will have those 2 individuals with all their babies for example. How do I deal with that?

I see in the csv file that they are four columns that I do not have in my record table: delta.time.secs, Beh_numeric_event, ID, EventID. I assume the delta.time.secs can be obtained using the package camtrapR and a threshold of 5 minutes for indepedent event can be used (everything less than 5 minutes is considered the same event or same individual and therefore will calculate the delta.time.secs between each photos). However in my data analysis I used a threshold of 60 minutes to filter for independent events, does it mean I need to use the same in this case? Also what does Beh_numeric_event represent? Are the two last columns used in the csv file of importance for the analysis? I assume the interval used in the R script is the delta.time.secs column.

Another question is how to incorporate that "mean time interval between triggers' into my data analysis. If we take for example the siamese fireback in my dataset (pheasant), which reacted to the camera by getting very close to the camera and staring at it for long period of time. I have the effort in days for each cameras deployed, and based on the activity histogram, we estimate the daily time activity from 5:45 am to 18:15 am, so 12 hours and 30 minutes daily. So the daily Tk is 12.5 hours and then we need to multiply it by the numbers of days camera were active and then divide by t, which is the time interval between snapshots, so 2 seconds here. I assume therefore I would only change this 't' and replace by the one obtained following the methodology from the article. But what about all the distance estimations I did at the interval of 2 sec on my images for each species ? In my case, I filtered for all images in the record table that finished by an even number for the time and only estimated the distances for these images. In the article, to do the data analysis, the first photo of each photo series was used a snapshot moment. Do I therefore need to review this step and do distance estimation for the first photo of each photo series? In my case, the settings used are the following: "Cameras were programmed to operate all day and to record one photo at each activation, with the minimum triggering interval between activations (0.6 seconds)." Or if I found for example a t of 6 seconds, do I need to take only images in the data analysis every 6 seconds ?

Sorry for all the questions... Hopefully I was clear enough in my explanation...

Thank you for all your help,

Best regards,

Emeline

Stephen Buckland

unread,

Nov 27, 2023, 12:47:01 PM11/27/23

to Emeline Auda, distance-sampling

Emeline, you have some excellent questions. I developed the modelling method and the R code, but the data were filtered and edited before I analysed them, so I didn’t have to deal with the issues you raise! My comments (below in black) therefore fall short of addressing your questions adequately! Hopefully, one of the users of the method will respond …

Steve Buckland

From: distance...@googlegroups.com <distance...@googlegroups.com> On Behalf Of Emeline Auda
Sent: Monday, November 27, 2023 6:55 AM
To: distance-sampling <distance...@googlegroups.com>
Subject: [distance-sampling] Correcting effective survey duration in camera trap survey using still images

Dear all,

I have recently read with great interest the following article:

First, congratulations for this great article that made me question how I calculated my camera trap survey effort in the case of still images.

- 2 squirrels species and one treeshrew

- 1 junglefowl species

- 1 pheasant species

- wild pig (population estimate currently being estimated)

- 1 macaque species (population estimate currently being estimated)

*I don’t have practical experience, so hopefully someone else will respond. Pragmatically, you could try both approaches, and see if you get similar answers.

*There is some discussion of this issue in the paper, especially with respect to wild boars. Again, I hope someone with practical experience will answer this.

*From memory, my code used only times between events in the filtered dataset, and truncated those values that exceed T. Thus times longer than T are ignored.

*It is not obvious to me that you would need to use the same t for estimating the detection function as for estimating density, so you may be able to use the distances you have already recorded corresponding to t=2, but use the estimated t in the formula for estimating density or abundance. Again, I’m hoping someone with practical experience of the method will respond!

Sorry for all the questions... Hopefully I was clear enough in my explanation...

Thank you for all your help,

Best regards,

Emeline

--
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.
To view this discussion on the web visit https://groups.google.com/d/msgid/distance-sampling/abccd006-00a1-4bbd-8a3b-6a16f52ef30bn%40googlegroups.com.

PABLO PALENCIA MAYORDOMO

unread,

Nov 28, 2023, 5:37:05 AM11/28/23

to distance-sampling

Hi,

Regarding animal reactions to the cameras, this paper could be useful. It has been recently published

Reducing bias in density estimates for unmarked populations that exhibit reactive behaviour towards camera traps - Delisle - Methods in Ecology and Evolution - Wiley Online Library

Hope it helps

Pablo

Eric Howe

unread,

Dec 8, 2023, 2:58:52 PM12/8/23

to distance-sampling

Good day Emeline,

Sorry, there is a lot here and I didn't have time to respond right away.

Your questions concern several different issues, including

1. Potentially long and variable camera recovery times , such that t (the time interval between snapshot moments at which images could be recorded) is unknown and must be estimated. Recovery times result from imperfect and variable performance of cameras.

2. Potentially long and variable "retrigger times" for animals that remain stationary within the field of view. This can cause animals to be missed when in the FOV at snapshot moments ("false negatives"), and hence negative bias in the encounter rate . Long or variable retrigger times are caused by animal behaviour (staying still for a few to several seconds), so are likely to be species-specific.

3. Reactions to cameras.

4. Estimating activity level with respect to your (daily) survey duration

5. Detections of animals in groups

The article you mentioned addresses only 1 and 2 above, though reactions to cameras and activity level are also relevant.

Some specific responses inline below. Steve and others please feel free to correct me or add to my comments.

On Monday, November 27, 2023 at 1:54:56 AM UTC-5 emelin...@gmail.com wrote:

Dear all,

I have recently read with great interest the following article:
Kühl, H. S., Buckland, S. T., Henrich, M., Howe, E., & Heurich, M. (2023). Estimating effective survey duration in camera trap distance sampling surveys. Ecology and Evolution, 13, e10599. https://doi.org/10.1002/ ece3.10599

First, congratulations for this great article that made me question how I calculated my camera trap survey effort in the case of still images.

We have finalized a big camera trap survey in Cambodia in one of our protected areas and I managed to calculate population estimates for 5 species and I just finished estimating distances for 2 other species:
- 2 squirrels species and one treeshrew
- 1 junglefowl species
- 1 pheasant species
- wild pig (population estimate currently being estimated)
- 1 macaque species (population estimate currently being estimated)

I would like to recalculate the survey effort following the methodology described in the article and see how it affects our population estimates. For the two squirrels and the treeshrew, no behavioural reaction of the animal to cameras were observed. In this case, can I use the experimentally derived camera recovery time or do you still recommend to calculate the effective survey duration for each species based on our dataset given that other factors can also potentionally influence the camera recovery time?

EJH: This is consistent with this recommendation from the article: "If it is clear, for example from prior surveys that there are no behavioural reactions of animals to cameras, retrigger delays from natural behaviour are negligible and experimentally derived camera recovery time shows little variation under different conditions, it should be sufficient to just use the experimentally derived value for t."

For the other species, reaction to cameras was observed. The time interval data should therefore be manually inspected for each of those 4 species.

EJH: Yes, to avoid bias when modelling the detection function and estimating encounter rate and abundance, detections in which animals react to cameras should be censored. See https://peerj.com/articles/13510/ and the article Dr. Palencia recommended.

Data from reactive animals are potentially informative about recovery times, but potentially misleading with respect to retrigger times. If animals remain in the FOV and move sufficiently to retrigger the camera while reacting, times between consecutive triggers are informative about camera recovery times. However the same animal behaving normally (not reacting) might move less and so the retrigger times for observations included in the data would be longer than for reacting animals. For this reason, I might recommend first censoring observations of animals reacting to cameras before estimating t from the data.

I would first need to clean a bit the dataset for each species and remove any images where I have other species on the images (we did see sometimes wild pig and pheasant or junglefowl on the same images, so these photos should be removed from the dataset).

EJH: Sorry, I'm not sure it's necessary to censor these data when modelling the detection function or estimating encounter rates and abundance, but species-specific retrigger times would be affected.

This is where I am not fully sure about the next step. From what I understood, I need to filter for only those pictures belonging to the same animal to get a ‘clean’ time interval distribution that is not contaminated with time intervals between pictures belonging to different animals.

EJH: You could, and this was the approach we took with the "checked dataset" subset of the data, (see section 2.4) but you don't necessarily need to. That's the advantage of the methods described in the article. From section 2.3:

"A direct approach to estimating t would be to examine consecutive images, determine whether successive detections are of the same individual and simply take the sample mean of the intervals between successive detections. However, this requires a lot of time and effort to track an individual within one passage through the field of view and to record the intervals within such a series. By contrast, a statistical approach to derive non-observation times (hereby referred to as ‘mean time intervals between triggers’) from time interval distribution data gives results very quickly."

I have in my record table the number of individuals recorded on images, especially for wild pig where we can have images with 7 individuals. Do I then need to first remove those images from the dataset (images with number of individuals higher than 1), and then look through the rest of the images to be sure it is the same animal all along?

EJH: With CTDS data, we do not recommend estimating group density and group size, because group size may be underestimated. We recommend treating individuals rather than groups as the unit of observation. The assumption that observations are independent is already severely violated due to small t and the inclusion of multiple observations from the same animal during a single pass through the area monitored by a CT. The data you use to estimate detectability and abundance should have a separate row and observation of distance for each animal in the FOV of an image included in the dataset. This should not affect estimation of t because they are all detected at the same time (no time interval between these detections). "Intervals" of 0 seconds for > 1 animal in the same image should not be included in data used to estimate t.

For example for wild pig, sometimes, I have one individual entering the FOV then going out, then maybe 5 seconds later a second individual entering the FOV. And then the two individuals again enter the FOV at the same time. Since I filtered for only images with 1 individual, I won't get that images with 2 individuals. But I do know in my case it is 2 different events because later on I will have those 2 individuals with all their babies for example. How do I deal with that?

EJH: Because the pigs travel in groups, it will be difficult to separate retrigger times of animals that remain stationary in front of the camera for a few seconds from consecutive triggers from different animals. That could make defining an appropriate truncation time difficult because distributions of retrigger times from the same animals, and from different animals, will overlap. This might require you to take the "direct approach" and filter for pictures belonging to the same animal. Don't filter for images with only 1 animal.

I see in the csv file that they are four columns that I do not have in my record table: delta.time.secs, Beh_numeric_event, ID, EventID. I assume the delta.time.secs can be obtained using the package camtrapR and a threshold of 5 minutes for indepedent event can be used (everything less than 5 minutes is considered the same event or same individual and therefore will calculate the delta.time.secs between each photos). However in my data analysis I used a threshold of 60 minutes to filter for independent events, does it mean I need to use the same in this case? Also what does Beh_numeric_event represent? Are the two last columns used in the csv file of importance for the analysis? I assume the interval used in the R script is the delta.time.secs column.

EJH: Sorry I don't quite follow here. We include non-independent observations in CTDS data, and in the data used to estimate t. I also didn't conduct the analysis so would have to check on those column labels.

Another question is how to incorporate that "mean time interval between triggers' into my data analysis. If we take for example the siamese fireback in my dataset (pheasant), which reacted to the camera by getting very close to the camera and staring at it for long period of time. I have the effort in days for each cameras deployed, and based on the activity histogram, we estimate the daily time activity from 5:45 am to 18:15 am, so 12 hours and 30 minutes daily. So the daily Tk is 12.5 hours and then we need to multiply it by the numbers of days camera were active and then divide by t, which is the time interval between snapshots, so 2 seconds here. I assume therefore I would only change this 't' and replace by the one obtained following the methodology from the article.

EJH: Multiple issues here. Observations of reacting animals should be censored. If animals are only active from 5:45 through 18:15, you can set daily Tk to 12.5 hours (in seconds) and multiply by camera days and divide by t to get temporal effort. However, this is insufficient to account for the effect of animal activity on abundance estimates unless animals are continuously active for all of those 12.5 hours. You should still estimate activity level (following Rowcliffe et al.) within those 12.5 hours and include this proportion when estimating abundance (A-hat in Eq 2 of the article). We estimate activity level within the time included in our survey duration (Tk). In practice, it's often simpler to set daily Tk to 24 hours and estimate activity level over 24 hours. For most animals, activity level over 24 hours is < 0.5 (they rest for about half the day but also rest during the half of the day when they are active).

EJH: Yes, when estimating effort we will divide by t, and your 2 seconds would be replaced by estimated t.

But what about all the distance estimations I did at the interval of 2 sec on my images for each species ? In my case, I filtered for all images in the record table that finished by an even number for the time and only estimated the distances for these images.
In the article, to do the data analysis, the first photo of each photo series was used a snapshot moment. Do I therefore need to review this step and do distance estimation for the first photo of each photo series? In my case, the settings used are the following: "Cameras were programmed to operate all day and to record one photo at each activation, with the minimum triggering interval between activations (0.6 seconds)." Or if I found for example a t of 6 seconds, do I need to take only images in the data analysis every 6 seconds ?

EJH: If your cameras were programmed to record a single image when triggered, and performed well enough to detect animals after < 2 seconds (i.e., if recovery time in the field really was < 2 seconds), and if many (non-reacting) animals contribute more than one observation of distance with t = 2 seconds, then perhaps you don't need the methods described in this article? You could use predetermined moments 2 seconds apart and select the images that most closely align with those moments. If recovery times are reliably short (< t) then they won't cause bias. Retrigger times would only be a serious issue if animals frequently remained stationary in the FOV such that they went undetected at snapshot moments, but you could check for that in your data and correct the data instead of applying these methods.

EJH: You should select the images that most closely align with snapshot moments. Filtering using even numbers sounds like it could be effective, provided you're not dropping observations because time stamps are not rounded to the nearest second. E.g. we don't want to discard both images if they occur at 1.5 and 2.1 seconds, we'd chose the one at 2.1 which closely aligns to a predetermined snapshot moments.

EJH: If you use an estimated t, you would include all images of non-reacting animals in the data used to estimate abundance. Some would be < estimated t apart, but that's fine because it's a mean across short and long intervals. So, if estimated t is 6 seconds, you should set t to 6 seconds not 2 seconds, but do not "take only images in the data analysis every 6 seconds".

EJH: If you can use a "checked data set" (consecutive images of only the same animals, excluding reacting animals) and find that estimated t is < 2 seconds, I don't think you need these methods. Just use t = 2 seconds and select the images that fall closest to predetermined moments for estimating the detection function and abundance.

Emeline Auda

unread,

Apr 22, 2024, 1:20:24 AMApr 22

to Eric Howe, distance-sampling

Hi Eric,

Sorry for the lack of answers on my side, as I was on long holiday and then started several new projects. I have finally time to go back to this camera trap dataset and continue the data analysis.

I have censored all data where animals are seen reacting to the camera (mostly attraction and freezing) and I will follow the methodology described in the article recommended by Dr. Palencia, the removal+ignore-HR method for CTDS for animals that reacted to the camera (siamese fireback, wild pig, macaque and junglefowl). When checking the data for the two squirrel species, they tend to remain stationary in the FOV and as you mentioned do not retrigger the camera until they move (they will not move like for 5 and 6 seconds and stay on the branch of a tree before moving again, which you can see in the timestamp). What do you mean in this case to correct the data instead of applying these methods?

After checking the data, I would say most cameras seems to work well and we get a t < 2 s (except for the squirrel species), so I indeed might not need to do the methodology described in the article. However, we started the survey a second time this year and the cameras seemed to have a lot more issue (as they were now bought 4 years ago) and seems to not be as reliable as before with the recovery time. In this case, I will probably have to use the methodology mentioned in your article for this second round of camera trap.

Yes for estimating activity level and treating individuals as unit of observation, this is indeed how I did it. However, I am now having an hesitation on how I selected my snapshot moment, as it seemed to not be the best way to do it. The issue is that I do not have the milliseconds information in my timestamp and in my EXIF data so I am indeed wondering if I am not dropping observations. How would you recommend to do in this case? I assume I might need to go back to the selection of snapshot moment and do it manually instead of how I did it. Also when I checked the images that were selected as snapshot moments (so where time finishes by an even number), I noticed that the metadata extracted by Exiftool and camtrapR for the time was sometimes slightly different from the timestamp written directly on the image. There seems to be a difference a 1 second, sometimes no difference, less frequently a difference of 2 seconds between the metadata of the image and the timestamp on the image. I suspect that the camera might round up to the nearest second and there could also probably some sort of delay during image processing in the camera, which leads to the time difference. Just wondering if it was the right call to work on metadata values (and not timestamp on the image) to extract snapshot moment. Given all these information, I am wondering how I should proceed to be as accurate as possible with the selection of snapshot moments.

Thanks as always for all the help, hopefully I will be able to complete the data analysis for this dataset soon based on your answers.

Best regards,

Emeline

--

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.

To view this discussion on the web visit https://groups.google.com/d/msgid/distance-sampling/0f19a556-79a8-4098-88f6-44878618f4b2n%40googlegroups.com.

Eric Howe

unread,

May 23, 2024, 10:06:32 AMMay 23

to distance-sampling

Good day Emeline,

Sorry for the delay, and apologies if I misinterpret some of what you wrote.

Firstly, snapshot moments are not relevant to estimating activity level. Activity level should be estimated from independent times of detection. Only one time of detection should be included from each independent encounter between an animal and a CT. Some researchers use start times of videos, others use a time criterion, such as to not include > 1 time of detection of a species, at a location, within any e.g. 15-minute time window. Small differences in times of detection (seconds) should not affect your estimates of activity level.

Great that you censored observations of animals reacting. I'm not certain the ignoreHR method will always be appropriate but it seems reasonable here.

"When checking the data for the two squirrel species, they tend to remain stationary in the FOV and as you mentioned do not retrigger the camera until they move (they will not move like for 5 and 6 seconds and stay on the branch of a tree before moving again, which you can see in the timestamp). What do you mean in this case to correct the data instead of applying these methods?"

EJH: If you're confident that it's the same squirrels remaining stationary, I suggest adding detections of that animal to the data at the snapshot moments when it was within the field of view but did not trigger the camera. The distance would be the same as the last observation before the animal stopped moving. This would remove bias caused by missed detections due to insufficient movement to trigger cameras.

"However, we started the survey a second time this year and the cameras seemed to have a lot more issue (as they were now bought 4 years ago) and seems to not be as reliable as before with the recovery time. In this case, I will probably have to use the methodology mentioned in your article for this second round of camera trap."

EJH: This sounds reasonable.

EJH: I'll assume the below is about selecting images from which to measure distances, not to estimate activity level.

"The issue is that I do not have the milliseconds information in my timestamp and in my EXIF data so I am indeed wondering if I am not dropping observations. How would you recommend to do in this case? I assume I might need to go back to the selection of snapshot moment and do it manually instead of how I did it. Also when I checked the images that were selected as snapshot moments (so where time finishes by an even number), I noticed that the metadata extracted by Exiftool and camtrapR for the time was sometimes slightly different from the timestamp written directly on the image. There seems to be a difference a 1 second, sometimes no difference, less frequently a difference of 2 seconds between the metadata of the image and the timestamp on the image. I suspect that the camera might round up to the nearest second and there could also probably some sort of delay during image processing in the camera, which leads to the time difference. Just wondering if it was the right call to work on metadata values (and not timestamp on the image) to extract snapshot moment. Given all these information, I am wondering how I should proceed to be as accurate as possible with the selection of snapshot moments."

EJH: Programming cameras to record video ensures we can extract images at predetermined snapshot moments, but if cameras are programmed to record single images or short bursts, the goal should be to select the images that most closely align with snapshot moments. E.g. if an animal remains in the FOV for 5 seconds, and you're using t = 2 seconds, then there should be at least 2 observations of distance recorded during those 5 seconds. We prefer predetermined snapshot moments to avoid positive bias in the observed distances, but as long as t is small relative to the speed of animal movement, such that many animals are detected more than once during a single encounter with a CT, then any bias in observed distances should be small. E.g. distances recorded e.g. 0, 2, 4, etc. or 1, 3, 5, etc. seconds after the camera was triggered will be similar to distances that would have been recorded at the snapshot moments.

Hope some of this helps,

Eric

Reply all

Reply to author

Forward

0 new messages