DSCT - snapshot moments

[Jamie McKaughan]

Hi all,

I saw a recent email chain on snapshot moments, but I wasn't sure it was quite the same query - apologies if it is!

I previously discussed using t = 1s with Eric Howe, with a camera set up of six rapid fire shots (0.3s apart) with a 1s recovery time between bursts. I then filtered my photos to take only photos every 1s apart, where with any photos that had the same timestamp I took the second image.
As has been found by Corlatti et al 2020 and several other people I have seen - cameras haven't necessarily performed as per their spec.

If for example my cameras were set to take 6 images 0.3s apart with a 1s delay before being able to trigger again and take six more, the camera's actual recovery time is normally at best 10s (not 1s), and consequently although I would get six images during an 11.5s total period (5x0.3+1x10) only one or two images actually fell on a 1s snapshot. The suggestion is that if I used a mean interval instead - I could instead have six 1.9s moments in that 11.5s period and therefore should use all 6 images obtained assuming that the mean interval between photos is 1.9 per camera trigger, even though the images themselves don't actually fall 1.9s apart. It is clear that data may/has been lost from where the animal is still present in front of the camera but the camera hasn't recovered quick enough to obtain further images of it, a clear issue with using still images instead of video, but I wondered if this approach would be robust as a way of trying to account for this.

I was wondering what the implication would be of taking this mean snapshot moment and then being able to use all the data I collected. I didn't quite feel right with this from how I understood the snapshot element, as I felt the snapshots were predetermined moments in time in the day rather than predetermined average time between photos and the mean would not really account for the 'snapshot' element at all leading to further violation of the assumptions and lead to an upward bias in density estimates - but I wanted to get your take on this.

Many thanks
Jamie

[Stephen Buckland]

Jamie, I can’t see any theoretical reason why taking the mean time won’t work.  It doesn’t matter that you miss some shots of an animal, provided animals don’t manage to pass right in front the camera and be missed altogether.  Others with practical experience might see a problem with doing this, but I don’t.

 

Steve Buckland

--
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/483768eb-ed02-4e1a-b479-bdc6427cbd67n%40googlegroups.com.

[Eric Howe]

Hi Jamie, and thank you Steve,

Steve, is it true that mean delay time should be calculated from observations of animals that are close to the camera (where g0 as a function of distance is close to 1.0) to avoid overestimating delay time?

I was going to suggest either

(1) including all photos and using the mean delay time to calculate "effective" t, in which case t  = (mean delay time + 1.5) / 6. Then, t = 1.9s if delay time = 10s as in your example Jamie.  

(2) selecting two photos from each triggering event, in which case we get only 2 opportunities to measure distances per delay time + burst time, so t = (mean delay time + 1.5) / 2. I only suggest this because if animals move slowly, repeat detection 0.3 seconds apart might not be more informative regarding the shape of the detection function than repeat detections e.g. 0.9 seconds apart. If animals don't move far between most images 0.3 seconds apart, but frequently trigger the same camera after > 10 seconds, selecting two images from each triggering event could be more efficient. 

Ideally, snapshot moments would be predetermined and independent of camera triggering times. If we dispense with predetermined moments, we risk some upward bias in observed distances. Excluding the first image of each burst might reduce the potential for bias. However, the effect of slight upward bias in observed distances on abundance estimates would be small compared to assuming animals were available for detection every 1 second during a 10+ second delay.

Eric

[Stephen Buckland]

You’ve thought more about this than me, and I suspect you’re right.  If camera properties are known well enough, it would be good to do a simulation study, with accurate simulation of what each type of camera does, and a suitable animal movement model, to assess bias, and how best to correct for it!

 

Steve

To view this discussion on the web visit https://groups.google.com/d/msgid/distance-sampling/a8a7c3dc-1e16-4d6b-a1c8-d2c775b644e4n%40googlegroups.com.

[Jamie McKaughan]

Thank you both for your responses - very helpful and good to understand the scale of effect the distance bias would possibly have compared to the availability assumption.

Best wishes,

Jamie

[Jamie McKaughan]

Hi all,

I have produced new density estimates following this change in snapshot interval (moving from using t of 1s to an effective t of 1.9s) and then using all my data, rather than just the images falling at 1s moments. The estimates I now have are huge and I wondered if this could purely be from this change? The upturn seems a lot more than I anticipated might occur from 'some upward bias' from not using the predetermined moments.

I don't quite understand why dispensing with predetermined moments risks some upward bias in observed distances, beyond the general principle of DS, that observations nearer the point will be made more easily and often than those further away.

I have pasted a table of my results below if they are useful as a reference.

Many thanks

Jamie

[Eric Howe]

Hi Jamie,

Moving from t of 1 to 1.9 reduces temporal effort by almost half, so would almost double the abundance estimates for the same number of observations. If you also added more observations, abundance estimates will increase proportionally. If we double t and the n, the abundance estimate will quadruple (assuming the new observations don't cause a big change in the detection function). Changes in t and n will have bigger effects on the abundance estimate than small changes in the distribution of observed distances. 

We expect upward bias in distances observed upon initial detection, e.g. if only the first image of each animal was included in the data. This is because animals approaching the camera from in front will be detected at the maximum sensor range, and not again, even if it proceeds towards the camera, where detection probability is supposed to be high (is assumed to be higher than at longer distances [monotonically increasing]). If cameras can be triggered again after recording an image, and we include those subsequent observations in our data, and the camera performs well enough to ensure that t is short enough that we get multiple observations of distance to some of the the same animals(s) as they move through the zone of potential detection, we're not as concerned about snapshot moments being predetermined. Observations of distance should be unbiased if (1) snapshot moments are predetermined, in which case we can chose any value for t without causing bias, or (2) t is small relative to how long it takes animals to travel through the zone of potential detection, such that the data describe animals' locations as they move through the sector, not only upon initial detection. 

Could you explain what the table of results shows? I assume the values are estimated densities but I don't understand the column headings. 

Eric

[Jamie McKaughan]

Hi Eric,

Thanks so much for your response. Sorry - I meant to change the headings! The first column is density per sq km using all photos and the effective t of 1.9s, the second column were my original estimates with a filtered data set for t=1s. I also applied an effective detection angle to both.

Looking at this then, I would suggest that the use of my effective t is not providing an accurate representation of my species, especially compared to my initial estimates. If I know that on average for every trigger of my camera I lose 10s of effort from my camera, i.e. it does not trigger for another 10s, is there any value in using a t=1s snapshot moment as I did initially, but for each camera location remove 10s x total number of triggers for that camera and use that as a basis for my total survey effort, effectively removing the remaining estimated time cameras were inactive for? Or is that not really scientifically sound enough?!

Many thanks

Jamie

[Stephen Buckland]

I think if your camera has a recovery time of 10s, then you cannot use a snapshot interval of less than 10s.

To view this discussion on the web visit https://groups.google.com/d/msgid/distance-sampling/1d474b6b-dd58-466a-9842-e8dc2e07450an%40googlegroups.com.

[Eric Howe]

Hi Jamie,

From your earlier post " If for example my cameras were set to take 6 images 0.3s apart with a 1s delay before being able to trigger again and take six more, the camera's actual recovery time is normally at best 10s (not 1s), and consequently although I would get six images during an 11.5s total period (5x0.3+1x10) only one or two images actually fell on a 1s snapshot. The suggestion is that if I used a mean interval instead - I could instead have six 1.9s moments in that 11.5s period and therefore should use all 6 images obtained assuming that the mean interval between photos is 1.9 per camera trigger, even though the images themselves don't actually fall 1.9s apart. "

We suggested you could include all images if you calculated effective t including delay time and burst time. Then we're saying we obtained those 6 observations over ~ 11.5 seconds, right? Trying to think about how this affects the data collected: If the animal actually remains in front of the camera for those 11.5 s, then we're accounting for the delay appropriately (we would have obtained 6 images at t = 1.9s if the camera functioned as intended, or as described by the model). However, if the animal moves through quickly and you get a burst of 6 images, we would only have obtained e.g. 1 or 2 images at t = 1.9, but we're including all 6. This might cause positive bias.

Also, the below, from one of your previous posts, suggests that with t set to 1 second you would only include one or two images per "burst".

"I previously discussed using t = 1s with Eric Howe, with a camera set up of six rapid fire shots (0.3s apart) with a 1s recovery time between bursts. I then filtered my photos to take only photos every 1s apart, where with any photos that had the same timestamp I took the second image

In that case, your more recent suggestion (below again) does make sense to me. Maybe more sense than pretending that 6 images obtained over 1.5 s were really obtained over 11.5 s. I think you'd want to include all triggers. E.g. if a non-target species or vegetation triggers the camera, you still miss the opportunity to detect target species for those 10 s. 

" is there any value in using a t=1s snapshot moment as I did initially, but for each camera location remove 10s x total number of triggers for that camera and use that as a basis for my total survey effort, effectively removing the remaining estimated time cameras were inactive for? Or is that not really scientifically sound enough?! "

 Sorry I can't give you a definitive answer about whether the estimates should then be accurate in this situation. 

All the best,

Eric

[Jamie McKaughan]

Great - thank you very much for your thoughts!

Best wishes

Jamie