Continuous-time acoustic spatial capture recapture with moving microphone arrays : estimating gibbon abundance with drone-borne microphones

Abstract

The focus of this PhD thesis is on finding an effective method to estimate the group density of gibbon species. Gibbons are challenging to observe in the wild because they are low-density species inhabiting the upper canopy of rain forests. Estimating gibbon density using visual surveys is thus not recommended as it would require enormous effort. However, gibbons’ territorial calls can be heard from large distances, creating the opportunity for acoustic surveys. Distance sampling is widely used to estimate gibbon density but it requires estimation of distances and a moderately large sample size to ensure reliability of estimates.

We developed a new acoustic spatial capture-recapture (aSCR) maximum likelihood estimator for linetransect surveys that are conducted continuously in time, we call it continuous-time aSCR, and we implemented it in an R package. The method does not require either the estimation of distances or a large sample size to yield accurate estimates. Unlike distance sampling, it uses information from received sound alone, including times of arrival (TOA), signal strength, and time of detection that was used to gather information about where the detector was when it did and did not detect gibbons.

Drone-borne microphones open the possibility to conduct line-transect acoustic surveys from the air, allowing surveying of regions that would be impossible to access from the ground due to difficult terrain and vegetation. We designed and built an acoustic drone, i.e. a special drone equipped with microphones and a tiny multi-channel recorder. We tested the acoustic drone during a fieldwork in Laos and we successfully conducted a playback experiment in Switzerland during which the drone collected acoustic data in a way that mimics a real line-transect survey. We devised a noise-cancelling software that amplifies the signal-to-noise ratio by partly removing drone noise and allowing to increase the number of detected calls by more than 40%. Noise-cancelling increases the number of detected calls but it can distort the time domain of the audio recordings. This makes sound source localization using only TOA imprecise. The use of signal strength, and especially the information about where the detector was when it did and did not detect gibbons, improved localization performance.

We used our aSCR software to estimate location, call rate and calling direction of a one directional sound source. Assuming that a source was omnidirectional led to an underestimation of the call rate and less precise localization. A directional acoustic model was able to yield accurate call rate, localization, and calling direction. We simulated the same scenario of the playback experiment and demonstrated that assuming omnidirectionality when the source is directional leads to a systematic underestimation of the call rate and overestimation of the detection probability.

We simulated different scenarios with multiple omnidirectional sources to evaluate the performance of our method in estimating gibbon group density under different conditions. Although the developed method models call times using a random pattern generated by the Poisson Process, we demonstrated that we can use it to get accurate group density estimates even if we assume that call times are following a clustered process closer to the real call pattern of gibbons. We also show that we can get accurate density estimates without information about TOA, using only received signal strength and time of detection. We show that knowing when the array was when it did and did not detect calls, not only improves localization when TOA are not informative, but it also allows us to estimate density when TOA are omitted. This opens up the possibility to use smaller single-microphone arrays.

Finally, compared to acoustic line-transect distance sampling, our estimator yields unbiased group density estimates also when the collected acoustic information is scarce. This allows one to calculate distances between source and microphones even if the source was detected only once (given some other sources that have been detected more than once), expanding the scope for reliable inference about population size to situations of data scarcity that would otherwise require more effort to yield reliable group density estimates. In theory, our method can be applied to vocalizing animals other than gibbons, either to just localize sources or estimate their abundance. This opens up the possibility to test continuous-time aSCR for wildlife monitoring on a broad scale.