My third publication from my PhD work is out: Temporally adaptive acoustic sampling to maximize detection across a suite of focal wildlife species, published as an open access article in Ecology and Evolution. The Github repository accompanying the paper is located at https://github.com/cbalantic/temporally-adaptive-sampling.

This might be the most novel and unique intellectual contribution from my PhD work. It was also the most “niche” application and the trickiest to explain and justify. The need for our approach likely wouldn’t be immediately obvious to most, because the methodology arose as a consequence of our use of smartphones as autonomous remote monitoring devices, which isn’t a widespread practice in ecological monitoring (yet, at least).

Let me back up. For my PhD work, we built 20 smartphone-based autonomous monitoring units that sat in weather-proof cases, attached to a solar panel for power. For about two years, these autonomous monitoring units hummed away independently collecting data, taking audio recordings at permanent stations in the desert of southeastern California.

Our original hardware approach involved some soldering (above, left), followed by basic assembly of the components within a Pelican case (above, two middle photos). The case containing the phone was attached to a u-post and connected to a solar panel for power (above, right).

Each day, the smartphones sent their recordings over the cellular network directly to our long-term storage, making the audio recordings available for analysis on a daily basis, even though we were physically located about 4,000 km from our field site. Pretty cool, and pretty convenient for performing rapid analyses. Sometimes I’m still amazed we were able to pull this off!

A drawback of using smartphones this way is that cellular data plans are expensive. You certainly can’t take recordings 24/7, and you probably don’t want to take recordings that are too long – large file sizes make it harder to send your data over the cellular network (particularly if your monitoring units are located in a low cellular signal area, which is another huge consideration when implementing smartphone-based monitoring).

So, cellular data plans are expensive and your budget is limited. You run into the familiar scientific conundrum of constrained sampling resources. In our case, with a prototype of 20 smartphone-based monitoring units, we had enough money to budget for 9 minutes of audio recordings per site, per day. Only 9 minutes a day! Doesn’t sound like a lot at first, but then I think back to past avian ecology field work I’ve done, where we visited each of our monitoring points for a 5-minute point count once a week for ~10-12 weeks during each bird breeding season. Compared to a point count, our smartphones in California were collecting nearly 13x as much data as that on a weekly basis. Nine minutes per day adds up quickly across many days and sites.

However, how should you allocate those 9 minutes across the day if you’re looking to detect more than just the standard dawn chorus? In our California work, we were not only looking to detect breeding birds that sing primarily on spring mornings; we were also interested in recording birds that are vocally active at night (Common Poorwill, Lesser Nighthawk), coyote howls, and the breeding call of Couch’s Spadefoot, an awesome desert amphibian only acoustically available after a monsoon rain in the late summer or fall. So, we wanted to sample in the morning sometimes, and at night sometimes, and under conditions when our desert toad might be available (if present).

We needed a way to remotely schedule audio sampling each day, in a way that would adapt based on weather predictions AND on the probability that other target species vocalizations had previously been captured on a recording (if the species is actually present at a monitoring site, that is!). We didn’t want to waste any precious sampling resources. Thus, temporally adaptive sampling was born, and if you’re interested in the specifics of the algorithm and workflow we came up with, I hope you’ll read the paper, and reach out to me if you have any questions. We ultimately were only able to deploy 16 remote monitoring units, and this small field sampling size meant we ended up conducting this work as a simulation-based study, though we did work out the protocol and mechanics in the field as well. Based on our simulations, the benefits of an optimized schedule are highest when your study season is short, your sampling budget is small, or your target species are only available under specific temporal or environmental conditions.

Smartphone-based monitoring methodology, and the temporally adaptive sampling opportunities it yields, is not necessary for all research circumstances – probably not for most research circumstances in 2019. Looking forward toward a 5G world and increasing telecommunications connectivity and technology, however, I think the general idea of temporally adaptive sampling could have broad appeal for a variety of research challenges in automated ecological monitoring.