Desert Oasis

This year I decided to expand the acoustic monitoring that I’ve been conducting at the River Fork Ranch in northern Nevada.  The Nature Conservancy has been supported of my efforts by graciously providing access to preserves, so I decided to try to monitor at other TNC preserves.  In April of this year, I headed south to the Amargosa Preserves near Beatty, NV.  This set of preserves protects a section of the Amargosa River, a unique desert stream that creates an incredibly rich habitat and important stopover for birds migrating up portions of the Pacific Flyway.  The Amargosa River itself is a fascinating system.  Not a river in the usual sense, it’s a long series of springs near the headwaters, then a long series of desert marshes that periodically goes underground.  It’s only recognizable as a stream in the stretches along Beatty, Nevada, and Shoshone and Tecopa, California.  Its ancient waters bubble up to create the marshes and springs at Ash Meadows Wildlife Refuge (home of the Devil’s Hole pupfish). After flowing more than 100 miles south from it’s headwaters, it turns north for another 80 miles where it terminates at Badwater Basin in Death Valley National Park, the lowest point in North America at 282 feet BELOW sea level.

The Nature Conservancy recently acquired the 7J Ranch, at the headwaters of the Amargosa.  This cattle ranch includes a series of man-made ponds, which provide water for cattle and wildlife.  The ponds support a lush habitat of cottonwoods and willows, as well as wet meadows and emergent pond vegetation.  The ponds and the river system through the surrounding Oasis Valley are biodiversity hotspots that are well-known to birders.  The area has its own unique toad species – the endangered Amargosa toad (unfortunately for a sound recordist, they don’t vocalize). The headwaters of the Amargosa River is also at the ecotone between two biomes: the Great Basin Desert and Mojave Desert.  All of these factors made the 7J Ranch a pretty exciting place to set up a monitoring station.

As I was really just exploring the feasibility of acoustic monitoring in this area, I left one recorder (Songmeter mini) for 5 weeks between mid-April and late May.  I expected that this would capture the peak, although not the entirety of the spring migration.  So unlike many bird surveys that quickly assess what birds are in an area (hopefully breeding), I was more interested in who was moving through and who was staying, as well as when different birds were singing.  In addition, with a permanent water source, I was curious what amphibians and mammals might be vocalizing. Very few acoustic monitoring programs have attempted to look at the phenology (seasonal changes) of birds in an area, so I was also curious if this method was a viable way to study migration.

I set the recorder to record 10 minutes of every hour, 24 hours a day, except during the middle of the day when it skipped a few hours.  When I collected the recorder in late May, it had made 554 10-minute recordings.  As soon as I got home, I set about going through the recordings and identifying the species vocalizing.  Like much of both the Great Basin and Mojave deserts, afternoons are typically windy, especially in the spring.  The winds hampered many of the recordings, so if I establish a more permanent recording station, I will have to find a way to better protect the recorder from the wind.  The site is also right outside of Nellis Air Force Base, and the noise from the military jets sometimes swamped the recordings.  Highway 95 was a few miles to the west, and if there was a westerly wind, the noise of semis and motorcycles appeared in the recordings.  So not the ideal place for a recorder, but in all, I detected 95 species, and observed a couple of species in the area that did not occur in the recordings.  Although it was a rather windy location, sometimes in the pre-dawn and dawn, the noise settled down revealing a great marsh soundscape:

I found the rare quiet evenings even more compelling, with a nice symphony of insects, frogs, birds (Common Poorwills, Ruddy Ducks, and Least Bittern), and mammals (burros):

The limiting factor for the project was the time it took to go through the recordings.  It took more than two months, although I did take several other recording trips during that time.  So, finding a way to analyze the recordings more quickly would allow me to expand to other areas and seasons.  Recently, the Cornell Lab of Ornithology has developed a desktop app, BirdNET, that can process sound files using machine-learning AI to identify the species of birds present.  This sounded like it might be very helpful, so I downloaded the program and tested it out.  Right away it was obvious that recordings with a lot of wind or jet noise really confused the program.  So, I limited the files to those recorded between 3 am and 9 am.  The output for each file is a list of the birds it identified (starting with a pre-populated list based on location), with a confidence estimate.  I played with the results quite a bit to try to find a confidence level that would include most of the birds I identified, and limit those that were simply not in the recordings (false positives).  This ended up being a confidence level of around 45%.  Other researchers have used much higher confidence levels, but they were trying to weed out potential transient birds, whereas I wanted to try to detect everything in the area.  For birds that BirdNET detected that were not on my list, I checked eBird to see if there were any detections in the area.  If eBird did not show that the bird had been detected in the local area, I removed it from the results.  This, of course, limits the “rare” bird sightings, but also limited the false positives. 

It was a lot of number-crunching to compare my results to those of BirdNET.  To simplify the results, I summarized the finding for each day, rather than each recording.  Overall, there was 88% agreement between my detections and BirdNET, and 12% disagreement.  That’s not too bad, considering not only the background noise, but confusion caused by birds like the Great-tailed Grackle, which produced loud, long, highly-variable songs.  Or by Northern Mockingbirds which were so skilled at mimicking other birds that it confused the program completely.  Luckily, mockingbirds only showed up on a couple of occasions.   

Although 88% agreement isn’t bad, the false negatives and false positives created some problems.  For example, the program confused the last couple of hoots of a Great Horned Owl call with those of a Long-eared Owl.  It was easy to see the error in a spectrogram.  But Long-eared Owls have been seen in the area (but not detected by me), so it would not be good to just reject all of the Long-eared Owl detections without reviewing each one.   For another example, the program missed most of the songs of the Red-winged Blackbird, a very vocal resident of the ponds.  It had no trouble with Yellow-headed Blackbird calls.  Apparently the program was trained on species mostly in the eastern US, and it doesn’t recognize a western Red-winged Blackbird. 

So while the program has great potential, I’m not yet comfortable with the level of error.  For now, I’ll stick with manually reviewing the files, which has the added benefit of detecting  vocalizing amphibians and mammals.  To keep from going totally insane, I’ll subset the files to just the early morning hours, with a couple hours in the late evening for frogs and owls.  The rest of the files will be stored in case a better identification program comes along or some volunteers step up to help me.

For my previous analysis of recordings at River Fork Ranch, I used a program called Arbimon. Arbimon compares small snippets of a recording to a template of individual calls of a species. This technique is good if you want to look for particular species, or even particular calls of a species, but it is not good for examining what is present in an area, as I was trying to do at the 7J. If it turns out that there is a particular species we want to monitor, for example, a Willow Flycatcher (declining species) that showed up briefly at the ponds, Arbimon might be a good tool for that.