Introduction

A total solar eclipse that leaves populated areas in darkness for several minutes is so rare that experiencing one is often a once-in-a-lifetime occurrence. Anecdotal evidence and research involving human observation, community science, radar analyses, and acoustical studies, suggest birds stop singing and engage in nighttime behaviors such as roosting during a total solar eclipse1,2,3,4,5,6,7,8,9. The April 8, 2024 total solar eclipse created a path of totality from Mexico through the United States and into Canada and was visible to millions of people, many of whom traveled great distances to experience this Great American Eclipse. Unlike the 2023 annular solar eclipse when solar obscuration reached a maximum of 90%10, this eclipse entirely blocked the sun with 100% obscuration along a wide swath of North America11. During an eclipse when the moon blocks the sun’s warming rays, relative humidity increases, air temperatures drop 1.5 °C with 97% solar obscuration and as much as 7 °C during totality12. The next total solar eclipse to cross largely populated areas of North America will not occur until August 12, 2045.

Anecdotes of how animals behave during total solar eclipses abound and many people have heard about or observed animals initiating their evening routines before or during totality, including birds roosting or altering their vocalization patterns3. These observations date back to 16th century writings which mention that “birds ceased singing” and “birds fell to the ground”1.

Anecdotal results do not universally support the idea of bird silence during totality. During the 1963 total solar eclipse, a team from the Cornell Lab of Ornithology recorded birdsong from a secluded spot with many singing birds but little human-caused interfering noise. While bird song decreased noticeably as darkness descended, their recordings showed that singing never stopped completely, and observations shared by others similarly indicated that birdsong diminished without completely ending2.

Due to the rarity of total solar eclipses, few peer-reviewed studies of animal13 or bird behavior during a total solar eclipse have been published. In 1932, the first known study gathered 498 observations from the public, game wardens and naturalists, 44% of which were conducted within the zone of totality. They found “the evidence is absolutely conclusive that birds of all kinds showed practically no unusual or abnormal behavior in regions where totality was 98% or less.” Additionally, they found overwhelming evidence that “most birds showed some reactions of an unusual nature to totality” with several reports of wild birds exhibiting nighttime behaviors such as roosting and silence. They also indicated that there was limited evidence of differing reactions between bird species1.

More recent studies of total solar eclipses include one in 1998 conducted via human observations which showed birds ceased foraging and sought roosts as totality approached4. During the 2017 total eclipse, researchers observed animals at a zoo that remained open to visitors. Lorikeets and cockatoos roosted and preened, behavior typically found near sunset, and flamingos exhibited signs of anxiety such as flocking and feather ruffling5. According to a news story, animal observations were conducted by these researchers during the 2024 total solar eclipse at a zoo closed to visitors that lacked 2017’s “fireworks, cheering and chaos.” This time, animals went through their evening behaviors but no signs of anxiety were observed, leading the study’s main author to note “This is seemingly really strong evidence that animals that exhibit anxiety during an eclipse are anxious not because of the eclipse itself, but because of the human reaction to the eclipse.”14.

The 2017 total solar eclipse also was studied using radar data to determine if there was a sunset-like increase in atmospheric bird activity around the time of totality. At some of the 26 sites experiencing 95–100% obscuration, decreased activity was noted leading up to totality, followed by a short low-altitude activity spike during totality, indicating that the eclipse suppressed diurnal activity but did not initiate nocturnal activity6. Preliminary, unpublished results from 13 weather radar stations suggested that the 2024 total eclipse suppressed the diurnal behaviors of aerial animals15,16.

One acoustical study of the late-summer, mid-day 2017 total solar eclipse made soundscape recordings from four sites within the path of totality. Distant sounds of howling, cheering, and yelling were recorded, in addition to birds and other natural sounds. Changes in bird vocalizations associated with the eclipse were found among late-breeding species that were likely still fertile and nesting at the time of the August eclipse. The authors suggest that the impact of the eclipse would have been more pronounced and impacted a greater number of species if it had occurred in the spring7.

Thus, anecdotes and research suggest: (i) the largest changes in bird behavior likely would be found at sites experiencing full totality; (ii) that observed behaviors were likely those associated with nighttime activities; (iii) that true measures of the eclipse’s impact on birds would be found at sites where human celebrations were not a factor; and (iv) that larger impacts would be found for species in breeding status.

Community (also called citizen) science was not a term used in 1932, but the study that year may be the first that relied on input from regular, non-scientific community members. Community science now is widely used for a variety of ecological research17,18 because it allows large, longitudinal studies of broad spatiotemporal scales and on private lands9. Participants also benefit from improved nature connectedness, happiness, and sense of a worthwhile life19. Challenges with community science-based research using platforms such as eBird and iNaturalist stem from errors resulting from variable observer quality and knowledge1, biases in sampling8, and sampling effort9.

While community science relies on active participation, autonomous recording units (ARU) equipped for passive acoustic monitoring (PAM) offer a cost-effective alternative that requires no participant input20,21. PAM use has expanded22 and shown great promise23 as a tool to non-invasively detect and monitor sound-producing animals’ abundance24, diversity25, distribution26, population changes27, and animal behavior13,28, especially in remote locations. PAM allows for simultaneous recordings at multiple sites, provides a permanent record of sound data, and reduces observer bias. However, it also requires time and resources to deploy and recover data from specialized equipment and to store and analyze vast quantities of bioacoustic data29.

PAM devices can now be equipped with neural networks capable of automatically identifying target species, including marine mammals30 and birds31,32,33. They are rapidly being adopted in ecology34 where they greatly reduce the storage and analysis challenges resulting from traditional PAM35. Challenges remain, including the need for large, labeled training datasets, and the ability to distinguish between target species and anthropogenic sounds36,37.

BirdNET is a deep artificial neural network (DNN) trained on thousands of bird and non-bird recordings that successfully identifies 984 North American and European bird species by sound32. BirdNET’s use with ARU38 greatly increases monitoring opportunities39 including those with a community science component40. As a new tool with great conservation potential, it has been tested, compared with point counts and surveys, and used with a variety of bird species39,41,42,43,44,45,46.

In this study we present the first results using the Haikubox network that we developed as a hybrid community science-based and PAM-centered project using machine learning to detect and classify bird vocalizations in real-time. When an individual consumer purchases and installs a Haikubox near their home, it connects to the Haikubox system and begins to acoustically identify birds. It accomplishes this using edge hardware running a moderately sized neural net and a cloud system running a custom-trained version of BirdNET designed to reduce false detections of human-related sounds. In this study we eliminated any Haikuboxes that had human detections from the start to the end of the eclipse. Thus, this is the first study to examine continent-scale responses of birds to an eclipse that also controls for their possible reactions to humans. This research could inform future work about how birds may respond to light pollution or sudden changes in light levels. Additionally, it shows the power of community science to collect wide geospatial animal PAM data which is free from human-induced behavioral responses.

Bird vocalization rates and solar obscuration

For this analysis, 344 Haikuboxes across North America (Fig. 1) were placed into one of five groups based on the percent maximum obscuration (0 to 50%, > 50 to 90%, > 90 to 95%, > 95 to 99%, > 99%). While the average bird vocalization rate during any 180 s (three minute) period was highly variable for all groups (Fig. 2), no large or sustained peaks or dips were seen on average in any groups with the exception of those where solar obscuration reached more than 99%. In this group, an increase in vocalization rate was detected 10–30 min prior to maximum obscuration, followed by a decrease in vocalization rate at the eclipse’s peak. A test statistic was calculated by comparing the pre- and peak obscuration mean vocalization rate at each site. “Pre-peak” includes 30–60 min before maximum obscuration and “during peak obscuration” includes the time of peak obscuration plus the following three minutes. This analysis and scatter plot showed high variability across all Haikubox sites (Fig. 3). Most sites where obscuration was greater than 99% showed a decrease in the vocalization rate during peak obscuration when compared to the pre-peak vocalization rate, however, some individual sites instead showed a small increase in vocalization rate during peak obscuration. The mean and standard error for each obscuration group were plotted (Fig. 4). The vocalization rate in the group experiencing totality showed significantly lower activity compared to the partial eclipse group (> 0–99% obscuration) using a permutation test (p-value = 0.0493; observed mean difference = -3.49; bootstrap 95% confidence interval − 5.79 to -0.91). There was no statistically significant difference between the 0.95–0.99 group and the sites with < 0.95 obscuration (p-value = 0.8).

Fig. 1
figure 1

Locations of Haikuboxes included in the analysis. The color of the dot denotes the maximum obscuration of the eclipse at that location, with yellow: 0–50% maximum obscuration, orange: >50–90%, purple: >90–95%, blue: >95–99%, and black: >99%.

Full size image
Fig. 2
figure 2

Bird vocalizations per 180s centered on the time of maximum obscuration from one hour before to one hour after the eclipse. Lines are the mean for Haikuboxes grouped by maximum obscuration as indicated in the legend.

Full size image
Fig. 3
figure 3

Difference in the average bird vocalization rate at obscuration peak (obscuration and three minutes after) to the average bird vocalization rate from 30–60 min before the eclipse. Negative values indicate a decrease in vocalization rate at the time of maximum obscuration compared with observations prior to maximum obscuration. a. all sites included in this study and the data scatter indicates that the difference between bird vocalizations before and during peak obscuration did not show a trend for most obscuration levels. Figure B zooms in on those sites at 90–100% obscuration and the difference between bird vocalizations before and during peak obscuration only decreased at those locations where obscuration was greater than 99%.

Full size image
Fig. 4
figure 4

Mean and standard error of the change in mean vocalization rate from before the eclipse (30–60 min before peak obscuration) to the obscuration peak (obscuration and three minutes after), grouped by obscuration. Those sites experiencing a total eclipse (greater than 99% obscuration) showed a significant change in the bird vocalization rate during peak obscuration compared to the partial eclipse groups, with a significantly lower rate of vocalizations. Data are binned as > lower bound and ≤ upper bound, for example, 0.95–0.99 means > 0.95 and ≤ 0.99. ‘n’ indicates the number of Haikubox stations in that group.

Full size image

Bird vocalization responses are highly variable

A detailed analysis of vocalization counts of bird species from Haikuboxes with more than 99% obscuration showed high variability in their responses to the eclipse, several examples of which are shown in Fig. 5. One or more American Robins (Turdus migratorius) sang throughout totality and dominated the soundscape at one site (Fig. 5A); Fig. 6 is a spectrogram from this site of an American Robin singing at peak totality. At another site, one or more Pine Siskins (Spinus pinus) which had been vocalizing to dominate the soundscape prior to totality ceased singing and calling during peak totality (Fig. 5B). Two sites where the Black-capped Chickadee (Poecile atricapillus) was the most vocal species during the study period showed different behavior. While vocalizations of this species began after peak obscuration at one site (Fig. 5C), the other site saw this species vocalize throughout peak obscuration (Fig. 5D). These examples highlight the variable responses seen between sites, both broadly among different bird species and within the same species.

Fig. 5
figure 5

Total bird vocalization rates for the two hours centered on the time of peak obscuration (time bin 0.0) at select sites. Each bar represents the total number of bird calls/songs identified for 180 s broken out by the top five most detected species at each location, along with an “Other” category that includes any additional bird species. Haikubox locations with 99.4% maximum obscuration, the top in (A) Kentucky and the bottom in (B) Maine, which highlight the variability between sites. The differing behaviors of one species, the Black-capped Chickadee, which was the most recorded bird at two Haikubox locations with 99.8% maximum obscuration, the top in (C) New York and the bottom in (D) Vermont. Species include American Robin (Turdus migratorius), Northern Cardinal (Cardinalis cardinalis), Carolina Wren (Thryothoros ludovicianus), House Finch (Haemorhous mexicanus), Brown-headed Cowbird (Molothrus ater), Black-capped Chickadee (Poecile atricapillus), American Crow (Corvus brachyrhynchos), Dark-eyed Junco (Junco hyemalis), Tufted Titmouse (Baeolophus bicolor), Song Sparrow (Melospiza melodia), Pine Siskin (Spinus pinus), American Goldfinch (Spinus tristis), Blue Jay (Cyanocitta cristata), and Boreal Chickadee (Poecile hudsonicus).

Full size image
Fig. 6
figure 6

Spectrogram of an American Robin singing during peak totality (99.36% obscuration) at one Haikubox site. This example shows the quality of bird vocalizations produced at some sites which experienced totality.

Full size image

The variability in the response to the eclipse at the sites where percent obscuration exceeded 99% showed no significant relationship between temperature, cloud cover, wind speed, humidity, and latitude and the change in vocalization rate (Fig. 7). An ordinary least squares regression of all these predictor variables had an adjusted r2 of 0.11, showing they have little explanatory power.

Fig. 7
figure 7

Relationship of weather and geographic factors to the change in vocalization rate for sites with greater than 99% peak obscuration.

Full size image

Discussion

The behavioral responses of birds to a solar eclipse over a large network of continuously monitoring devices showed varying responses, and we believe it is the first study capable of reviewing bird vocalization behavior to look for differences at sites with a range of solar obscuration. This granularity mattered: a broad decrease in vocalization rates was consistently found only at locations which experienced more than 99% total solar obscuration. If we consider that peak obscuration only lasts a few minutes and even 95% obscuration results in light levels approximating what a bird might experience during a normal cloudy day, it should not be surprising that birds did not consistently alter their behaviors in response. Studies have shown that the timing of the dawn chorus of birds can be influenced by factors including cloud cover, temperature, and precipitation47.

In addition to changes in light levels, there is a lagging decrease in temperature and wind speed accompanying a total solar eclipse, with minimum temperatures reached at 13 min7. Interestingly, the decrease in vocalization rates at sites experiencing totality started at the time of totality and extended for about 12 min. The decrease in vocalization was not centered on totality, but the time period just after peak obscuration. This study cannot determine whether it was the decreased light levels, temperature or wind or a combination of these factors that led to decreased vocalization rates. Another study using radar to study avian activity found that activity increased with decreased temperature, wind, and visibility, and increased cloud cover48. In the context of an eclipse one might expect an increase in activity due to decreased temperature and wind, but that might also be counteracted by decreased activity associated with cloudy conditions. Since our study focused on vocalizations, it is possible that there could have been increased flying activity associated with less vocalization.

The responses of individual species varied between Haikuboxes in areas of totality, so it is not possible to say that all birds or even specific species universally responded to the sudden darkness in the same way. Even on a normal day, many factors may impact whether an individual bird vocalizes or whether it can be recorded. These include the individual’s breeding or hormonal state, interactions with other birds or animals (including humans), and their movement into or out of recording device’s range (e.g., an individual Haikubox). It is likely that one or more of these factors influenced bird behavior in this study. For example, the American Robin, an early migrant and breeder, and Black-capped Chickadee, a non-migratory species, were recorded singing nearly continuously during totality, possibly to demonstrate their breeding readiness through dusk chorus vocalizations49. The early April timing of the eclipse meant that many migrating birds likely had not yet reached northern Haikubox sites.

Human influence on bird vocalization behavior was controlled for in this study. Haikubox owners were reminded via notices on the website and through newsletters to be quiet if possible during the eclipse, yet only 53% of the Haikuboxes (344) stayed connected during the eclipse and had no evidence of human sounds or celebrations. In our study, many locations with less than 99% obscuration with no human interference showed no detectable change in bird vocalization rates. It is likely that some of the effects on animal behavior reported elsewhere are at least in part due to an unusual increase in the number of enthusiastic humans outside.

This study shows the value of a geographically large, connected network of devices continuously detecting and classifying animal sounds to answer questions about their behavioral responses to large-scale environmental perturbations. This is the first of what we hope are many new discoveries that can be made from this growing spatiotemporal database; Haikubox data are available to scientists studying effects of wildfires, storms, and climate change.

There are limitations to this study due to the timing of the eclipse and use of remote sensing devices. The total solar eclipse occurred on the afternoon of April 8, 2024 and covered a swath of North America from Texas through Maine and Newfoundland. This means that the eclipse happened prior to full spring migration. Results may have been different if the eclipse occurred during the breeding season or during peak migration. Additionally, because acoustic data alone were collected from autonomously deployed Haikuboxes, the study did not collect or analyze any other bird behaviors. Birds may have demonstrated other behaviors that were not captured, including moving, roosting or preening in unusual ways. Without control over how individual Haikubox owners installed their device, there also could have been variability based on their placement choices.

Methods

Haikubox edge hardware

Haikubox contains an ESP32-S3-WROOM-1 dual core microcontroller module running at 240 MHz with either a Vesper MEMS microphone (model: VM3000; sensitivity − 26 dBFS) or an Infineon MEMS microphone (model IM73D122; sensitivity − 26dBFS). Vesper ceased production of their microphone in 2023. Power is provided by a 5 V DC power supply and the Haikubox is connected to the owner’s WiFi using Bluetooth. The owner decides where to install their Haikubox. When the Haikubox is connected to the router the latitude and longitude is obtained from the smartphone and stored in a cloud database. Any Haikubox owner who did not agree to share their data was excluded from the dataset.

Haikubox edge neural net

The Haikubox edge neural network was trained with 96 classes, including 83 bird species, environmental sounds (e.g., rain, engines, power tools, fireworks, and dog barking), and human-related sounds (e.g., vocalization, whistling, coughing). Since there are many more bird species, the edge net also includes a general bird class. TensorFlowLite is used to perform real-time inference using a custom-trained edge neural net quantized to int8 containing over 3 million parameters on 3-second, non-overlapping audio segments sampled at 24 kHz. The spectrogram for the edge net is generated with a 512-point FFT with a 280-point step size and Hanning window. When a ‘bird’ class is detected on the edge net, but it is not in the list of species to be reported directly, the audio clip is sent to a server to be analyzed with BirdNET.

BirdNET for Haikubox

The Haikubox system runs a custom-trained version of BirdNET in the cloud that includes over 1,000 bird species and other classes including human non-vocal, human vocal, human whistle, frog, and insect classes. If a species score was above 0.25 and the species was in the list of expected species, it was counted as a detection.

Detection algorithm

The algorithm for determining whether there is a bird detection is as follows:

  1. 1)

    The edge net detects a ‘bird’ class with a confidence greater than 90.5%.

  2. 2)

    If the top scoring species confidence is greater than 77.3% and the species is in a list of 30 common birds for that location, the audio clip is sent to the server and the detection of that species is stored in the cloud database. Only one audio clip is sent per minute for detection of common species, but each detection is counted.

  3. 3)

    If the top scoring species confidence is less than or equal to 77.35% or the top scoring species is not in the list of common birds for that location, then the audio clip is sent to a cloud server for inference using BirdNET for Haikubox. Note that if two contiguous clips satisfy this step, then only the first clip will be sent to the cloud server (i.e., there is always a gap of at least three seconds between inferences performed in the cloud).

Eclipse metadata

The time and amount of maximum obscuration was obtained for each Haikubox location using timeanddate.com.

Data set

During the eclipse we uploaded a three-second audio clip each minute, regardless of whether there was a detection by the edge net. These audio clips were used to verify that no human sounds were detected during the eclipse analysis period.

During the eclipse, 655 Haikuboxes experiencing at least a partial eclipse were connected to the internet and thus capable of identifying and sharing bird identification data without interruption. An initial screen was used to remove Haikuboxes that rebooted or where the edge net detected four or more human sounds above 0.25 confidence during the eclipse, yielding 472 Haikuboxes. These Haikuboxes were manually audited for human sounds and fireworks by investigating the spectrograms of the three second clip uploaded every minute. After this manual audit, 344 sites remained in the data set. We verified that none of these Haikuboxes were moved by their owners for the eclipse and were operating during the entire analysis period.

Statistical analysis

Permutation test for difference of means

To evaluate the effect of the eclipse on bird vocalization activity, we utilized a permutation test to compare the mean bird vocalization rate changes between total eclipse and partial eclipse locations. The permutation test is a non-parametric method that allows for comparison without relying on the assumptions of normality that are required by traditional parametric tests.

We first calculated the observed difference in mean bird call vocalization rate change between the total eclipse and partial eclipse locations. To assess the significance of this observed difference, we generated a distribution of differences under the null hypothesis that there is no difference in bird vocalization activity between the two groups. This was done by randomly reassigning the detected bird call rate changes to total eclipse and partial eclipse groups and recalculating the mean difference for each permutation. We repeated this process 10,000 times to build a robust distribution of mean differences under the null hypothesis.

Bootstrap confidence interval for difference of means

In addition to the permutation test, we employed a bootstrap method to estimate the confidence interval for the difference in mean bird vocalization rate change between total eclipse and partial eclipse locations. Bootstrap resampling is a non-parametric technique that allows us to estimate the sampling distribution of a statistic by repeatedly resampling the data with replacement.

To calculate the confidence interval, we performed bootstrap resampling by randomly selecting data with replacement from both the total eclipse and partial eclipse groups, maintaining the original sample sizes. For each of the 10,000 resamples, we calculated the difference in mean bird vocalization activity rates between the two groups. The resulting distribution of these differences allowed us to determine the 95% confidence interval.

Weather data

Weather data were downloaded from Open-Meteo.com which uses an algorithm to interpolate between weather stations to estimate weather conditions at a specific location50. The weather data for the three hours around the eclipse were averaged for each location with greater than 99% obscuration. Temperature was temperature at 2 m. Wind speed was wind speed at 10 m. Relative humidity was relative humidity at 2 m. Cloud cover was percent cloud cover. Longitude was not used in the statistical analysis as it would be tightly correlated with latitude along the path of the eclipse.