Rats in urban areas: an acoustic analysis of the rat population within the Flevopark, Amsterdam
Bachelor thesis, Future Planet Project 2021 Major: Future Earth
Subject: Urban ecology
Supervisors: Dr. Ir. E.E. (Emiel) van Loon, dr. C. E. (Caitlin) Black, dr. R. P. J. (Renske) Hoondert Student: Mathijs Blom
Acknowledgements
I would hereby like to thank all those who have helped and supported me throughout the writing of this thesis. Firstly, I want to thank both dr. Caitlin Black and dr. Renske Hoondert for their feedback, support, and guidance. Furthermore, I want to express my gratitude to my peers Antonia van der Grinten, Amee van Boheemen, Ilja van Vuuren, Lune Walder and Sascha Rem. Their dedication to teamwork and communication greatly improved the quality of this thesis study.
Abstract
The brown rat (Rattus norvegicus) has thrived in European cities for hundreds of years, ever since their introduction from Asia around the 18th century. Even though they are one of the most common species living within cities, there still exists a knowledge gap on the role they fill within urban ecosystems. Acoustic analysis has been used to study animal species in the past and has proven to be of value when researching animal vocalization in relation to certain behavioral contexts. As a means to fill in the knowledge gap surrounding rats in urban ecosystems, and to gather information that could help construct a method for quantifying urban rat populations, an acoustic analysis was performed on the brown rat population within the Flevopark, Amsterdam. Two acoustic loggers (AudioMoths) were set up in the park along with two cameras to gather data on local rat vocalizations. Since rats communicate on ultrasonic frequencies, a deep learning program that is trained to filter ultrasonic vocalizations (USVs) from background noise called DeepSqueak was be used to analyse and process the raw data. The data showed that USVs were picked up within two frequency ranges typically associated with rat calls, namely 18-32 kHz and 32-96 kHz. However, since camera footage not only confirmed the presence of rats, but also wood mice, it is not possible to label all USVs as rat calls. Since this was the first implementation of DeepSqueak outside of a lab setting, these early results show promise for future applications to population studies within urban areas.
1. Introduction
The abundance of rats in urban green spaces is no new phenomenon and has been accelerated in the past by the rise of global trade. Rats were an originally an invasive species that were introduced to Europe from Asia around the 18th century (Strand & Lundkvist, 2019), and quickly grew in numbers. Rats and other vermin often travelled along with long sea voyages and carried a variety of pathogens to all corners of the globe. Past pandemics such as the bubonic plague were largely blamed on the abundance of rats, as they carried infected fleas which enabled the disease to spread (McEvedy, 1988). Currently, the sight of rats still brings about a sense of disgust and unease, unlike other vermin such as squirrels, which are just as likely to carry diseases or bacteria (Millins et al., 2015). The presence and sightings of rats in urban areas has even been proved to have a negative effect on the mental health of those who live near rat infested areas (Byers et al, 2019). Furthermore, there exists a knowledge gap on the ecological impacts of rat populations, as cities such as Amsterdam are struggling to quantify local populations. Given the fact that rats are known to reproduce at a rapid rate, up to five litters per year which can contain four to eight pups (Feng & Himsworth, 2014), their numbers are able to grow rapidly when not controlled. Additionally, the expansion of urban areas around the globe provides an increasing number of habitats and food for rat populations. At the moment there are no estimations on the total number of rats within the population in the municipality of Amsterdam. Instead, accounts of rat sightings or nuisances are being registered in an attempt to chart rat activity (GGD Amsterdam, 2020).
One possible method of studying rat populations in urban areas could be in the form of audio recorders, the usage of audio recorders to study the behavior of animal populations has proven to be of value in addition to other conventional techniques such as cameras or traps. For example, an acoustic analysis was done on the vocalizations of meerkat populations made when foraging for food (Townsend et al., 2011). Where cameras have always shown how meerkats warn their group mates when foraging for the presence or absence of any threat, this analysis showed that there are significant differences in how they signal for terrestrial or aerial predators (Townsend et al., 2011; Townsend et al., 2014). Additionally, an acoustic analysis has been performed on Mexican free-tailed bats, with the goal to determine whether or not there are differences in vocalizations that can be coupled to certain behavioral contexts (Bohn et al., 2008). The results showed how these bats use different vocalizations for scenarios such as infant isolation, mating calls, distress signals and some foraging calls (Bohn et al., 2008). These cases show how acoustic analysis can benefit in the better understanding and monitoring certain animal populations.
1.1. Aims and research question
The aim of this study is to contribute towards finding an efficient method suited to quantifying and studying the brown rat (Rattus norvegicus) population of Amsterdam, and to fill the knowledge gap that currently surrounds the dynamics of rodents in urban ecosystems (Banks & Smith, 2015). This study will consist of an experimental phase in which audio recorders will be used to study the vocalizations and presence of rats, along with other experimental methods such as cameras, chew cards and studies in surrounding flora and fauna. After the experimental phase, the collected data will be analyzed. The joint effort of these experiments and corresponding data are expected to provide a framework with which future research can attempt to quantify rat numbers in the city. Specifically, the results of these projects will serve as pilot data for an upcoming study by Dr. Caitlin Emily Black. Given the fact that rats are likely to carry pathogens that are harmful to humans, the results of this study are not only be beneficial for the municipality of Amsterdam, but also to both the GGD (Gemeentelijke gezondheidsdienst) and the RIVM (Rijksinstituut voor Volksgezondheid en Milieu). This first stage of experiments was conducted in the Flevopark, located in the east of Amsterdam. Since this study focused on studying the local rat population with audio recorders, the main aim was to confirm the presence of brown rats found in the Flevopark. The results of this experiment were inherently coupled with the results of the camera observations and the results of feeding lures, as analyzing the audio data to confirm rat presence can be supported if the audio data can be coupled to visual clues such as eating, mating, or fighting. The research question this study will attempt to answer is as follows: ‘To what extend can audio recorders aid in confirming the presence of wild brown rats within the Flevopark, Amsterdam’. Audio recorders have been used extensively to research rat behavior, mostly however in a lab context as opposed to studying rat populations. This study attempts to present audio recorders as a non-invasive method as an additional path to fill the knowledge gap surrounding rats in urban areas. Gaining information on which type of calls rats produce when undisturbed could prove useful when attempting to identify locations with an abundance of rats.
2. Materials and methods
2.1. Study area
The outcome of this study should be of use to study the rat populations of the municipality in Amsterdam. Therefore, the initial research area was chosen to be an urban green space within the municipality. Figures 1 and 2 show a schematic view of the study area, consisting of the Flevopark located in the east of Amsterdam. The area covers 0.359 km2 including various bodies of surface water, benches, and trash cans, making it a suitable habitat for the brown rat which thrive in and around urban environments and prefer to have habitats near the water where birds are being fed (Traweger et al., 2006).
Figure 1 Schematic view of the study area. (retrieved from: maps.google.com) Figure 2 Zoomed in view of the study area showing both study locations
(retrieved from: maps.google.com)
2.2. Data collection
In order to confirm rat presence within the park, two ‘AudioMoths’ were placed on the two locations shown in figure 2, the experiment consisted of eight study locations in total with two containing audio recorders. AudioMoths are acoustic loggers that are capable of registering not only audible, but also ultrasonic frequencies. The usage of these devices provides us with a non-invasive method for monitoring rat activity and its effectiveness is based in scientific research (Hill et al., 2018). While rats are known to make audible sounds, among themselves they almost exclusively communicate within the ultrasonic range (Brudzynski, 2009). The exact locations were coupled with the placement of two ‘Spypoint Force-Dark trail’ cameras and stations to measure the giving up density (GUD). This will help with confirming the presence and specific behavior or actions of the rats. One AudioMoth was be placed within 10 meters of a nearby trash can along with one camera and a station to measure the GUD. The other audio recorder was also placed within 10 meters of a trash can, but this time the GUD station is absent. The two cameras were encased in secured boxes, the Spypoint security box SB-200, and were attached to nearby trees or shrubbery with Python cables and will be suspended 10-30 cm above the ground. The cameras were put in ‘multi-shot mode’, which means that six photos per detection were taken, these settings were chosen as they are believed to provide the clearest image of the rats’ behavior.
Additionally, a ‘no-glow’ setting has been applied to all cameras which removes the flash effect from the taken photos, since the flash could influence the animals’ behavior. Since rat activity is highest between dusk and dawn, the AudioMoths will be set to only collect data between 6 P.M. and 8 A.M.
2.3. Data analysis
The AudioMoths store uncompressed ‘.wav’ files on a microSD card with a built-in memory of 32 GB. Since several species of rodents including the brown rat communicate on ultrasonic frequencies, ‘DeepSqueak,’ a learning-based system that is geared towards detecting and analyzing ultrasonic vocalizations (USVs; Coffey, 2019), was used to analyze the collected data. The DeepSqueak program is operated through MATLAB version R2021a. The program has been specifically trained to recognize rodent USVs found in audio files and helps with isolating vocalizations (Coffey et al., 2019). After loading the raw .wav files into DeepSqueak, the program converts the data into sonograms. As a result of a learning process undergone by the program, the sonograms are divided into sections that are passed on to further classification. The subsections of the sonogram are again passed through a denoising network created by DeepSqueak and is eventually labelled as an USV or as background noise.
The call statistics of the processed data are then exported to Microsoft Excel; these files contain the number of calls found by DeepSqueak per audio file with their corresponding statistics. The variables that will be considered are the number of accepted calls, call duration, principal frequency, minimal frequency, and maximal frequency. Additionally, a confidence score between 0-1 is generated after denoising which can be used as a threshold for further classification. Since adult rat vocalizations typically fall into two frequency ranges, being 18-32 kHz (’22 kHz-vocalizations’) and 32-96 kHz (’50 kHz-vocalizations’) (Portfors, 2007), two data subsets containing the calls within these ranges were created using the program R (see data repository). By analyzing whether the collected data identified calls within these frequency ranges, the presence of brown rats within the area could be confirmed using camera traps. Both AudioMoth locations are also equipped with their own Spypoint camera which could enable the acoustic analysis to be confirmed by additional footage of rats provided by these cameras. The observations made by all cameras were annotated manually using the program ‘Agouti’, these results were subsequently loaded into R. Each sequence recorded by the cameras was reviewed and any visible animals were identified where possible. The observation data was merged with the data containing the camera locations in order to determine what species were observed per location.
3. Results
As mentioned, the DeepSqueak-generated call statistics were exported to Excel and subsequently loaded into R for analysis (see data repository). Table 1 and 2 show the results of this analysis, the table 1 shows the results separated per location and table 2 shows the results of merging and subsetting the total dataset. The collection of data took place over a period of nine days, recording from 6 P.M. until 8 A.M. which resulted in a total of 126 hours’ worth of audio. The data shown in table 1 was retrieved by combining all call statistics per day per location resulting in 9 different datasets, while the data in table 2 is the result of combining and subsetting these daily datasets of all detected call files. Both tables show the number of accepted calls, the mean confidence score attributed by DeepSqueak, the mean call length, the mean principal frequency, and the low/high frequency. The principal frequency reported by DeepSqueak is defined as the median frequency of the USV contour (Lenell, 2021). Since the cameras at both locations did not manage to capture footage containing rats, the acoustic analysis cannot be supported by visual confirmation of rat presence. Both locations merely showed records of several bird species, humans, dogs, and lastly wood mice. However, cameras at other study locations within the park have managed to confirm rat presence.
Table 1. Descriptive statistics retrieved from the data associated with the audio collected at location three and four.
Date
Location Accepted
calls
Mean
score
Mean call
length (s)
Mean
principal
freq.
(kHz)
Low
freq.
(kHz)
High
freq.
(kHz)
2021-03-26 3 30 0.7156 0.02373 44.74 22.71 52.17 2021-04-07 3 23 0.7124 0.04165 37.06 17.86 95.26 2021-04-08 3 214 0.7642 0.03524 38.72 16.91 61.92 2021-04-14 3 862 0.8317 0.03998 44.17 14.44 60.53 2021-04-15 3 24 0.8203 0.0402 40.40 18.85 58.76 2021-04-16 3 312 0.7046 0.03372 38.39 17.90 74.46 2021-03-26 4 962 0.8665 0.02691 45.17 17.35 62.73 2021-04-07 4 2353 0.5828 0.03986349 27.78 16.74 92.28 2021-04-08 4 1633 0.6618 0.04276693 29.18 15.42 95.40Table 2. Descriptive statistics retrieved from the merging of all created datasets, in addition to those of the two subsets made. The merged dataset contains the call statistics of both locations over all days of data collection.
Dataset/subset Accepted
calls
Mean
score
Mean call
length
Mean
principal
freq.
(kHz)
Low freq.
(kHz)
High freq.
(kHz)
Merged dataset 6413 0.6929 0.03815 33.99 14.44 95.40 Subset 18-32 kHz 3242 0.5929 0.04091 27.44 18.01 31.99 Subset 32-96 kHz 2475 0.8333 0.03108 45.07 32.00 95.40The camera traps throughout the park managed to capture approximately 30 animal species (see data repository), including the Rattus norvegicus (brown rat) and the Apodemus sylvaticus (wood mouse). Since the footage captured by these camera traps can be linked to audio collected around the same time, making this connection could prove useful in confirming presence of certain species with a higher degree of certainty. However, since the cameras at location three and four did not capture any rats, this connection could only be made with wood mice sightings. The observation data from all camera locations shows that a wood mouse was sighted at location three on April 8th at 19:40:29, since the raw audio data from that moment has a unique id which in this case is ‘606F5C10’ it can be linked to its detection file created by DeepSqueak. Figure 3A shows camera footage of location three at the time of a wood mouse sighting, with figure 3B showing a call fragment found by DeepSqueak which was labeled as a ‘USV’.
Figure 3A, 3B Figure 3A shows the recording of a wood mouse in the bottom left corner, whereas figure 3B shows a call fragment found by DeepSqueak at the same data, at approximately the same time (2021-04-08 19:40). The time on figure 3A does not correspond with the observation time since the internal clock of the camera did not adjust for daylight savings.
4. Discussion
Since this study is the first recorded use of DeepSqueak to analyze data that has not been recorded in a lab setting, some technical challenges in calibrating the program and processing the data needed to be overcome. As expected, the sonograms from the collected data often contained significantly more background noise than the provided example recordings of mouse and rat vocalizations. Due to this fact, and the large number of call detections that needed to be categorized, extra trust had to be put in the ‘Post Hoc denoising’ feature provided by DeepSqueak. This is a trained network which separates the call fragments and labels them either as ‘USV’ or ‘background noise’, providing another level of accuracy in addition to the confidence score generated in the primary detection. For calls to be considered for analysis, they both had to be labelled as ‘USV’ and have a confidence score of above 0.50 since this is the standard threshold used by DeepSqueak. When rejecting certain data based on a threshold, the value of said threshold should always be well considered. Future research could explore the possibility of raising the threshold to 0.60, 0,70 or even 0.95, to determine whether this yields more accurate results. Table 2 shows how there is a significant difference in confidence scores between the two frequency range subsets. One possible explanation for this could be the fact that lower range frequency calls are more likely to be polluted by background noise, even though DeepSqueak attempts to filter this noise it is still possible that there are some inaccuracies. Figures 4 and 5 show an example of a lack of accuracy when working with a program that has not been optimized for outside usage; two sonograms which DeepSqueak has both accepted and attributed a roughly equal score of ~0.6 but differ greatly visually. Yet another example of inaccuracy is shown in figures 6 and 7, where two sonograms who are visually similar have gotten drastically different scores (0.74 and 0.37 respectively) where figure 6 shows the accepted sonogram and 7 the one which was rejected.
Figure 4A, 4B Two sonograms generated by DeepSqueak which were both retrieved from data collected 2021-04-08 at location four. Figure 4A shows a sonogram which does not resemble any form of vocalization that DeepSqueak labeled as USV, whereas 4B shows an accepted call that more closely resembles a rodent vocalization. Both fragments were attributed similar confidence scores of ~0.6.
Figure 5A, 6B Two sonograms generated by DeepSqueak which were both retrieved from data collected 2021-04-08 at location four. While figure 5A and 5B appear similar regarding sonogram shape, vastly different confidence scores were attributed to both fragments. The call shown in 5A received a score of 0.74 and was thus accepted, whereas the call shown in 5B was rejected with a score of 0.37)
As mentioned, the data shown in the results section were determined using this 0.50 threshold and show that there are a total of 6413 call fragments that DeepSqueak has labeled ‘USV’. Of these 6413, 5717 calls fall into either the 22 kHz- or 50 kHz vocalization categories that were previously described. However, attributing all calls to brown rats is not possible since mice are also known to produce USV’s within ranges of 30-110 kHz (Portfors, 2007). Since both the presence of wood mice and brown rats within the study area has been confirmed, further distinction between sonograms needs to be made which is hindered by the fact that these sonograms are often less clear due to high amounts of background noise. DeepSqueak provides four networks used for call recognition: one for general USV recognition, one for short rat calls, one for long 22 kHz rat calls, and finally one for mouse calls. Given the amount of audio data and the fact that this is the first-time trained networks are used to detect USV’s outside a lab setting, the general USV network was used for data analysis. Making this distinction between rodent species using additional monitoring methods such as additional cameras, chew cards or traps allows for acoustic analysis to not only aid in confirming rat presence but also allude to specific behaviors.
Literature suggests that the two categories of rat vocalization are associated with either a negative or a positive response; 22 kHz vocalizations were typically produced in lab settings where rats were confronted with inescapable pain or other negative stimuli (Litvin, 2007), whereas 50 kHz vocalizations are linked to positive response such as socializing, being ‘tickled’ by lab personal or during sexual
Future implications of this research could prove useful in finding new methods to study the impact of rats in urban areas. The existence of the knowledge gap that currently still exists regarding the role that rats play in urban ecosystems is partly the result of a lack of non-invasive methods to study their behavior. As mentioned, analyzing UVSs can indicate specific behaviors since certain frequency ranges are associated with either positive or negative stimuli. Additionally, rat pups are known to emit vocalizations in the range of 40 kHz when separated from their mothers (Portfors, 2001), in lab settings these findings enable scientists to determine whether the pups or adult rats are in stressful situations which might influence study outcomes. When studying rats in urban areas, one possible application of these findings is a new method of population charting; when catching one rat for example, microphones could be placed at the catching location to determine if there are pups around that are ‘calling for their mother’ by emitting 40 kHz-vocalizations. Additionally, audio recorders could provide a method of quantifying a trend in abundance by analyzing different urban areas and comparing the number of calls found per area. However, since there is still a significant lack of research done on the analysis of rodent vocalizations outside of lab settings this study could provide the first steppingstone for future projects.
Finally, a problem that arose during the data collection phase meant that no data could be collected at location four, since the entire setup had been destroyed. The SD cards of both the microphone and camera could not be recovered hence the gap in data from the 8th of April onward.
5. Conclusion
The data retrieved from the acoustic analysis shows promise regarding future application in confirming rat presence and in monitoring their populations. USV’s have been detected within the frequency ranges that are typically associated with rat calls. However, the calls cannot be attributed to rats alone since there were no rats found on the camera footage and there were wood mice found at both locations. Since the aim of this research has been to determine to what degree an acoustic analysis could aid in confirming rat presence, the results would have to indicate that an acoustic analysis would have added value to existing methods such as cameras, traps, and chew cards. The results of this study on its own were not able concretely confirm rat presence, although it is likely that there were in fact rat calls picked up by the audio recorders. With more experimentation and further exploration of the possibilities within DeepSqueak, acoustic analysis could provide a non-invasive method of not only confirming rat presence but also provide an indication on behavior, or abundance.
6. Data management
The raw data collected by the audio recorders is stored in the form of .wav files and will is located on a local hard drive and will be made available upon request. The call statistics retrieved via DeepSqueak, the camera observation data, and the script used to analyze all data were uploaded to the public online data repository figshare (www.figshare.com). This data can be found by searching for ‘Mathijs Blom’, ‘Rodent call statistics generated by using DeepSqueak’, ‘R scripts used to analyze rodent call statistics generated by 'DeepSqueak', or ‘Annotated camera observations of wildlife within the Flevopark, Amsterdam’.
7. References
GGD Amsterdam. (2020, 27 juli). Bruine rat en zwarte rat. GGD Amsterdam. https://www.ggd.amsterdam.nl/dierplagen/bruine-rat-zwarte/
Banks, P. B., & Smith, H. M. (2015). The ecological impacts of commensal species: black rats, Rattus rattus, at the urban–bushland interface. Wildlife Research, 42(2), 86.
https://doi.org/10.1071/wr15048
Bohn, K. M., Schmidt-French, B., Ma, S. T., & Pollak, G. D. (2008). Syllable acoustics, temporal patterns, and call composition vary with behavioral context in Mexican free-tailed bats. The Journal of the Acoustical Society of America, 124(3), 1838–1848. https://doi.org/10.1121/1.2953314
Brudzynski, S. M. (2009). Communication of Adult Rats by Ultrasonic Vocalization: Biological, Sociobiological, and Neuroscience Approaches. ILAR Journal, 50(1), 43–50.
Https://doi.org/10.1093/ilar.50.1.43
Byers, K. A., Cox, S. M., Lam, R., & Himsworth, C. G. (2019). “They’re always there”: resident experiences of living with rats in a disadvantaged urban neighbourhood. BMC Public Health, 19(1), 1-13.
Coffey, K. R., Marx, R. G., & Neumaier, J. F. (2019). DeepSqueak: a deep learning-based system for detection and analysis of ultrasonic vocalizations. Neuropsychopharmacology, 44(5), 859–868. https://doi.org/10.1038/s41386-018-0303-6
Feng, A. Y., & Himsworth, C. G. (2014). The secret life of the city rat: a review of the ecology of urban Norway and black rats (Rattus norvegicus and Rattus rattus). Urban Ecosystems, 17(1), 149-162.
Hill, A. P., Prince, P., Piña Covarrubias, E., Doncaster, C. P., Snaddon, J. L., & Rogers, A. (2018). AudioMoth: Evaluation of a smart open acoustic device for monitoring biodiversity and the environment. Methods in Ecology and Evolution, 9(5), 1199–1211. https://doi.org/10.1111/2041-210x.12955
Lenell, C., & Johnson, A. M. (2021). The effects of the estrous cycle, menopause, and recording condition on female rat ultrasonic vocalizations. Physiology & Behavior, 229, 113248. https://doi.org/10.1016/j.physbeh.2020.113248
Litvin, Y., Blanchard, D. C., & Blanchard, R. J. (2007). Rat 22kHz ultrasonic vocalizations as alarm cries. Behavioral Brain Research, 182(2), 166–172. https://doi.org/10.1016/j.bbr.2006.11.038
Massei, G., Lyon, A. J., & Cowan, D. P. (2002). Conditioned Taste Aversion Can Reduce Egg Predation by Rats. The Journal of Wildlife Management, 66(4), 1134. https://doi.org/10.2307/3802945
McEvedy, C. (1988). The bubonic plague. Scientific American, 258(2), 118-123.
Millins, C., Magierecka, A., Gilbert, L., Edoff, A., Brereton, A., Kilbride, E., Denwood, M., Birtles, R., & Biek, R. (2015). An Invasive Mammal (the Gray Squirrel, Sciurus carolinensis) Commonly Hosts Diverse and Atypical Genotypes of the Zoonotic Pathogen Borrelia burgdorferi Sensu Lato. Applied and Environmental Microbiology, 81(13), 4236–4245. https://doi.org/10.1128/aem.00109-15
Mulder, C. P. H., Grant-Hoffman, M. N., Towns, D. R., Bellingham, P. J., Wardle, D. A., Durrett, M. S., Fukami, T., & Bonner, K. I. (2008). Direct and indirect effects of rats: does rat eradication restore ecosystem functioning of New Zealand seabird islands? Biological Invasions, 11(7), 1671–1688. https://doi.org/10.1007/s10530-008-9396-x
Parsons, M. H., Banks, P. B., Deutsch, M. A., Corrigan, R. F., & Munshi-South, J. (2017). Trends in urban rat ecology: a framework to define the prevailing knowledge gaps and incentives for academia, pest management professionals (PMPs) and public health agencies to participate. Journal of Urban Ecology, 3(1), jux005.
Portfors, C. V. (2007). Types and functions of ultrasonic vocalizations in laboratory rats and mice. Journal of the American Association for Laboratory Animal Science, 46(1), 28-34.
Strand, T. M., & Lundkvist, Å. (2019). Rat-borne diseases at the horizon. A systematic review on infectious agents carried by rats in Europe 1995–2016. Infection Ecology & Epidemiology, 9(1), 1553461.
Townsend, S. W., Zöttl, M., & Manser, M. B. (2011). All clear? Meerkats attend to contextual information in close calls to coordinate vigilance. Behavioral Ecology and Sociobiology, 65(10), 1927–1934. https://doi.org/10.1007/s00265-011-1202-6
