[/heading][divider line_type=”No Line” custom_height=”4″][toggles accordion=”true”][toggle title=”Q: Has the SonoBat “SonoBatch“ Auto-classification performance been “peer reviewed“ for accuracy? If so, what is the false-positive rate, especially for T&E species?” color=”Default”]
A: Acoustic species discrimination is still a relatively new “science” and it must be understood that any auto-classifier provides guidance and inference to species, but by virtue of the call variation from all species, the process of classification does produce false positives on species ID. The species ID process essentially amounts to a probabilistic process because the plasticity and resulting call repertoires of species results in so many areas of overlapping data space.
Although some species do show a 100% correct rate for some situations, do note that rate describes the classifier’s ability to recognize ideal reference (i.e., good quality) files of the type on which the classifier was built, and a finite data set. You should expect actual field data to present unknown call types and sounds outside of the known library, and any such signals can potentially result in misclassifications.
Only a smaller set of species-discriminating call varieties can provide confident species ID, e.g., the short, flat, above 26 kHz call types of Lano, and these are the type of calls that you would use to make confident decisions about presence . However many other call types from species fall into overlapping data space, e.g., the longer curved call types of Epfu and Lano, and SonoBat will output what can only be considered a probabilistic decision, the correctness of which will depend upon quality, the distance of the bat from the microphone, and probably the bat itself. For some species like Myso and Mylu back east we have yet to even recognize a species-discriminating call variant and the entire call repertoire apparently overlaps, even though the classifier shows some statistical difference.
The take home message is that bats exhibit considerable plasticity in their calls and SonoBat and any other classification system must be considered fallible and serve merely as a guide and inference to classification.
A: Yes, each regional classifier employs a slightly different decision process based on the species suite included in the classifier. When you can exclude ambiguous species of concern then you gain much greater performance in classifying the species that you expect to be present in a geographic area. Therefore, applying this classifier to a geographic region outside the range of these species may result in some misclassifications of the out of range species.
A: There is currently is no system available that can provide a complete certainty and process everything for you, but SonoBat comes closest to that ideal than any other system. Bats have considerable plasticity in the calls they produce and within each species’ repertoire there can be some data space that overlaps in characteristics with parts form the repertoire of another species. Definitive species recognition relies upon the unique subset of each species’ repertoire. This also means that no one can identify each and every recording; for reasons of call plasticity and call quality. Noise and distance from the microphone greatly affects signal quality and classification in the same way that distance and backlight can make it difficult to identify birds. Unfortunately that’s the nature of the beast. It’s not so different than most other wildlife monitoring methodologies in that it requires some level of expertise.
The current SonoBat classifiers provide batch processing and automated classification with idealized rates of performance listed in the classification notes accompanying the program. Even with the high quality data on which the US classifiers were built (11,000 call samples), species classification rates range from just below 90% to 100%, with acceptable classifications (those that come out with an acceptable level of confidence, e.g., in discriminating data space) range from 27% to 97%. So statistically, even with a 98% correct rate (and that’s ideal- for good data), the overlapping characteristics of calls will bring up some misclassifications for any substantial data set, particularly if recorded under less than ideal conditions.
But, that is the state of the art, and most agencies that require this sort of survey work understand that, but do expect best possible practice. I hope this helps to understand this endeavor. The only reasonable solution to get data analyzed thoroughly at this time would involve working with someone that already has the expertise to oversee the analysis.
According to the 2014 Indiana Bat Summer Survey Guidelines developed by the USF&WS: “At a minimum, for each site/night a program considered Indiana presence likely, review all files from that site/night. Qualitative analysis must also include a comparison of the results of each acoustic ID program by site and night (including: number of call files flagged as probable Indiana bats by each tool used; an evaluation of other species identified by the acoustic ID program; individual file level agreements and disagreements on Indiana bats between programs; and a qualitative analysis of ALL probable Indiana bat call sequences to further evaluate that the correct ID has been recommended by the program used).” And, also according to the USF&WS: “Individuals qualified to conduct qualitative analysis of acoustic bat calls typically have experience: (1) gathering known calls. This provides a valuable resource in understanding how bat calls change and the variation present in them; (2) identifying bat calls recorded in numerous habitat types; (3) familiarity with the species likely to be encountered within the project area; and (4) individuals must have multiple years of experience and must have stayed current with qualitative ID skills. A resume (or similar documentation) must be submitted along with final acoustic survey reports for anyone making final qualitative identifications.”
A: Automated call processing can do much to reduce the workload of eliminating noise and non-bat files unsuitable for analysis, and point users to files with good potential for confident identification, and suggest those identifications. However, with so much unexpected variability when interpreting the SonoBat outputs, responsible users should manually vet any files that come up as an unusual, unexpected, or a result of one or only a few species at a site, and manually check these files to see if they contain characteristics of a confident type of call sequence. The classifier operates by comparing the parameterized data from calls and how they fit into the parameter space of known species. Although SonoBat uses a number of redundant checks and attempts to recognize spurious signals and noise, some data still gets through that falls into some known data space and results in a misclassification. And of course that happens more often with files recorded under less than ideal conditions.
A: SonoBat rejects calls deemed unreliable based on several factors: (1) A quantified quality measure based on the total points of the sonogram above a threshold value, and uses this value internally to assist in the call trending analysis of strong and weak call signals. Generally, the best and most reliable calls to use for species recognition will have a quality value above 0.90 and calls with values below 0.90 will become increasingly less reliable with lower and lower quality ratings. This is a synthesized measure found to work well for discarding unreliable signals. It is a basically a tweak of a combining signal to noise ratio and dynamic range measurements. It has a user defined threshold in the Prefs panel.
Also: (2) SonoBat will reject calls based on poor call trends, i.e., too much small scale bouncing around of the trend line, indicating noisy, unclear, or poor calls that produce unreliable results. E.g. a jumping trend can happen at the end of a call from overlapping echoes and this can result in the trend missing the downward toe that may possibly be essential for discriminating an big brown bat from a low frequency Myotis species or a little brown bat from a red bat.
And finally: (3) SonoBat rejects calls after a species determination if they do not meet minimum characteristics found to provide reliable results for a given species. One example that can result in a rejected call is too little bandwidth, because that can indicate a call fragment. Under quiet ambient conditions calls can come out with a high quality rating but the bat can still be too far from the microphone to provide full tonal content essential for a confident ID and this will eliminate the chance of misclassification with the fragment of another species. E.g., the fragment from the body of the call from a Myev falls into the data space of short calls of Epfu. With the full bandwidth of a Myev call and details like resolving its toe, the discrimination becomes more confident.
A: You can never get more than +/- 1 or 2 kHz precision with bat calls because of Doppler shift effects from their movement relative to the microphone, and we can’t ever control how closely any given bat approaches our microphone. The bats also vary their calls considerably. They are quite flexible, in fact. And if two conspecifics fly together, then they will shift frequency to accommodate each other to parse out bandwidth, and if the other one is beyond the range of your microphone then you only get an odd sequence. Precise analytics do not apply in quantitative species ID, rather overall ensembles of features and trends must be used.
A: A lot of recordings fail SonoBat’s call quality settings because users set up their equipment near the ground or in areas with lots of objects around the microphones that generate echoes resulting in lots of poor quality signals. This scenario can be improved upon to make cleaner recordings by using externally mounted (cabled) microphones that are deployed at least 20-feet (6 meters) off the ground. The higher above the ground, the better. Consider that the microphone picks up good recordings from bats up to about 20-feet, and few bats fly within a 5 feet of the ground. Then you might as well get the mic up 25-40-feet to pick up the bats 20-feet below it, and 20-feet above it. That way you’ll cover more volume of flight space and minimize ground echoes.
A: Ultrasound recordings made with microphones on Wildlife Acoustics SM2BAT and EM3 detectors have subtle differences in low level noise compared with those of other detectors. A lot of the SM2 calls fail at the last stage in the SonoBatch decision algorithm because of how noise affects call quality and the inherently limited bandwidth of SM2 (or EM3) microphone response. You can increase your throughput of identified calls if you adjust your SonoBat preference for acceptable call quality downward from the default setting of 0.80 to 0.70 to compensate and enable the consideration of more calls. But, use caution to avoid too many low quality calls to pass through as acceptable as that may compromise correct classification with SonoBat.
And, to increase the recognition of sequences for tallying bat passes, reduce the acceptable quality to tally passes from the default setting of 0.20 to 0.10 or perhaps as low as 0.05. The ambient noise and acoustic conditions will affect how low you can adjust this setting and avoid errant tallies of non-bat signals. If uncertain of how this setting will perform with your data, run some test batches and inspect the results to guide your selection. Most users will probably want to err on the side of caution and use a higher value that avoids tallying any non-bat signals.
Finally much time and energy and effort has gone into creating compensators to adjust for the MEMS microphones used by older Wildlife Acoustics detectors. Unfortunately, not enough benefit was realized from these efforts to make them efficient or effective for large-scale acoustic analysis. More benefits can be realized by proper detector and microphone placement and knowing when to reject low-quality calls from analysis. A better choice is the new SM3 detector avail be from Wildlife Acoustics.
A: Choose a classifier that contains as many of your regional species as possible, understanding that those not present will need to be manually inspected. But, you can still use the classifier to help to automate the identification process in a couple of different ways. (1) You could run batches in “parameterize” mode and sort the call parameter spreadsheet output to find recordings having characteristic frequencies and call durations consistent with species in your area not covered by the classifier, or (2) You could choose a classifier that has a congener (i.e., related species) for un-included bat(s) in your area. For example, running the western U.S. classifier in southern Florida will identify Molossus as Eumops.
In the end, it is not recommended to use a classifier with species that you don’t expect to have, as you will get a certain number of misclassifications and false positives. Run a classifier without spurious species to avoid this headache, unless you want to keep species of interest in there to alert you for potential records of those species that you can manually confirm.
A: There are actually no substantial regional differences between geographically separated individuals of the same species that would affect the classifier. For example, the classifier built on western big and little browns does fine to classify eastern North American ones, and vice versa. Early reports of regional variation in the literature were not “apples to apples” comparisons, e.g., bats in clutter in one region compared to bats in the open in another region, and then analysts concluded the differences were due to geographical locale. Upon discerning the full call repertoires of species in each region, the geographical variations essentially evaporate into the intra-specific repertoire variation.
The expectation for regional variation follows from birds, who select their songs as identifiers, and have no constraint on complexity. Bats select their calls to optimize information acquisition to suit their tasks, and those tasks remain similar within a species. The physics of sound do not change across the continent, and as such the selective forces on call morphology operate to maintain consistent species call repertoires across regions.
Ambiguous species pairs and other difficult recordings
[/heading][divider line_type=”No Line” custom_height=”4″][toggles accordion=”true”][toggle title=”Q: SonoBat keeps giving me Cora/Coto classifications on calls that clearly do not belong to this species upon manual inspection. Is there something wrong with my classifier?” color=”Default”]
A: Coto and Cora are indistinguishable calls, so a classification of one can mean the other. So what you have depends upon what you expect and where you record them. Moreover, SonoBat produces Cora/Coto results somewhat liberally with the intent to serve as an alert for a potential detection, but with the expectation that these results require manual confirmation, and the relative rarity of Cora/Coto doesn’t present much of a burden to do that. More often than not, other signal sources will generate Cora/Coto output as this species pair has a very simple, common denominator, call types, and many species of bats which are echolocating just out of range of the best volume of detection of a microphone will produce call fragments that mimic a harmonic-less Cora/Coto sequence.
Cora/Coto produces “feature-thin” calls that just run from about 40 to 20 kHz, and out of range bats of numerous other species can easily leave fragments that leave pieces like that. E.g., If you have good confirmed Myth sequences at your site, you could readily have some out of range Myth sequences that leave fragments that come up as Cora/Coto. Also, Tabr can produce calls that have characteristics like that during feeding buzzes. So to vet Cora/Coto calls, check that you have a complete sequence of search phase calls, hopefully with harmonics typical of and readily recorded for Cora/Coto (compare calls classified as Cora/Coto with the reference views). If you suspect Tabr feeding buzzes you should also see a transition of call types in the sequence.
Also, everything goes out the window if you record near a roost as you do not get routine search phase calls and you get a lot of social vocalizations of great variety. The rule of thumb is to ONLY analyze obvious “search phase” calls and eliminate recordings with “approach phase” or “feeding buzz” call pulses from analysis or when manually vetting decisions out-putted by a SonoBatch.
Finally, do note that the caveat screen and Classification Notes in the SonoBat “help” sections recommend manual vetting of any Cora/Coto result. In other words, consider the SonoBat output of results for this species to automatically point you to which recordings to check for confirmation.
A: Only a smaller set of species-discriminating call varieties can provide confident species ID between this species pair. E.g., the short, flat, above 26 kHz call types of Lano, and these are the type of calls that you would use to make confident decisions about presence. Full spectrum recordings of the shorter duration calls for this species pair are essential, especially with fully-formed call pulses that include harmonics. Epfu high-frequency maximum values under these conditions will exceed 55kHz. Conversely those of Lano will almost always be below 50kHz. However many other call types from species fall into overlapping data space, e.g., the longer curved call types of Epfu and Lano, and SonoBat will output what can only be considered a probabilistic decision, the correctness of which will depend upon quality, the distance of the bat from the microphone, and probably the bat itself.
A: Free-tails and hoary bats both make very loud calls that travel far, get filtered through a lot of airspace, and lose details. When all you get is a piece from the strongest part of the call. i.e., the fragment, this fragment can fall into Tabr data space and come out classified as that. Such out of range classifications require manual confirmation (and that’s acoustically out of range and geographically out of range). If such recordings do not have Tabr species-discriminating characteristics (e.g., an upturn into the call or a long downward turn out of the call) then do not accept the classification. Separating Laci-Tabr has about a 5% error rate even for very good data, so when you toss in not so good data you can expect more misclassifications.
A: Yellow bats would mostly likely fall into the data space that the Ozark-nGA classifier knows as Lasionycteris noctivagans, although they will also have some overlap with Eptesicus fuscus. But, unlike Lano and Epfu, yellow bats will have the typical curvilinear, reverse J-shaped call type that fluctuates erratically in Fc from pulse to pulse in a sequence, so Epfu and Lano calls would have to be hand-vetted if yellow bats are suspected, and any ambiguous (e.g., No-ID results) with Epfu or Lano pulses identified somewhere in the sequence should also be viewed.
A: Unfortunately, Lasiurus seminolus and L. borealis appear acoustically indistinguishable. Refer to a range map and consider a result for Labo as either where these two species’ ranges overlap and Lase if you do not expect Labo where the recordings were made.
A: The SonoBat classification process operates probabilistically, and even with a 98% correct rate (and that’s ideal- for good data), the overlapping characteristics of calls will bring up some misclassifications for any substantial data set, particularly if recorded under less than ideal conditions. As recommended in the software caveats, you should vet any spurious or suspicious classifications for this reason.
Labo presents trouble because out-of-range Myotis species calls, mostly Mylu, can leave simple curved fragments that fall into Labo data space. The farther out of range a Mylu the less myotis features are left, such as the downward ending tail. As a rule out west it is best to personally check everything that comes out Labl. This is easier in the West than in the east as Labo are generally more rare.
A: Acoustically, Myme will be the same as Myci. In fact, there was no Myme until recently when the molecular systematists declared that part of the Myci range would now become separate realms of Myci and Myle. A couple of decades a go they did that to break all of North American Myle into Myci (western) and Myle only in eastern North America. Acoustically, eastern Myle and western Myci are essentially acoustically ambiguous, too. So, refer to a range map and if you recorded a call where the map shows Myme instead of Myci (or Myle!), consider any SonoBat Myci result to have the genes that the systematists call a Myme.
A: Disambiguation is more tricky with Nyhu as it just has a simple curved call shape that overlaps so much with the simple curves that Labo can make. Nyhu will tend to have a steeper (i.e., more vertical) start if you have full, good samples, and look for the Nyhu characteristic of regularly alternating Fc up-down-up-down, different than Labo that will vary Fc up and down more randomly.
A: Unfortunately, Pisu’s make a simple curvy call that can be mimicked by call fragments from other species like Labo and even Mylu. In general, Pisu has more consistency in Fc across a sequence than Labo, but both can be confused with slightly out-of-range Mylu where the typical Myotis species downward trending, low intensity “toe” is attenuated by distance or otherwise obscured by echo or other over-riding noise. In cases of poor quality calls, ambiguous classifications will result and not every recording can be confidently identified to species, nor can SonoBatch classification results on poor-quality recordings be confidently reported.
A: Unexpected results from the SonoBat classifier should always be manually reviewed. Sometimes Myotis californicus, especially longer duration calls, do not contain much of a diagnostic downward trending toe, especially if the bat is echolocating at quite a distance from the microphone, and this typically low-intensity portion of the call is being attenuated by distance. You can simulate this distance-induced attenuation by moving the blue “threshold” slider upward, which is located on the right hand side of the SonoBat window. View the “Myca_WawonaCA-01” call pulse from your SonoBat reference call collection for an example of an ambiguous Myca/Pihe call type.
A: The Indiana bat, Myotis sodalis (Myso), and the little brown bat, M. lucifugus (Mylu), present substantial overlap in their echolocation call characteristics that render only a small portion of their repertoires with a tendency toward discriminating characteristics. Although SonoBat may report a result indicating a greater likelihood of one species over the other, e.g., 0.85 Myso versus 0.15 Mylu, such a result only indicates the relative distances from the centroid of the known multivariate data space for each species. Because these species have their centroids buried in the multivariate data clouds of the other species, they never clearly separate, and either species could just have well vocalized a call producing those results, despite lying closer to the mean values of one over the other.
To prevent outputting null species identification results, the SonoBat classifier uses this rubric: when a species decision for either of theses species does not exceed the threshold discriminant probability setting (DP, SonoBat uses 0.90 as the default setting), and if the second potential species comes out as the opposite of this pair, and their combined discriminant probability score meets or exceeds the threshold setting, then SonoBat will output this result using the ambiguous designation “LuSo.” This will indicate the call or sequence probably came from one of these two species, but presented call characteristics within overlapping data space that prevented disambiguation. However, with most Myotis species, longer duration calls provide more robust and consistent data that enhances discrimination performance. As a first rule of thumb, have little confidence in any results from calls less than 5.5 or 6 msec and have increasing confidence for calls longer than that. And of course, have more confidence in sequence results (based on longer calls) than from individual calls. The sequence decisions that SonoBat generates integrate the combined information from all calls in a sequence.
A: Nyhu and Tabr are both difficult to classify as even the very best calls can have simple, largely feature-less shapes that make it easy for other species to mimic, i.e., fall into that data space, especially with poorer quality or out of range calls. That is, calls missing features distill can down to having the characteristics of those species. We see this too with the overlap of Laci and Tabr. It’s a headache. In practice, don’t trust the acoustic data to confidently think you have any given species unless you have manually vetted the sequence to see if it has discriminating characteristics. For example, definitive roll into the call and roll down out of the call for Tabr (i.e., “chaise lounge” shaped call pulses). If just feature-thin flat calls they are ambiguous. For data like that SonoBat can only really point out the files that you need to look at and then see if they past muster.
A: Tabr causes a lot of ambiguous classification results because it can act most like a mockingbird among our bats, i.e., it has a tremendous variety of call types that can overlap with calls of other species. Manual inspection reveals non-discriminating call types for Tabr. With shorter-duration calls that are diagnostic for Tabr, look for calls that have “upswing into call and downswing out of call” making the pulse appear “chaise lounge shaped.” This is a (mostly) reliable distinguishing characteristic. This is qualified with “mostly” because you can often find any bat eventually doing something odd. The more pronounced the upswing/downswing the more definitive the Tabr. Do note that the beginning and ending details (where you get the upswing/downswing) have lower amplitude than other parts of the call so you first lose some of those details as bats get further from the detector. That’s how you often pick up ambiguous call pulses that are confused with Epfu and/or Lano. Realize that as you get farther from the detector, the volume of the airspace you cover increases with the cube of the distance! Therefore you are much more likely to record bats with some missing low amplitude details than those that fly close to your detector and give you their full information content. With longer-duration calls that are diagnostic for Tabr, look for nearly flat, quasi-constant-frequency (QCF) or constant-frequency (CF) calls with “carrot shaped” oscillograms because Tabr (unlike potentially ambiguous Laci) are “power forward,” turning on quickly at the beginning of the call, then slowly decreasing in intensity.
A: You can identify “social calls” or “directives” in a bat pass as typically longer duration, lower frequency, and atypically shaped call pulses. These calls generally do not repeat at regular intervals as do echolocation calls. They are typically interspersed among echolocation calls at random intervals. Social calls or directives are only now being studied in attempts to understand their significance and potential species- and/or individual-specific qualities. They are currently not used in SonoBat classification algorithms, but we are collecting evidence of potential social calls, so feel free to submit any *.WAV recordings you suspect of containing them to us if you wish. Directions for submitting *.WAV files to us for archiving, comment, or identification can be found in the My Account section of this site.