iBatsID (Europe) is a classification tool which uses ensembles of artificial neural networks (eANN's) to classify time-expanded recordings of bat echolocation calls from 34 European bat species (see Walters et al. 2012).
This classification tool has been trained to identify calls from the following species: Barbastella barbastellus, Eptesicus bottae, E. nilssonii, E. serotinus, Hypsugo savii, Miniopterus schreibersii, Myotis alcathoe, M. bechsteinii, M. blythii, M. brandtii, M. capaccinii, M. dasycneme, M. daubentonii, M. emarginatus, M. myotis, M. mystacinus, M. punicus, Nyctalus lasiopterus, N. leisleri, N. noctula, Pipistrellus kuhlii, P. nathusii, P. pipistrellus, P. pygmaeus, Plecotus auritus, P. austriacus, Rhinolophus blasii, R. euryale, R. ferrumequinum, R. hipposideros, R. mehelyi, Tadarida teniotis, Vespertilio murinus.
Given the intrinsic uncertainty in the eANN classifications for some species, we do not recommend accepting species-level classification for all species. We recommend to exclude species-level classifications for species in the lowest quartile of probability of correct classification (less than 70.65%). The following species should therefore only be classified to subgroups:  M. myotis/M. blythii/M. punicus;  M. bechsteinii/M. brandtii/ M. daubentonii/ M. mystacinus;  M. emarginatus/M. alcathoe).
For all classifications, applying a threshold of 80% to the eANN outputs increases median correct classification rate across the remaining species/ groups to 95%. Further increasing the threshold does not significantly increase correct classification rate. We therefore recommend applying a threshold of 80% to the eANN outputs, so that calls with eANN outputs below 80% are left unclassified at that level of the hierarchy. Classification should then be assigned by the last stage of the hierarchy for which the output is greater than 80%.
See  Walters et al. 2012 for details and methods used in constructing the classification tool.

How to use iBatsID (Europe)  

Getting started 

To open iBatsID click on the link:     iBatsID
This should open a page which looks like this:

Inputting Data

Calls should be measured using SonoBat software (see www.sonobat.com).

SonoBat outputs a text file (see example below), which can then either be copy/pasted directly into iBats ID in the exact format output from SonoBat (use Ctrl+A to select all, Ctrl+C to copy and Ctrl+V to paste), or the file can be loaded by clicking 'Load Data From File' and then navigating to the file on your computer.
NB. Please ensure decimal points are displayed as full stops (.) rather than commas (,). If commas are used rather than full stops, IbatsID will not be able to correctly identify calls. Commas can be changed to full stops in the text file using the 'find and replace' function in a text editor. 
The file data should appear on the screen. You can now click 'Classify Calls' to run the classification.
Example SonoBat output file is available for download below.

Classifying calls

The default method of classifying calls, used in Walters et al. submitted, is to run calls through one stage of the network hierarchy at a time. The output can then be assessed manually at each stage, and you can decide whether to apply thresholds to the classification probabilities.

To run the network in this way, select ‘Use selected network’ then in the  dropdown menu, select ‘CallTypeGroups’ first.

You can select ‘Download result as file’ to generate a .csv file output, or leave this unselected to see results appear on the screen.
Click ‘CLASSIFY CALLS' to run the classification.  

Interpreting Results

The results will either download as a .csv file, or appear on the screen, depending on whether the ‘Download result as file’ box is checked.
If results are displayed on the screen, these can be downloaded to .csv after by clicking the ‘Download as CSV’ link at the bottom of the page.
The results page has a number of columns, in comma delimited format. This is easier to read when opened in excel.

The first three columns give the call id number, the filename, and the start time in file so the call can be traced back to the original recording.  

The next two give the species or group classification that has been assigned to the call, and the associated classification probability.

The rest of the columns give the probability of classification to each of the remaining categories for that stage of the hierarchy.

For Example:

This call is classified as ‘Group 5’ with a classification probability of 0.980240541. The probability this call is in ‘Group 2’ is 3.5 x10-05, the probability it is in ‘Group 4’ is 1.01 x10-04 and so on.

At this stage the user can decide which network stage the call should be run through next. This will usually be the network associated with the classification given in the ‘Classification’ column in the results panel (i.e. the output with the highest classification probability). In the example above this is Group 5. Here, the user can decide if the classification probability has reached a particular threshold level, and therefore whether or not to classify the call further.

Alternative method

An alternative method of classifying calls is to run calls through every stage of the network at once. The eANN output probabilities for each stage of the network are multiplied, so calls are given a probability of classification as each of the 34 possible species. This propagates the uncertainty in classification at each stage of the hierarchy, allowing an assessment of the confidence of classification.  However this does not allow the user to manually set a threshold for each stage of the hierarchy.
To run the network in this way select ‘Use Full Network’
 The results will show a number of columns, as before, with the first 3 being the call ID, filename and start time in file, and the next 2 being the classification assigned to the call and the associatied classification probability. The rest of the columns show the classification probabilities for each of the 34 possible species.

Neural Network Configuration Files

The Proj Files.zip folder, downloadable below, contains text files showing the parameter settings used in training each neural network, as well as the classification rates of each individual network.


17 Feb 2012, 08:30
Proj Files.zip
20 Apr 2012, 03:45