Following ecological quick scans, bat surveys often need to be carried out. Currently, this is a time-consuming activity and also very costly. In addition, there are too few ecologists available for all surveys to meet the demand. This is not surprising: it is night work, which not everyone likes.
For Aquila Ecology, this was reason to start working on an AI model to identify bats. In collaboration with the NDFF, we created an AI model that can identify bat individuals by species with 91% accuracy based on sound. These sounds must be recorded with special recording equipment that can record high sound frequencies. Then this can be determined by the developed AI model. But how does developing such an AI model actually work technically?
First, we need a large dataset of bat sounds whose determination has been established with certainty. For some bat species this is easy, for example there are many recordings of Common Dwarf bat. However, there are also species that make sounds that are difficult to distinguish from other species. These sounds are often recorded before or after animals are captured and identified by appearance. Of these, far fewer sounds are available.
Machine learning actually needs as wide a variation of sounds as possible. For instance, the computer learns to distinguish exactly what is typical about the call of a particular bat species, and what happens to be a variation or background noise. To artificially increase the amount of data, we use a trick called data augmentation in jargon. Here, we mix the sounds the AI is trained on with different background sounds to still add variation. We had to collect these background sounds first, because they do need to be recorded in high frequency.
All the data are then repeatedly 'looked at' by an algorithm. This algorithm adjusts itself slightly each time based on the data. It looks for patterns in the data in a complex way. Each time the algorithm 'looks' at the data, it gets a little better at recognising bats. Aquila Ecology's software indicates this while training the model:
The next step is to make this technique useful for analysing large data sets of, say, several weeks. For this, the software must not only be able to distinguish bat species, but also know when there is no bat on the recording. We do this by recording a lot of environmental noise and teaching the software to distinguish those.
