Utilizing Neural Networks to Resolve Individual Bats and Improve Automated Counts

Accurate population counts are essential for understanding the status of species and for researchers studying various phenomena including monitoring the relationship between environmental stresses and the spread of disease within populations. Both small roosts and large colonies of bats provide challenges when attempting to determine an accurate population count. Recently, there have been a number of new video analysis software applications, that are available on the internet, which can be used to provide population counts. When software-based counts are compared with manual counts, the software provides counts that are substantially less labor intensive, determined substantially more quickly, and have the potential to be more accurate. This short paper discusses the use of neural networks to determine the number of bats that there are in a region when multiple bats may overlap. The work discussed in this manuscript demonstrates that the counts of multiple overlapping bats can be improved using trained neural networks. This is a critical improvement for providing accurate counts in high density videos. This manuscript contains the biological motivations, and a brief overview of how artificial intelligence is being implemented. The results discussed compare the accuracy values of neural networks for a few case studies including cross-comparisons of data trained on different video types and for different animals which can have accuracy values above 90 % for comparable video types. Finally, the generation and use of synthetic images, to increase the amount of data in a training set, is also discussed, which resulted in a trained neural network that produced an accuracy value of 80% on 12 unbiased categories.