A Tiny CNN Architecture for Identifying Bat Species from Echolocation Calls

Abstract

Effective monitoring of bat populations will contribute towards the United Nations' SGD 15 which is tied to maintaining biodiversity and SGD 3 which is about maintaining good health and well-being. Bat species are particularly sensitive to anthropogenic pressures and monitoring bat populations trends can serve as a good indicator of an ecosystem's health. Bats have also been linked to at least 60 strains of viruses (including Covid-19) that can infect humans. Monitoring bats can help contribute towards achieving both SGD goals. However, monitoring bats is a difficult and resource-consuming task. This paper investigates how monitoring can be enhanced by automatically identifying bats using audio data from their echolocation calls. Such a system will ease assessing their populations status, trends and habitats. A Convolutional Neural Network (CNN) was developed to identify eight different bat species based on their vocalizations. Alternative CNN models using Short-Term Fourier Transforms (STFTs), Mel-spectrograms Filter banks (MSFB), and Mel Frequency Cepstral Coefficients (MFCC) were developed. The CNN models were optimized using Hyperband. The best model used MSFB features and had only 220K parameters (0.892 MB) and can easily be embedded into a small handheld device. Using 10-fold testing, the best CNN model had an average Accuracy of 97.51% and Average F1-score of 0.9578.