Acoustic Signal Detection of search-phase echolocation bat calls with CNN

Varnit Shree , Affan Ahmed , Poornima Kulkarni

Acoustic Signal Detection, Search-phase Echolocation, Bat Calls, Convolutional Neural Network (CNN), Spectrogram, Time-frequency Representation, Denoising, Sliding Window, Precision, Recall, ROC Curves, PR Curves, Deep Learning, Bioindicators, Biodiversity, Endangered Species, Passive Acoustic Monitoring

Bats are one of the most diverse species on earth. Due to deteriorating environmental conditions bat fauna across the globe are endangered. One of the important characteristics of Bats is bioindication. Not only do they provide highly beneficial services to nature such as pollination and pest control, they provide useful insights on the changes occurring to the ecosystem as a result of anthropogenic change. Therefore, bat tracking helps in conservation of endangered bat species and also to measure trends in biodiversity. The existing systems till now have focused only on the task of species classification neglecting the task of first localizing bat calls. This project develops a convolutional neural network based pipeline for automatically detecting searchphase calls produced by echolocating bats in noisy, real world recordings. The proposed CNN architecture is made up of four convolutional layers. The ReLU activation function is used, and both max and average pooling have been incorporated after each convolutional layer. Finally, the output of the convolutional layers are fed into a fully connected layer, and then to a softmax layer which returns the probability of the audio files containing bat call. This CNN is applied in a Sliding window manner to accommodate audio files of various durations. Since CNN network can only be used on images, audio files are first converted into spectrogram or time-frequency representation and then denoised. The results of our proposed model were compared with other existing models on different evaluation metrics like precision, recall, ROC curves and PR curves. The proposed model performed better than the existing systems on three different acoustic datasets. Around 500 more bat calls were detected across all 3 datasets compared to that of the existing systems, with significant increase in recall of the proposed model, as high as 11% compared to existing systems. Though the detection process was effective, the main drawback of this work was the lack of graphical support it provided. Providing a GUI would make this process user friendly. A future enhancement to improve detection performance would be to create an ensemble CNN model trained with both STFT and MFCC representation separately

"Acoustic Signal Detection of search-phase echolocation bat calls with CNN", IJRTI - International Journal for Research Trends and Innovation (www.IJRTI.org), ISSN:2456-3315, Vol.9, Issue 6, page no.284 - 290, June-2024, Available :https://ijrti.org/papers/IJRTI2406040.pdf

Volume 9 Issue 6, June-2024

Pages : 284 - 290

Paper Reg. ID: IJRTI_190014

Published Paper Id: IJRTI2406040

Downloads: 205,624

Research Area: Engineering

Country: Bengaluru, Karnataka, India

Published Paper PDF: https://ijrti.org/papers/IJRTI2406040

Published Paper URL: https://ijrti.org/viewpaperforall?paper=IJRTI2406040

ISSN: 2456-3315 | IMPACT FACTOR: 10.57 Calculated By Google Scholar | ESTD YEAR: 2016

An International Scholarly Open Access Journal, Peer-Reviewed, Refereed Journal Impact Factor 10.57 Calculate by Google Scholar and Semantic Scholar | AI-Powered Research Tool, Multidisciplinary, Monthly, Multilanguage Journal Indexing in All Major Database & Metadata, Citation Generator

UGC CARE Approved — Transparent Peer-reviewed journal aligned with UGC’s 2025 suggestive parameters.

Publisher: IJRTI (JW Publication)