Automated detection of dolphin whistles with convolutional networks and transfer learning

Abstract

Effective conservation of maritime environments and wildlife management of endangered species require the implementation of efficient, accurate and scalable solutions for environmental monitoring. Ecoacoustics offers the advantages of non-invasive, long-duration sampling of environmental sounds and has the potential to become the reference tool for biodiversity surveying. However, the analysis and interpretation of acoustic data is a time-consuming process that often requires a great amount of human supervision. This issue might be tackled by exploiting modern techniques for automatic audio signal analysis, which have recently achieved impressive performance thanks to the advances in deep learning research. In this paper we show that convolutional neural networks can indeed significantly outperform traditional automatic methods in a challenging detection task: identification of dolphin whistles from underwater audio recordings. The proposed system can detect signals even in the presence of ambient noise, at the same time consistently reducing the likelihood of producing false positives and false negatives. Our results further support the adoption of artificial intelligence technology to improve the automatic monitoring of marine ecosystems.

Building Ensemble of Resnet for Dolphin Whistle Detection

“…The protocol used in our experiments mirrored that proposed in [28]. However, we have used the validation set described in section 2.3.2 to learn which data augmentation methods to apply and the weights of the weighted sum rule.…”

“…The dolphin whistle dataset developed by Korkmaz et al [28] is a welldeveloped and relatively large set of patterns containing 108,317 spectrograms, of which 49,807 are tagged as noise and 58,510 as dolphin whistles. The test set contains 6,869 spectrograms.…”

“…As mentioned in the introduction, the training set [28] contained all the spectrograms obtained from audio files recorded between June 24th and June 30th, while the test set was composed of the spectrograms of audio files recorded between July 13th and July 15th, a three-day window. The rationale given by the authors for dividing the training and test sets in this manner was mainly to test the generalizability of models using completely disparate sets of recordings, as this would better assess detection accuracy amidst varying sea conditions.…”

“…The original training and test sets in [28], as described in section 2.1.2, were used in this study. Unlike the original authors, however, we extracted a validation set from the training set using all the spectrograms related to the three-day recording period of June 24th to June 26th.…”

“…The goal of this work is to continue exploring data augmentation techniques for the task of dolphin vocalization detection. Towards this end, we use the benchmark dolphin whistles dataset developed by Korkmaz et al [28] but apply data augmentation on the original test set of spectrograms to enlarge it rather than on the training set. The training set contains all spectrograms obtained from audio files recorded between June 24th and June 30th, while the test set is composed of the spectrograms of audio files recorded between July 13th and July 15th, a three-day window.…”