Reccurent Neural Network for the Detection of Odontocetes
This research project explores the ability of a Recurrent Neural Network to detect Odontocetecalls. With the current degree of climate change impacting the planet, being able to better track and locate highly mobile endangered species is imperative. An example of such a species is the Sperm Whale, w...
| Main Author: | |
|---|---|
| Format: | Text |
| Language: | unknown |
| Published: |
University of New Hampshire Scholars' Repository
2022
|
| Subjects: | |
| Online Access: | https://scholars.unh.edu/thesis/1646 https://scholars.unh.edu/context/thesis/article/2686/viewcontent/Foster_unh_0141N_11495.pdf |
| Summary: | This research project explores the ability of a Recurrent Neural Network to detect Odontocetecalls. With the current degree of climate change impacting the planet, being able to better track and locate highly mobile endangered species is imperative. An example of such a species is the Sperm Whale, which is listed as endangered under the United States Endangered Species Act, that migrate annually and have a range that extends through all the world’s oceans. Current methods of detection require a great deal of manual interaction and skill on behalf of the identifier. Automating this process would enable easier tracking and aid conservation and study efforts. The system of concern is the use of a Bilateral Long Short-Term Memory Recurrent Neural Network (B-LSTM) to detect odontocete calls in a data set provided by the 2018 Workshop on Detection, Classification, Localization and Density Estimation of Marine Mammals using Passive Acoustics. This data set included 92,050 Odontocete calls that were utilized in this experiment. This data set was randomly partitioned into a training set and a test set for two neural networks, one using B-LSTM and the other using a basic Feedforward Neural Network (FNN). Analysis was conducted using Precision-Recall and False Positive Rate-Recall curves to compare the performance of the two networks. While the B-LSTM model was able to successfully detect odontocete calls, it did not perform better when compared to the FNN model in all cases. The B-LSTM network performed better than the FNN for the recall range of [0.253, 0.817], achieving a peak precision of 0.904. When there were significantly more false negatives than true positives (low recall) as well as when there are significantly less false negatives than true positives (high recall), the B-LSTM network achieved less precision than the FNN. |
|---|