North Atlantic right whale acoustic signal processing: Part II. improved decision architecture for auto-detection using multi-classifier combination methodology

Autonomous signal detection of the North Atlantic right whale (NRW), Eubalaena glacialis, is becoming an important factor in monitoring and conservation for this highly endangered species. Both online and offline systems exist to help study and protect animals within this population. In both cases auto-detection of species-specific calls plays a vital role in localizing individual animal by searching time-frequency passive acoustic data. This research presents an experimental system, referred to as the NRW-CRITIC, for automatic detection of the NRW contact call. In general, the CRITIC uses a combinatorial classifier approach to integrate a series of existing machine learning algorithms; each designed specifically for NRW contact call identification. The proposed configuration consists of several recognition methods running in parallel; these include linear discriminant analysis, artificial neural network (NET) and classification regression tree (CART). This paper presents the details for the NRW-CRITIC and discusses the approach used to combine multiple independent decisions into a single result. A side-by-side performance comparison, between the CRITIC and a well-known method, the feature vector testing (FVT), is summarized. Performance metrics are evaluated based on a large database of acoustic recordings consisting of over 58,000 NRW contact calls from various locations, including two critical habitats, Great South Channel and Cape Cod Bay. Results indicate the FVT algorithm yields a 74.7% detection probability with an error rate of 4.35%. In comparison the CRITIC, operating at similar information level yields a 78.02% detection probability with a 3.25% error rate, exceeding the performance of the FVT. Performance was also measured using data from a multi-channel acoustic array located in Massachusetts Bay. A side-by-side comparison of array presence is discussed for two separate days. Results show that with the FVT and CRITIC operating at 0% error for array pres- - ence, the FVT method had 18,769 and 24,469 false positives for the Massachusetts Bay datasets respectively. With the same 0% error condition the CRITIC provided successful detection with significantly lower number of false positive rates: 1,072 and 2,324 calls, respectively. Future extensions of this experimental work are also discussed.