Houegnigan, L. | bioRxiv Science

bioRxiv2024

Automatic parameter estimation and detection of Saimaa ringed seal knocking vocalizations

Passive acoustic monitoring (PAM) offers a non-invasive method for monitoring elusive pinnipeds, but manual analysis of large recording datasets limits its scalability. For the endangered Saimaa ringed seal ( Pusa saimensis ), PAM provides a rare opportunity to study breeding-season behavior beneath seasonal ice cover. We evaluated automated methods for detecting and characterizing the species' distinctive knocking vocalizations using recordings from Lake Saimaa. Annotated data ( n = 12 565 calls) were used to develop pulse repetition rate (PRR) estimation and call-detection systems. The best-performing PRR estimator matched manual measurements with a mean absolute error of 1.50 Hz, while a spectrogram-based convolutional neural network detected knocking calls with a mean F1-score of up to 99.28%. These results show close agreement with manual approaches, indicating that automated detection and characterization are achievable at a standard that could substantially reduce PAM analysis effort. This represents an important step toward scalable, long-term acoustic monitoring of this endangered seal.

ecology↗

bioRxiv2022

Wildlife and Marine Mammal Spatial Observatory: Observation and automated detection of Southern Right Whales in multispectral satellite imagery

The Wildlife and Marine Mammal Spatial Observatory is a joint research effort for the census of wildlife and particularly of marine mammals in satellite imagery. In that context, this paper illustrates the development of a high accuracy algorithm for the detection of right whales in sub-meter resolution multispectral satellite imagery with the constraint of a relatively small sample support of 580 southern right whale images. A significant space is devoted to exploratory data analysis to describe the statistical structure of right whale pixels and ocean surface pixels across multispectral bands. Observations of southern right whale in satellite imagery are divided into typical and atypical right whale forms and the first observations of right whale mother and calf pairs in satellite imagery are presented. Measurements of whales are furthermore automatically extracted from whale observations (major axis length, minor axis length, etc). A significant space is also devoted to statistical data exploration, a step frequently overlooked in machine learning solutions, that yet offers interesting insight into the structure of animal detection in satellite imagery. The extracted statistics can readily be used by researchers to develop detection solutions even with low sample support. The adopted solution for detection consists of feature extraction with a convolutional neural network followed by classification with a support vector machine. 20 different convolutional neural networks were tested for feature extraction. Biostatistics parameters (accuracy, sensitivity, specificity and precision) were measured for comparison. Most architectures generally achieved high performance with low false positive and false negative rates. 100% accuracy is achieved in the case of 2 convolutional neural networks, Nasnet Large and Inception V3, and only with a specific selection of multispectral bands. NB: This is a preprint that does not include satellite imagery due recent reviews

bioinformatics↗