Deep Learning Bioacoustic Detection of Threatened Amphibian Populations

Abstract

Introduction: Conservation and animal management are informed by repeated assessments of threatened amphibian populations and environmental factors affecting their behavior. Classical observation is effective, but sporadic and hard to conduct in wide or remote territories. Materials and Methods: Artificial intelligence (AI)-assisted estimation of animals' activity based on passive acoustics recordings and meteorological parameters including rainfall, temperature, humidity, water depth, and vegetation is presented here. Experimental data set consisted of 18,600 ten-second audio fragments recorded at 12 wetland sites during 14 weeks. Recorded data was filtered, synchronized and split into training, validation, and testing datasets. Results and Discussion: Proposed methodology yielded 96.0% accuracy, 95.5% precision, 95.1% recall and 95.3% F1 score. Significant correlations between animals' behavior and environment were found for humidity, water depth, temperature, vegetation and anthropogenic noises. The method enhanced identification of behavioral pattern in different environmental conditions and provided interpretable warnings for ecological analysis. Conclusion: Presented results prove that artificial intelligence can enhance non-invasive observations if model outputs are related to environmental context and expert knowledge