Using Machine Learning for Automated Identification of Nocturnal Wildlife Species

Thesis Abstract

The accurate identification of nocturnal wildlife species remains a significant challenge in conservation biology and biodiversity monitoring due to their elusive behavior, limited visibility during night hours, and the substantial labor and expertise required for manual identification. This study aims to develop and evaluate a machine learning-based framework for automated identification of nocturnal wildlife species utilizing audio-visual sensor data collected in natural habitats. The specific objectives include designing a comprehensive dataset comprising audio recordings and infrared video footage, training and optimizing convolutional neural networks (CNNs) and support vector machines (SVMs) for species classification, and assessing the model's accuracy and robustness across varying environmental conditions. A quantitative research design underpins the study, employing a cross-sectional approach to analyze data gathered during nocturnal surveys conducted across multiple forest reserves. The targeted population encompasses representative populations of key nocturnal species, including small mammals, reptiles, and amphibians, with a total sample size of 2,500 audio-visual recordings obtained over a six-month period. Within this dataset, stratified random sampling was employed to ensure balanced representation across species, habitat types, and temporal variations. Data collection instruments included synchronized audio recorders, infrared cameras, and environmental sensors to capture multiple modalities of wildlife activity as well as contextual environmental variables. Data preprocessing involved noise filtering, augmentation techniques such as spectral transformation and image enhancement, and feature extraction protocols utilizing Mel-frequency cepstral coefficients (MFCCs) for audio and convolutional features for visual data. For machine learning model development, the study applied deep learning architectures—specifically several CNN configurations—including ResNet, VGG, and custom lightweight models, alongside traditional classifiers like SVMs for comparative analysis. Model evaluation was conducted using metrics such as accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve (AUC-ROC), with k-fold cross-validation (k=10) employed to ensure robustness. Additionally, gradient-weighted class activation mapping (Grad-CAM) techniques facilitated interpretability of CNN outputs, aligning model decision-making with salient features. Anticipated findings suggest that the integrated multimodal approach will significantly outperform traditional manual identification methods, achieving classification accuracies exceeding 85% across multiple species. The models are expected to demonstrate high robustness under diverse environmental conditions, such as varying illumination and background noise, through data augmentation and model regularization. The results are projected to show a considerable reduction in time and cost associated with nocturnal wildlife monitoring, with potential application in conservation management, ecological research, and policy formulation. This research contributes to the existing knowledge by advancing the application of machine learning techniques—particularly deep learning frameworks—in automating species identification from nocturnal sensor data, thus filling critical gaps in remote wildlife monitoring. It extends theoretical understanding by integrating ecological models with technological innovation, particularly through the lens of the Social-ecological production system (SEPS) and Object Detection Theory. The study also offers practical insights into designing scalable, context-aware systems for biodiversity conservation efforts. In conclusion, the study emphasizes the potential of AI-driven solutions to transform nocturnal wildlife monitoring, advocating for wider adoption of automated identification frameworks to facilitate timely and accurate ecological assessments. Recommendations include further refinement of models through expanded datasets encompassing additional species and habitats, integration with real-time monitoring platforms, and policy development to promote technological adoption in conservation strategies. Future research directions should explore the deployment of such systems in diverse ecological contexts and evaluate their long-term impacts on conservation outcomes.