Cross-Biome Biodiversity Assessment and Anomaly Detection Using AI-Enhanced Acoustic Monitoring

Abstract

This study proposes a novel AI-powered eco-monitoring framework that integrates acoustic ecology, deep learning, and low-cost IoT devices to enable scalable, real-time biodiversity assessment and ecological anomaly detection across diverse environments. The primary objective is to automate species classification and environmental monitoring using passive audio data captured by solar-powered IoT sensors, thereby reducing reliance on manual ecological surveys. The framework comprises four modules: acoustic data acquisition, dual-representation preprocessing Short-Time Fourier Transform (STFT) and Mel-Frequency Cepstral Coefficients (MFCCs), species classification using CNN and CNN-LSTM models, and anomaly detection via autoencoders and one-class SVM. Field validation and multi-dataset testing were conducted across 250+ species from temperate forests, wetlands, and urban areas. The CNN-LSTM model achieved the highest performance with 93.7% accuracy, 93.0% precision, and a 92.5% F1-score, while anomaly detection reached 89.7% precision with an AUC of 0.94, effectively identifying irregularities such as invasive calls, mechanical noise, and species absence. A forest case study demonstrated the system’s ability to detect circadian acoustic patterns (e.g., dawn chorus of sparrows, nocturnal owl calls), and real-world disturbances with 91% expert validation agreement. The novelty of this work lies in its hybrid AI architecture with real-time unsupervised anomaly detection, cross-biome generalization capability, and deployment readiness on low-power edge devices like Raspberry Pi and Jetson Nano. Inference times as low as 18 ms per sample and bandwidth usage under 3 MB/hour make it feasible for continuous, remote deployment. The framework offers a robust and adaptable solution for conservation efforts, environmental policy, and climate resilience initiatives. Future directions include integrating multimodal data sources and transformer-based continual learning for broader ecological impact. These findings position the system as a scalable and intelligent tool for next-generation, AI-driven environmental monitoring.