Intelligent Systems for Urban and Natural Environments: Advanced Deep Learning for Traffic Signal Control and Bioacoustics Analysis

Abstract

This dissertation explores the use of advanced deep learning frameworks to address complex problems in two main areas: intelligent transportation systems and bioacoustics. By leveraging modern architectures such as deep reinforcement learning (RL), attention mechanisms, and state-space models (SSMs), this research develops innovative solutions that outperform existing methods in efficiency, scalability, and performance. The first part aims to reduce urban traffic congestion and its environmental impact. It introduces two frameworks. FogLight is a cooperative traffic signal control system that employs fog computing and deep reinforcement learning to coordinate multiple intersections. Traffic signals are grouped, and the RL agent makes collective decisions. Simulations with both synthetic and real-world data demonstrate that FogLight decreases vehicle waiting times compared to current state-of-the-art distributed RL methods. Building on this, the dissertation presents GreenLight. This advanced traffic control framework uses a novel indexed additive self-attention mechanism to analyze detailed Vehicle-to-Infrastructure (V2I) data describing each vehicle’s real-time status. Powered by the attention mechanism, GreenLight significantly reduces traffic delays, fuel consumption, and emissions. The second part shifts focus from urban areas to natural environments, addressing challenges in bioacoustics analysis when labeled data is scarce. This research introduces BioMamba, an end-to-end audio large language model (LLM) based on Mamba, designed for bioacoustics tasks such as animal sound classification and detection. Pre-trained on a large collection of general audio clips through a two-phase self-supervised learning scheme and fine-tuned on 12 diverse bioacoustics benchmarks, BioMamba achieves performance comparable to state-of-the-art Transformer-based models while being more efficient in computation and memory. Overall, this dissertation highlights the versatility and strength of modern models in solving real-world issues. The proposed frameworks provide scalable, practical solutions for intelligent urban infrastructure management and ecosystem monitoring, emphasizing the transformative impact of deep learning across different applications.