Date of Award

12-2025

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

Electrical Engineering

Committee Chair/Advisor

Dr. Fatemeh Afghah

Committee Member

Dr. Yongqiang Wang

Committee Member

Dr. Yongkai Wu

Abstract

Wildfires have increased in frequency & severity over the past several decades, demanding faster, safer, & more continuous situational awareness than traditional wildfire management methods that rely on satellite, manned aircraft, & ground-sensor pipelines can reliably provide. This thesis designs, develops, evaluates & presents a fully autonomous UAV agent trained in a high-fidelity digital twin environment for wildfire detection & monitoring. The FIRETWIN digital twin environment used in this thesis incorporates the Cosys-AirSim plugin in Unreal Engine 5.3 and renders 3D terrain, vegetation, and physics-based fire & smoke modeling, enabling reproducible training & evaluation in a realistic environment. The proposed model architecture combines Proximal Policy Optimization (PPO) modeled reinforcement learning with a frozen pretrained vision–language model to deliver semantic-based reward shaping & directional guidance from dual onboard cameras, using a segmented top-down view for immediate directional navigation guidance & an angled horizon view for longer-range scene cues.

This final dual-camera, VLM-guided agent is compared against a selection of five ablated model variants as benchmarks, including a baseline PPO agent without VLM-guided reward shaping, and evaluated on metrics including the total evaluation episode reward, the percentage of time the UAV detects wildfire in its field-of-view, and the time-to-detection. These six models are evaluated under five different scenarios to evaluate their ability to detect & localize wildfire in the immediate field-of-view, conduct short-range & long-range search & exploration, and respond to adverse environmental conditions such as high wind & low lighting. Overall, the simulation results presented in this thesis validate the core objective of this research, demonstrating that semantic reward shaping from VLM guidance has the potential to significantly improve autonomous, AI-enabled wildfire search & monitoring, developing a robust & efficient trained model with the potential for future real-world implementation and field deployment.

Available for download on Thursday, December 31, 2026

Share

COinS