UAV trajectory optimisation for pre- and post-bushfire disaster assessment using metaheuristic and deep learning algorithms

Abstract

The MAs presented in this thesis are nature-inspired and swarm intelligence–based algorithms that provide the best transportation cost with the least computational time for pre-disaster assessment. Moreover, for the DL-based algorithms, the proposed you only look once (YOLO) algorithm analyses the real-time images and differentiates between fire and no-fire regions, providing the greatest accuracy in a shortened time for post-disaster assessment. In a disaster scenario like a bushfire, the telecommunication infrastructure is vulnerable and considerable concerns arise for communication of UAVs to the ground station. Therefore, to provide continuous and robust communication, advanced IoT connectivity can be incorporated to provide a dedicated, smart, and automated system for bushfire disaster management. This thesis focuses on the UAV trajectory optimisation for pre- and post-bushfire disaster assessment using MAs and DL algorithms while considering a. the dynamics of computing environment, b. selection of associated parameters, c. analysis of exploration and exploitation groups, d. comparative analysis, e. flexible deployment of UAVs, f. robust connectivity, g. bushfire detection, and h. shortened computation time and improved transportation costs. The main contributions of this thesis are as follows: 1. a nature-inspired metaheuristic-based technique is compared for UAV trajectory optimization using advanced IoT connectivity in smart cities. The proposed SFOA method outperforms other algorithms based on the chosen parameters, saving up to 24.5% in transportation costs and 13.3% in computational time. An autonomous UAV trajectory optimization for pre-bushfire disaster assessment using state-of-the-art swarm intelligence metaheuristic algorithms, where the DG- BCO algorithm outperforms others using exploration and exploitation phases and achieves a 25.2% improvement in transportation costs and a 20% reduction in computational time. four distinct environmental scenarios – general environment, condensed obstacle environment, maze environment, dynamic environment – are explored and com- pared to demonstrate the resilience and versatility of our proposed model. a lightweight DL-based YOLO network, incorporating an improved bottleneck CSP module and PAN layers, achieves a higher accuracy of 97.4%, the least false positive rate of 1.258%, and a minimum NMS of 3ms during post-bushfire detection.

Date of Award	2023
Original language	English
Awarding Institution	Western Sydney University
Supervisor	Khoa Le (Supervisor)