AI-based synthesis of aerial imagery

Abstract

This thesis investigates how modern text-to-image di!usion models can be leveraged to
synthesise aerial wildlife imagery for training and evaluating the BAMBI animal detection system. Unmanned aerial vehicles (UAVs) equipped with RGB and thermal sensors
are increasingly used in wildlife monitoring and anti-poaching, but collecting and annotating su"cient real data is costly and time-consuming. Existing synthetic approaches
largely rely on game engines and struggle to reproduce the complexity of real Earthobservation imagery. The work addresses three research questions: (1) Which di!usion
or generative AI models are suitable for generating synthetic data to test the BAMBI
classifiers? (2) Which model architectures are most appropriate for this wildlife application? (3) How can models be adapted to generate convincing aerial RGB, heatmap
and thermal imagery? To answer these questions, two state-of-the-art generative models, Stable Di!usion and FLUX, are fine-tuned using Low-Rank Adaptation (LoRA) on
BAMBI-related aerial datasets. Multiple datasets with short and long prompts, as well
as three image styles (RGB, heatmap, thermal), are constructed. Model performance
is evaluated quantitatively using Kernel Inception Distance (KID), Fréchet Inception
Distance (FID) and CLIP-Score, and qualitatively via a YOLO-based Bambi classifier
applied to generated thermal images. The experiments show that LoRA fine-tuning consistently improves both models, but Stable Di!usion systematically outperforms FLUX
across all datasets and styles in terms of KID, FID and CLIP-Score. RGB imagery
is the easiest modality, while thermal images remain most challenging; nevertheless,
Stable Di!usion achieves near-zero KID on larger datasets and produces scenes that
frequently trigger plausible animal detections. Overall, the results demonstrate that diffusion models particularly LoRA-tuned Stable Di!usion o!er a promising pathway to
generate statistically and semantically meaningful synthetic aerial wildlife imagery, reducing reliance on purely manual data collection and providing a scalable complement
for training and testing UAV-based conservation systems.

Date of Award	2025
Original language	English (American)
Supervisor	David Christian Schedl (Supervisor)