Composite Augmentations for Semantic Segmentation in Aerial Images with Few Samples
Date
2025-04-09
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
ORCID
0000-0003-2162-4106
Type
Thesis
Degree Level
Masters
Abstract
Remote sensing has the potential to enable comprehensive population monitoring to inform wildlife and biodiversity conservation. However, annotated wildlife datasets in situ are often difficult, expensive, and time-consuming to procure. The accuracy of these remote sensing models is also paramount for their application on live data; composite datasets are often used to supplement real training data. However, the quality of these datasets is often compromised by perspective artifacts, inconsistencies, and other anomalies that arise from the composite nature of the data sources. This thesis proposes a computational and data-efficient method to synthesize composite images to supplement real-world data in data-sparse environments with few positive samples. We validate our pipeline through extensive experiments on benchmark remote sensing datasets, observing trends in results on datasets supplemented by these pipelines, as well as design recommendations for future composite augmentation research aimed at improving the fidelity and reliability of remotely sensed data.
Description
Keywords
Remote sensing, deep learning, composite images, synthetic data, machine learning
Citation
Degree
Master of Science (M.Sc.)
Department
Computer Science
Program
Computer Science