Remote sensing applications of machine learning processes: satellite imagery road extraction using few shot segmentation

• Main Author: Ong, Grace Hui Lee
• Other Authors: Long Cheng
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision
• Online Access: https://hdl.handle.net/10356/168247
• Institution: Nanyang Technological University

Road extraction from aerial images is a fundamental task in the field of remote sensing. Much of the deep learning models for road extraction rely on convolutional neural networks (CNNs) and their derivative architectures. CNNs are able to capture higher-level representations in the images’ raw pixel data content and perform well on various image tasks. However, many machine learning as well as CNN-based models tend to perform poorly on data from unseen classes. Additionally, these models require a large amount of data, time and annotations, especially in segmentation tasks, where dense labels are hard to annotate. Few-shot segmentation approach is recently gaining popularity due to its potential ability to learn from limited data, annotations and manual parameter tuning, by recognizing object masks from unseen classes with only one or a few annotated support examples as supervision. However, there is a dearth of studies associated with applying few-shot segmentation onto road extraction tasks. Hence, in this project, two main models: i) U-Net, a popular model for binary segmentation, and ii) PFENet, a state-of-the-art few-shot segmentation model, are trained to learn a classifier to segment roads from non-roads, and tested on a set of data from unseen classes. To facilitate the model learning performance evaluation, a web application is developed to allow users to query and evaluate the models. The results strongly suggest that the PFENet model is feasible for road extraction tasks. For future studies, hyperparameter tuning and other loss functions such as shape-aware loss and combo loss should be explored to further improve the efficacy of the model learning architecture.