Efficient Multi-Class Segmentation in Remote Sensing Imagery Using a Customized Segment Anything Model

Abstract

This thesis explores the challenges of semantic segmentation in remote sensing imagery, focusing on overcoming issues such as class imbalance, small object detection, and diverse spatial patterns. Leveraging the Segment Anything Model (SAM), a custom multi-class segmentation model was developed, incorporating enhancements such as bounding boxes, grid points, and point labels to improve segmentation accuracy. Fine-tuning the model on benchmark datasets like Potsdam and Vaihingen revealed substantial improvements, particularly in handling underrepresented classes and intricate boundaries. The results demonstrate the model’s efficacy in remote sensing applications, contributing valuable insights to the field of high-resolution image segmentation.

Date of Award	16 Dec 2024
Original language	American English
Supervisor	Hasan Al Marzouqi (Supervisor)