Date of Award

Spring 2026

Abstract

The application of AI/ML to shallow bathymetric mapping with airborne lidar remains limited by a lack of high-quality labeled datasets, extreme spatial variability, and noise introduced by environmental conditions and hardware limitations. These issues reduce the generalizability of traditional statistical approaches for extracting bathymetric points from lidar point clouds, with additional challenges for supervised neural networks arising due to the amount and quality of labeled data required to properly capture broad seafloor characteristics. Recent advances in point cloud machine learning methods have the potential to improve conventional bathymetric processing workflows. This work examines modern self-supervised learning (SSL) methods for bathymetric lidar point cloud processing. Specific emphasis is given to exploring whether using SSL is effective in the airborne lidar bathymetry domain, and assessing if SSL can reduce the amount of labeled data required for model training while maintaining or possibly improving performance across differing survey sites. This approach splits a typical neural network training pipeline into two phases: An encoder is first pretrained using a contrastive SSL loss function on unlabeled data, followed by fine-tuning on labeled data in phase two. Due to the nature of SSL objectives focusing on the input data itself for a learning signal, no additional ground-truth labels are required during the pre-training stage. Our findings demonstrate that robust bathymetric point embeddings can be effectively extracted without any labels. Furthermore, this work suggests that leveraging unlabeled datasets during pre-training yields better downstream performance by enhancing generalization across unseen survey sites.

Document Type

Master's Thesis

First Advisor

Thomas Butkiewicz

Second Advisor

Kim Lowell

Third Advisor

Laura Dietz

Degree Name

Master of Science

Download

Share

COinS