Hyperspectral and Lidar Fusion and the Evaluation of Sampling Methodologies in Remote Sensing

Date of Award

12-2024

Document Type

Thesis

Degree Name

Doctor of Philosophy (PhD)

Department

Department of Electrical and Computer Engineering

First Advisor

Brett J. Borghetti, PhD

Abstract

This dissertation explores methodologies for enhancing the processing, fusion, and semantic segmentation of multimodal hyperspectral and lidar datasets using neural networks, as well as improving sampling methodologies for image-based remote sensing data. The first study introduces composite fusion-style neural network architectures that integrate both pixel- and point-based convolutional layers, facilitating joint processing of 2D hyperspectral images and 3D lidar point cloud data. This approach addresses the challenge of effectively merging diverse data types to improve model performance. The second study expands on this by exploring the direct processing of hyperspectral data in a 3D point cloud format, eliminating the need for feature projection transformations and enabling a unified point-based convolutional approach for both data modalities. The final study systematically examines sampling methods used in remote sensing, highlighting the prevalence of spatial correlation issues caused by random sampling techniques and assessing the effectiveness of various algorithms in mitigating these biases. By establishing a set of desirable characteristics for evaluating sampling methods, this study provides practical guidance to enhance the reliability of model performance assessments. Collectively, these studies offer solutions to current challenges in remote sensing data processing, advancing the field toward more accurate analytical approaches.

AFIT Designator

AFIT-ENG-DS-24-D-032

This document is currently not available here.

Share

COinS