Abstract

Recent advances in digital electronics have led to an overabundance of observations from electro-optical (EO) imaging sensors spanning high spatial, spectral and temporal resolution. This unprecedented volume, variety, and velocity is overwhelming our capacity to manage and translate that data into actionable information. Although decades of image processing research have taken the human out of the loop for many important tasks, the human analyst is still an irreplaceable link in the image exploitation chain, especially for more complex tasks requiring contextual understanding, memory, discernment, and learning. If knowledge discovery is to keep pace with the growing availability of data, new processing paradigms are needed in order to automate the analysis of earth observation imagery and ease the burden of manual interpretation.

To address this gap, this dissertation advances fundamental and applied research in deep learning for aerial and satellite imagery. We show how deep learning---a computational model inspired by the human brain---can be used for (1) tracking, (2) classifying, and (3) modeling from a variety of data sources including full-motion video (FMV), Light Detection and Ranging (LiDAR), and stereo photogrammetry. First we assess the ability of a bio-inspired tracking method to track small targets using aerial videos. The tracker uses three kinds of saliency maps: appearance, location, and motion. Our approach achieves the best overall performance, including being the only method capable of handling long-term occlusions.

Second, we evaluate the classification accuracy of a multi-scale fully convolutional network to label individual points in LiDAR data. Our method uses only the 3D-coordinates and corresponding low-dimensional spectral features for each point. Evaluated using the ISPRS 3D Semantic Labeling Contest, our method scored second place with an overall accuracy of 81.6\%. Finally, we validate the prediction capability of our neighborhood-aware network to model the bare-earth surface of LiDAR and stereo photogrammetry point clouds. The network bypasses traditionally-used ground classifications and seamlessly integrate neighborhood features with point-wise and global features to predict a per point Digital Terrain Model (DTM). We compare our results with two widely used softwares for DTM extraction, ENVI and LAStools. Together, these efforts have the potential to alleviate the manual burden associated with some of the most challenging and time-consuming geospatial processing tasks, with implications for improving our response to issues of global security, emergency management, and disaster response.

Publication Date

4-29-2019

Document Type

Dissertation

Student Type

Graduate

Degree Name

Imaging Science (Ph.D.)

Department, Program, or Center

Chester F. Carlson Center for Imaging Science (COS)

Advisor

Carl Salvaggio

Advisor/Committee Member

Hossein ShahMohamad

Advisor/Committee Member

Dave Messinger

Campus

RIT – Main Campus

Share

COinS