Abstract

Light detection and ranging (lidar) forest models are important for studying forest composition in great detail, and for tracking objects in the understory. In this study we used DIRSIG, a first-principles and physics-based simulation tool, to turn the lidar data into voxels, towards classifying forest voxel types. A voxel is a 3D cube where the dimension represents a certain distance. These voxels are split into categories consisting of background, leaf, bark, ground, and object elements. Voxel content is then predicted from the provided simulated and real National Ecological Observation Network (NEON) data. The inputs are 3D neighborhood cubes which surround each voxel, which contain surrounding lidar signal and content type information. Provided simulated data are from two sources: a VLP-16 drone, which collects discrete lidar data close to the canopy, and the NEON Airborne Observation Platform (AOP), which is attached to an airplane flying 1000 m above ground level and collects both discrete and waveform lidar data. Different machine learning algorithms were implemented, with 3D CNN algorithms shown to be the most effective. The Keras library was used, since creating the layers with the sequential model was regarded as an elegant approach. The simulated VLP-16 waveform data were significantly more accurate than the simulated NEON waveform data, which was attributed to its proximity to the canopy. Leaves and branches exhibited acceptable accuracies, due to their relatively random shapes. However, ground and objects in both cases had very high accuracy due to the high intensities and their rigid shapes, respectively. A sample of real NEON waveform lidar data was used, though the sample primarily focused on the canopy region; however, most of the voxels were correctly predicted as leaves. Additional channels were added to the input voxels in order to improve accuracy. One input parameter which proved to be very useful were the local z-values of each input array. Additionally, the Keras Tuner framework was used to obtain improved hyperparameters. The learning rate was reduced by a factor of 10, which provided slower, but steadier convergence towards accurate predictions. The resulting accuracies from the predictions are promising, but there is room for improvement. Different ML algorithms that use the point cloud should also be considered. Further segmentation of forest classes is another possibility. For example, there are different types of trees and bushes, so each tree or bush could have its own unique classes, which would make predicting the shapes much easier. Overall, discovering a method for accurate object prediction has been the most significant finding. For the ground truth models, the best object precision is approximately 99% and the best recall is 78%.

Library of Congress Subject Headings

Optical radar--Data processing; Neural networks (Computer science); Convolutions (Mathematics)

Publication Date

7-18-2023

Document Type

Thesis

Student Type

Graduate

Degree Name

Imaging Science (MS)

Department, Program, or Center

Chester F. Carlson Center for Imaging Science (COS)

Advisor

Jan van Aardt

Advisor/Committee Member

John Kerekes

Advisor/Committee Member

Carl Salvaggio

Campus

RIT – Main Campus

Plan Codes

IMGS-MS

Download

Share

COinS