Utilization of full waveform LIDAR and hyperspectral data for forest structure characterization

Jinha Jung, Purdue University

Abstract

Understanding characteristics of forest structure at regional, continental, and global scales is of increasing importance because of its critical role in addressing diverse issues such as global carbon cycle model, forest management, and studies of wildlife habitat. Structural characteristics provide valuable insights related to forest ecosystem function at individual tree, stand, forest, and global scales. Although field sampling based approaches are reliable and accurate for individual trees and at the plot scale, these methods are time consuming, expensive, and limited to local scale studies. Results may also be biased due to human interpretation and variable field conditions during an extended field campaign. To overcome the limitations of the field sampling based approaches, remote sensing technologies have been widely utilized to characterize forest structure at both local and global scales. Recently, the use of multi-sensor data for characterizing forest structure has gained significant attention because of the capability to exploit complementary information on targets. Although several combinations of multi-sensor data and various fusion approaches have been investigated for mapping forest structure, little research has focused on the integration of full waveform LIDAR and hyperspectral data. While synergism exists between hyperspectral and full waveform LIDAR data for characterizing forest structure, issues related to different acquisition modalities and extraction of relevant predictive features must be addressed. These issues motivate development of a new approach for characterizing forest structure using full waveform LIDAR and hyperspectral data. In the proposed approach, full waveform LIDAR data are co-registered with hyperspectral data by reconstructing representative waveforms over a common grid structure, and low dimensional features are extracted using unsupervised feature extraction methods. The extracted features are then used to develop predictive regression models in combination with field measurement data. For development of the new approach, this dissertation focuses on three primary research areas. First, a sequential LIDAR waveform decomposition algorithm is developed since the co-registration step involves decomposing a waveform into a mixture of Gaussians. The proposed algorithm utilizes a sequential approach, where the goal is to reduce computation, but provide a good approximation to the waveform. It improves computational efficiency by utilizing a region growing algorithm for initial parameter estimation and applying a series of increasingly more robust, but computationally demanding optimization techniques. The sequential algorithm is applied to ICESat waveform data, and experimental results demonstrate that the proposed algorithm utilizes a smaller number of components to decompose the waveforms, while it provides better approximation than traditional decomposition algorithms, especially for complex waveforms. Second, a new waveform reconstruction scheme is developed so that unsupervised feature extraction methods can be directly applied to extract lower dimensional features from the LIDAR waveform data and performance of linear (PCA) and nonlinear (Isomap) feature extraction methods are investigated. Experimental results reveal that the Isomap transformation successfully discovers a set of low dimensional features that can be used to characterize forest structure without manual interpretation. Finally, the proposed approach to utilize full waveform LIDAR and hyperspectral data for forest structure characterization is applied to data acquired over the La Selva Biological Station which is one of the most actively studied tropical rain forests. LAI (Leaf Area Index) measurements acquired directly from a modular tower are used as ground reference data. Experimental results indicate that the best fitting models can be obtained when features extracted from two data sets were included in the prediction models. Although inclusion of two data sets yielded improved prediction capability, contribution from the hyperspectral data was greater for the high canopy sites than for the low canopy sites. The main contribution of the proposed approach can be summarized as two factors; (1) the ability to co-register full waveform LIDAR with hyperspectral data at the pixel level, and (2) automatic discovery of meaningful lower dimensional features from the integrated data using unsupervised feature extraction algorithms.

Degree

Ph.D.

Advisors

Crawford, Purdue University.

Subject Area

Forestry|Environmental engineering|Remote sensing

COinS