Abstract
Forests play a significant role in the world's environment, economy and society. Forest inventory aims to provide forest attributes such as location, species composition, timber volume at stand level, and also aims to provide individual tree parameters such as the location and species of each tree, crown size, and tree health. Modern remote sensing techniques provide great potential for forest inventory and analysis to be conducted automatically, accurately and cost-efficiently. This dissertation investigated the use of remotely sensed data in automated forest inventory at both forest stand and individual tree levels.
At the forest stand scale, the research investigated the synergistic use of high spatial resolution multispectral imagery (i.e., QuickBird: 2.4 m ground sampling distance) and LIDAR data (3 m nominal posting) for forest species classification using an object-based approach. Machine learning decision trees were used to build classification rule sets. The results showed the integration of spectral and LIDAR data, in both image segmentation and object-based classification, significantly improved forest classification compared to using either data source independently.
At the individual tree scale, automatic tree crown detection and delineation algorithms were considered. Subsequent to reviewing the published approaches, the research compared three representative methods--watershed segmentation, region growing and valley-following--for tree crown detection and delineation on softwood and hardwood sites using a vertical aerial image and QuickBird panchromatic imagery with an 11° view angle. The research also demonstrated a standardized accuracy assessment framework under which the three algorithms were compared. Building on the understanding of current algorithms, a new approach for individual tree crown detection and delineation applicable under various imaging conditions was developed. The algorithm was developed based on an active contour model and a hill-climbing algorithm. It considered both spectral and geometric characteristics of tree crowns under various imaging conditions, and also considered the expert knowledge of the forest stand to improve crown detection accuracy. In comparison with the region growing algorithm, the newly developed algorithm provided higher accuracy in tree detection and delineation using three sets of images acquired in different imaging conditions. Ground validation also showed the new algorithm provided accurate crown diameter estimations.