Date of Graduation
Eberly College of Arts and Sciences
Geology and Geography
Tree species composition and health are key attributes for rural and urban forest biodiversity, and ecosystem services preservation. Remote sensing has facilitated extraordinary advances in estimating and mapping tree species composition and health. Yet previous sensors and algorithms were largely unable to resolve individual tree crowns and discriminate tree species or health classes at this essential spatial scale due to the low image spectral and spatial resolution. However, current available very high spatial resolution (VHR) remote sensing data can begin to resolve individual tree crowns and measure their spectral and structural qualities with unprecedented precision. Moreover, various machine learning algorithms are now available to apply these new data sources toward the discrimination and the mapping of tree species and health classes. The dissertation includes an introductory chapter, three stand-alone manuscripts, and a concluding chapter, each of which support the overarching theme of mapping tree species composition and health using remote sensing images.
The first manuscript, now published in the International Journal of Remote Sensing, confirms the utility of combining VHR multi-temporal satellite data with LiDAR datasets for tree species classification using machine learning classifiers at the crown level in a rural forest the Fernow Experimental Forest, West Virginia. This research also evaluates the contribution of each type of spectral, phenological and structural feature for discriminating four tree species: red oak (Quercus rubra), sugar maple (Acer saccharum), tulip poplar (Liriodendron tulipifera), and black cherry (Prunus serotina). The second manuscript investigates the performance of tree species classification in urban settings with three contributions: 1) 12 very high resolution WorldView-3 images (WV-3), whose image acquisition date covering the growing season from April to November; 2) a large forest inventory providing sufficient calibration/validation datasets in Washington D.C.; 3) object-based tree species classification using the RandomForest machine learning algorithm. This manuscript identifies the incremental losses in classification accuracy caused by iteratively expanding the classification to 19 species and 10 genera. It also identifies the optimum pheno-phases and spectral bands for discriminating trees species in urban settings. Building on these promising results from the second manuscript, the third manuscript detect a signal of statistical difference among individual tree health conditions using WorldView-3 images from June 11th, July 30th and August 30th , 2017 in Washington D.C.. It examines six vegetation indices calculated from WorldView-3 images to describe three health condition levels in good, fair and poor, and discusses the effects of green-down phenology for tree health analysis.
Overall, this dissertation research contributes to remote sensing research by combining data from both active and passive sensors to discriminate tree species in rural forest. For the species-rich urban settings, this dissertation illustrates the importance of phenology for tree species classification at crown level using VHR remote sensing images. Finally, this dissertation provides important insights on detecting statistical differences among tree health conditions at individual crown-level in the urban environment using VHR remote sensing images.
Fang, Fang, "Crown-level mapping of tree species and health from remote sensing of rural and urban forests" (2019). Graduate Theses, Dissertations, and Problem Reports. 4068.