Classifying FIA Forest Type from a Fusion of Hyperspectral and LiDAR Data

Abstract

In this study we develop a methodology for classifying FIA defined forest type across the Tanana Inventory Unit (TIU) using a fusion of hyperspectral and LiDAR data. The hyperspectral and LiDAR data used in this study were collected as part of the 2014 acquisition with the NASA Goddard's LiDAR, Hyperspectral & Thermal Imager (G-LiHT). In order to determine the best classification method, we tested 5 classification algorithms: Naive Bayes Classifier, K-Nearest Neighbor, Multinomial Logistic Regression, Support Vector Machine, and Random Forests. Each model was trained and validated using the forest type corresponding to each FIA subplot, alongside raw hyperspectral data (114 spectral bands in total), hyperspectral vegetation indices, and selected LiDAR-derived canopy height and topography metrics. Six different combinations of this input data were tested to determine the most accurate classification algorithm and model inputs. A 3-fold cross validation was performed in order to ensure that all data was included in both training and validation, but never within the same model. Of the five models and six model input combinations tested, we found Random Forest with hyperspectral vegetation indices as well as topography and canopy height metrics as model inputs had the highest accuracy at 77.53% overall. With the completion of this work, we hope to use this “best” model to classify forest types across the Tanana Inventory Unit in central inland Alaska where there is G-LiHT coverage. There are three primary sections of this thesis document. The first is this section, an informal introduction to the study wherein we introduce the study, the people involved, and the lessons we learned along the way. The second section is the research performed. This is the publication-ready paper which details the work performed in this study. It provides an overview of the relevant literature, methods and materials used, and discusses the implications of the findings from this study. The third and final section is an extended discussion of future work. It gives detailed descriptions of future research that can be undertaken following this study, and gives open and detailed critique of the methods and materials used in this study in order to inform future research. It also provides a detailed description of the bootstrap aggregation process that was attempted to improve accuracies.