Large-scale feature selection from imaging spectrometer data using a genetic algorithm for invasive species mapping and monitoring

Invasive species pose the single greatest threat of natural disaster in this century. With rare exceptions, the estimated annual cost of invasive species to the U.S. exceeds all other natural disasters combined. Invasive plant species can rapidly displace native vegetation and upset the balance in an ecosystem. Accurate and rapid identification and monitoring of effected areas are critical to maintaining control in environments where invasive species have or could develop.The goal of this research is to develop a method for selecting the near-optimal spectral band subset from hyperspectral imagery (HSI) data to improve classification accuracy and feature transparency. The research presented in this dissertation is focused on the implementation and testing of a genetic algorithm (GA) to reduce the number of spectral bands that are needed to produce an accurate and rapid classification map using HSI data. This algorithm was applied to the invasive species mapping and monitoring problem and can be used for similar "needle in a haystack" problems.This research has also provided a sensitivity analysis for a number of the parameters used by a GA including: the chromosome population size the number of generations required achieve an optimal feature set the crossover and mutation rates and the size of the feature subset. This research should provide a base for additional research into the application GA's for feature selection and extraction in high dimensional remote sensing data. Statistics of the time it takes to run the algorithm have also been captured in order to give users a sense of how this algorithm compares to the runtime for other feature extraction methods such as the Minimum Noise Fraction (MNF) algorithm.The contribution of this research is a demonstrated method to provide a high accuracy feature selection technique to identify invasive species that provides direct feature traceability to class spectra, through the development of a genetic algorithm. The method can be easily adapted to high dimensional remote sensing problems beyond the one addressed in this research such as extending it to the selection of optimal features from multiple remote sensing data types (spatial and texture features in panchromatic or SAR data in addition to spectral).