Evaluating the capability of machine-learning algorithms and object-oriented classification techniques using hyperspectral remote sensing for the discrimination of Australian native forest species in south-eastern Australia
Doctor of Philosophy
School of Earth and Environmental Sciences
Shang, Xiao, Evaluating the capability of machine-learning algorithms and object-oriented classification techniques using hyperspectral remote sensing for the discrimination of Australian native forest species in south-eastern Australia, Doctor of Philosophy thesis, School of Earth and Environmental Sciences, University of Wollongong, 2013. http://ro.uow.edu.au/theses/4082
With growing environmental awareness in the face of climate change, more accurate quantitative information on vegetation species/community composition is required now more than ever before for the conservation of native vegetation species and the sustainable management of native natural ecosystems. The traditional approach to obtain substantial information of forest species involves labour-intensive and timeconsuming fieldwork and has usually been limited to small spatial extents. Hyperspectral remote sensing enables the discrimination of forest species at both a higher spectral resolution and a much broader continuous spectrum than does conventional multispectral remote sensing. It provides an alternative method for obtaining detailed forest species information over relatively large geographical landscapes, and in remote and inaccessible areas. This thesis explored the capability of machine-learning algorithms and object-oriented classification techniques for the discrimination of specific Australian native forest species, particularly eucalyptus species (Eucalyptus spp.) at the leaf, canopy and community levels using hyperspectral remote sensing on the Beecroft Peninsula, New South Wales, Australia.
The study started with the statistical discrimination of spectral reflectance to demonstrate that the studied species can be theoretically discriminated from each other, and to investigate the key wavelengths suitable for the discrimination. A preliminary classification was then carried out using four traditional classification algorithms and pixel-based techniques. The best preliminary classification result was set as the benchmark against which the following tested algorithms were evaluated.
Subsequently, the state-of-the-art machine-learning classification algorithms, Support Vector Machine (SVM), AdaBoost and Random Forest (RF), applied using objectoriented classification techniques, were investigated for their ability to discriminate between the studied species. Additional textural information (Geographic Object-based Image Analysis, or GEOBIA) acquired from HyMap imagery and data transformation algorithms (Derivative, Continuum Removal and Wavelet) were assessed as key factors in the potential improvement of the machine-learning classification results.
The study results showed significant differences among the studied species at wavelengths of 497 nm, 680 nm, 725 nm, 780 nm, 850 nm, 1578 nm, 1690 nm, 2170 nm, and 2356 nm in the spectral reflectance data measured under laboratory conditions. All machine-learning classification algorithms significantly outperformed traditional algorithms in terms of classification results. At the leaf level, RF achieved the best classification accuracy, while SVM outperformed the other algorithms at both the canopy and community levels. The use of either textural information (GEOBIA texture analysis) or data transformation algorithms can effectively improve machine-learning classification results.
The results of this research provide important insights into the benefits and challenges of the classification of Australian native forest species. It is a major challenge to obtain accurate classification results at the community level, because of the high variation of the species composition. Studies at the community level should include calculation of textural information to ensure adequate discriminative information for classification. Classification at the canopy level does not have as much composition variation as the community level, and this improves the classification accuracy dramatically. However, because of the open canopy structure of eucalyptus species, the background (e.g., soil, twigs) increases the variation in the spectral reflectance, which limits the classification accuracy at the canopy level to some degree. The object-oriented classification could isolate the gaps of tree canopies presented in the imagery, and the canopy-level study should consider using data transformation methods to enhance the discriminative features for classification. There are a number of benefits that justify these studies. This thesis presents the first attempt to combine data transformation algorithms and texture analysis with a machine-learning algorithm for discriminating eucalyptus species. The classification results illustrate that the methodology presented here can be effectively used for the discrimination of specific Australian native forest species, which results in more accurate vegetation information. This will help forest managers to quantify the different tree species available within a forest, or within a given section of a forest, within a shorter time, and with greater accuracy. This could lead to improvements in environmental sustainability through better management and protection of native vegetation communities. Accurate species information derived from mapping is important for determining the spatial distribution of wildfire fuels and the type, location and sensitivity of previous forest burns, and can also be used in the design of fire behaviour models. In addition, accurate species information can help to investigate the fire disturbance of forest species. This is very important for forest conservation. This could lead to improvements in environmental sustainability through the better management and protection of native vegetation communities.