Field spectroradiometer data : acquisition, organisation, processing and analysis on the example of New Zealand native plants : a thesis presented in fulfilment of the requirements for the degree of Master of Philosophy in Earth Science at Massey University, Palmerston North, New Zealand
The purpose of this research was to investigate the acquisition, storage, processing and analysis of hyperspectral data for vegetation applications on the example of New Zealand native plants. Data covering the spectral range 350nm-2500nm were collected with a portable spectroradiometer. Hyperspectral data collection results in large datasets that need pre-processing before any analysis can be carried out. A review of the techniques used since the advent of hyperspectral field data showed the following general procedures were followed: 1. Removal of noisy or uncalibrated bands 2. Data smoothing 3. Reduction of dimensionality 4. Transformation into feature space 5. Analysis techniques Steps 1 to 4 which are concerned with the pre-processing of data were found to be repetitive procedures and thus had a high potential for automation. The pre-processing had a major impact on the results gained in the analysis stage. Finding the ideal pre-processing parameters involved repeated processing of the data. Hyperspectral field data should be stored in a structured way. The utilization of a relational database seemed a logical approach. A hierarchical data structure that reflected the real world and the setup of sampling campaigns was designed. This structure was transformed into a logical data model. Furthermore the database also held information needed for pre-processing and statistical analysis. This enabled the calculation of separability measurements such as the JM (Jeffries Matusila) distance or the application of discriminant analysis. Software was written to provide a graphical user interface to the database and implement pre-processing and analysis functionality. The acquisition, processing and analysis steps were applied to New Zealand native vegetation. A high degree of separability between species was achieved and using independent data a classification accuracy of 87.87% was reached. This outcome required smoothing, Hyperion synthesizing and principal components transformation to be applied to the data prior to the classification which used a generalized squared distance discriminant function. The mixed signature problem was addressed in experiments under controlled laboratory conditions and revealed that certain combinations of plants could not be unmixed successfully while mixtures of vegetation and artificial materials resulted in very good abundance estimations. The combination of a relational database with associated software for data processing was found to be highly efficient when dealing with hyperspectral field data.