The Role of Environmental Factors in Hyperspectral Measurement and Mineral Classification for Earth resources Exploration
The research presented in this thesis combines systematic laboratory research and quantitative analysis of spectral absorption features and classification repeatability. It investigates the effects of environmental factors on spectral measurements and subsequent mineral classification and establishes guidelines for mineral identification using hyperspectral data. The objective is to understand the systematic effects of varying illumination zenith, moisture content, and shadows. Spectral variation is quantified using reflectance values and absorption feature parameters such as centre wavelength positions, feature depth, and feature area. Based on these observations, the responses of different methods within a spectral classification chain are evaluated.
This research first evaluates the effect of using different endmember selection strategies on the repeatability of multi-temporal classification (chapter 2). The study employs the spectral angle mapper (SAM), a widely used classifier, to classify three multi-temporal AVIRIS hyperspectral images. Seasonal and temporal variations in image acquisition lead to differences in measurement conditions. Two classifications for each image are conducted: one using endmembers extracted from each of the images and another using endmembers averaged from the extracted images. Image-to-image registration and change detection are applied to assess classification repeatability. The results demonstrate that using averaged endmembers achieves approximately 65% repeatability, surpassing the 40% achieved using endmembers extracted from each image.
The effects of changing illumination zenith on hyperspectral measurement and subsequent mineral classification are investigated in chapter 3. The zenith angle defines the angle between the illumination ray and the objectives, ranging from 0° (direct normal) to 90° (horizontal). In this research, spectra from seven pure mineral powders are measured under nine zenith angles (5° to 85° in 10° increments) using an ASD Fieldspec®3 Spectroradiometer in a dark room, calibrated with a Spectralon® white reference panel for each set angle. We create a dataset of pure mineral spectra and their calculated linear mixtures, analyse spectral variation through spectral feature parameters, and classify it using three classifiers: SAM, Euclidean Distance Classifier (EDC), and an Expert System Decision Tree (ESDT). Endmembers for SAM and EDC are taken at a 5° zenith angle, and ESDT is designed using absorption features from the 5° spectra. Results show that, while the spectral feature centre positions remained consistent, spectral reflectance increased with larger zenith angles, with continuum removal extending this effect. SAM and ESDT maintain a 100% repeatability in classifying pure mineral spectra, while EDC shows 95.24% repeatability due to misclassifying spectra at 85°. Classifying linearly mixed spectra is more complex, with SAM, EDC, and ESDT showing varying repeatability levels depending on the mixing fraction. We conclude that changing the zenith angle affects mineral spectra by changing reflectance and altering absorption feature depths, but spectral mixture and classifier selection have a greater impact on the eventual mineral classification than changes in zenith angle.
Changing moisture content causes errors in hyperspectral measurements and subsequent mineral classification. The research presented in this thesis systematically studies the effect of increasing moisture content in different minerals in chapter 4. This study measures reflectance spectra from six different pure mineral powders with different moisture content, ranging from complete saturation to oven-dried in steps of ~5 wt\%. Variation in the shape of spectral features as moisture content increased is quantified using parameters: centre position, absorption depth and area. The spectra are classified using the full wavelength range (350 - 2500 nm) and a subset (2100 - 2360 nm) with the same three classifiers: SAM (sensitive to colour changes), EDC (sensitive to colour and brightness), and an ESDT (using only the centre position of diagnostic features). Results indicate that below a certain moisture threshold, the centre position of spectral features remained consistent, but beyond this threshold, water absorption features dominated and significantly altered absorption feature parameters. Moisture content influences mineral classification, with mineral combinations, classifier types, and wavelength ranges used being influential factors. We conclude that the centre position of spectral features is the least affected parameter, and applying continuum removal in a narrow wavelength range helped mitigate the impact of changing moisture content.
The knowledge gained in chapters 2 - 4, along with the effects of changing shadows on spectral measurement and classification, are tested in a semi-controlled outdoor setting in chapter 5. A camera positioned perpendicularly above measures hyperspectral images from four piles of rock fragments --- two rock types (white marble and dark basalt) in dry and wet conditions --- at multiple times in an afternoon. By analysing spectral feature parameters, such as centre position and absorption depth, we evaluate the effects of changing shadow, moisture, and illumination angle on mineral spectra. Three classifiers --- SAM, EDC, and ESDT --- are used to classify the measured spectra. Results indicate that shadows have the most significant impact on multi-temporal hyperspectral analysis. It is also confirmed that the knowledge gained in Chapters two to four remains valid in an outdoor environment. Endmember selection strategy plays an important role in the classification of data affected by measurement conditions. Classifiers sensitive to spectral colour are more robust under changing conditions than those sensitive to brightness.
In conclusion, the research presented in this thesis systematically studies the effects of varying environmental conditions, illumination zenith, moisture content, and shadows, on the accuracy and repeatability of spectral measurements and subsequent mineral classification.
The findings in this thesis emphasise the significance of understanding and mitigating the systematic effects of changing measurement conditions on hyperspectral data acquisition and analysis. Among the tested factors, shadows display the most significant effect on measurement and classification accuracy, followed by moisture content, with illumination zenith having the least effect. To enhance classification repeatability and reliability, it is crucial to select appropriate spectral wavelength ranges, apply continuum removal based on a full understanding of the data, and use classifiers that are less sensitive to spectral brightness but focus on the centre position of diagnostic spectral features. The incorporation of expert knowledge into data analysis significantly improves the robustness of hyperspectral mineral classification. This research not only advances our understanding of Earth's surface but also provides valuable tools for a wide range of applications in remote sensing.