Institute of Industrial Information Technology Applicability of Hyperspectral Fluorescence Imaging to Mineral Sorting Optical Characterization of Materials, March 19, 2015 Sebastian Bauer, M.Sc. (Head: Prof. Puente), Faculty of Electrical Engineering and Information Technology KIT University of the State of Baden-Wuerttemberg and National Research Center of the Helmholtz Association www.iiit.kit.edu
Motivation minerals are difficult to discriminate by means of color images different techniques available: laser-induced breakdown spectroscopy (LIBS) near-infrared hyperspectral imaging (NIR-HSI) X-ray fluorescence (XRF) X-ray transmission (XRT) all those methods have their own advantages and disadvantages in terms of accuracy, speed, cost, handling etc. are there other discrimination techniques with less disadvantages? 2 30 March 2015
Agenda 1. Motivation 2. Light-emitting phenomena 3. Experimental setup 4. Hyperspectral fluorescence images and their characteristics 5. Classification and its results 6. Conclusion and outlook 3 30 March 2015
Agenda 1. Motivation 2. Light-emitting phenomena 3. Experimental setup 4. Hyperspectral fluorescence images and their characteristics 5. Classification and its results 6. Conclusion and outlook 4 30 March 2015
Other physical light-emitting phenomena intensity idea: low light phenomena? X-rays ultraviolet visible infrared microwaves wavelength 5 30 March 2015
Photoluminescence incident photon excites electron electron emits photon when falling back to ground level emitted photon carries less energy than incident photon shorter wavelength fluorescence: short-lived kind of photoluminescence (<10-7 s) excited states (singlet) internal loss spin conversion triplet state absorption ground state (singlet) fluorescence phosphorescence 6 30 March 2015
Fluorescence characteristics factors enabling/enhancing fluorescence: presence of enough photoluminescent centers with activators, e.g., certain kinds of ions quenching: factors eliminating fluorescence collisional quenching: collisional encounters between the fluorophore and quencher static quenching: fluorophore and quencher form a complex that does not fluoresce Relatively small proportions of the sample determine the fluorescence Fluorescence color depends on the origin of the sample 7 30 March 2015
Agenda 1. Motivation 2. Light-emitting phenomena 3. Experimental setup 4. Hyperspectral fluorescence images and their characteristics 5. Classification and its results 6. Conclusion and outlook 8 30 March 2015
Experimental setup camera filter light source: 300 W Xenon lamp (~220-2000 nm) optical filter: acousto-optic tunable filter (AOTF) camera: electron-multiplying charge-coupled device (EMCCD) mirror sample light source with monochromator 9 30 March 2015
Image acquisition excitation: 260 60 nm in steps of 20 nm (full width at half max.: 15 nm) acquisition: 450 790 nm, full width at half max. and increment 4 nm each exposure time: 24 113 s, depending on excitation wavelength excitation acquisition 200 400 600 800 wavelength/nm 10 30 March 2015
Samples class number type 1 magnesite, white 2 2 magnesite, yellow 2 3 magnesite, cauliflower 2 4 magnesite, dark 2 5 magnesite, middlings 2 6 talc, partly with pyrite 7 7 dolomite 2 8 granite 2 9 chlorite 3 10 quartz 2 number of samples 11 30 March 2015
Agenda 1. Motivation 2. Light-emitting phenomena 3. Experimental setup 4. Hyperspectral fluorescence images and their characteristics 5. Classification and its results 6. Conclusion and outlook 12 30 March 2015
Image example: excitation 320 nm image at 526 nm original image at 526 nm filtered image at 526 nm 13 30 March 2015
Image example: excitation 320 nm, acquisition 526 nm granite granite talc talc talc talc/pyr talc/pyr talc/pyr talc/pyr yellow magnesite white magnesite chlorite chlorite chlorite 14 30 March 2015
average pixel value/counts Magnesite sample mean spectra (360 nm exc.) 700 600 500 400 300 200 100 origin 1 origin 2 origin 3 class 1 white class 1 white class 2 yellow class 2 yellow class 3 cauliflower class 3 cauliflower class 4 dark class 4 dark class 5 middlings class 5 middlings 0 450 500 550 600 650 wavelength/nm 700 750 800 15 30 March 2015
relative mean value Magnesite sample mean spectra normalized origin 1 origin 2 origin 3 class 1 white class 1 white class 2 yellow class 2 yellow class 3 cauliflower class 3 cauliflower class 4 dark class 4 dark class 5 middlings class 5 middlings 0 450 500 550 600 650 wavelength/nm 700 750 800 16 30 March 2015
average pixel value/counts Talc sample mean spectra (360 nm exc.) 700 600 500 400 sample 1 sample 2 sample 3 sample 4 sample 5 sample 6 sample 7 300 200 100 0 450 500 550 600 650 wavelength/nm 700 750 800 17 30 March 2015
relative mean value Talc sample mean spectra normalized sample 1 sample 2 sample 3 sample 4 sample 5 sample 6 sample 7 0 450 500 550 600 650 700 750 800 wavelength/nm 18 30 March 2015
average pixel value/counts Chlorite sample mean spectra (360 nm exc.) 90 80 70 60 sample 1 sample 2 sample 3 50 40 30 20 450 500 550 600 650 700 750 800 wavelength/nm 19 30 March 2015
Spatial variability magnesite 280 nm exc. 300 nm exc. 800 1000 400 500 1200 0 320 nm exc. 1800 340 nm exc. 360 nm exc. 2000 600 900 1000 0 500 600 700 500 600 700 500 600 700 20 30 March 2015
Spatial variability dolomite 280 nm exc. 300 nm exc. 250 350 100 200 0 320 nm exc. 250 300 340 nm exc. 360 nm exc. 250 100 150 100 0 500 600 700 500 600 700 500 600 700 21 30 March 2015
Agenda 1. Motivation 2. Light-emitting phenomena 3. Experimental setup 4. Hyperspectral fluorescence images and their characteristics 5. Classification and its results 6. Conclusion and outlook 22 30 March 2015
band 2 Classification used classifiers linear discriminant analysis (LDA) k-nearest neighbors algorithm (KNN) spectral angle mapper (SAM): : test spectrum : reference spectrum band 1 23 30 March 2015
Classificator training one sample of each class: training data other samples: test data LDA training: all pixels of the training sample used as training pixels KNN training: N pixels of each training sample used as training pixels for fair comparison (N: number of pixels of smallest training sample) SAM training: mean spectra of the training samples used as class reference spectra 24 30 March 2015
Classification results excitation LDA KNN SAM 260 nm 0.56 0.75 0.63 280 nm 0.50 0.69 0.69 300 nm 0.38 0.50 0.75 320 nm 0.44 0.31 0.50 340 nm 0.44 0.31 0.25 360 nm 0.44 0.38 0.25 low classification results due to no preprocessing similar spectra class composition 25 30 March 2015
Agenda 1. Motivation 2. Light-emitting phenomena 3. Experimental setup 4. Hyperspectral fluorescence images and their characteristics 5. Classification and its results 6. Conclusion and outlook 26 30 March 2015
Summary of hyperspectral mineral fluorescence images properties hyperspectral fluorescence images can be a useful tool for discriminating spectra spectra vary between different mineral samples of the same class origin dependence intensity information allows for, e.g., normalization hyperspectral investigations provide valuable information about spatial properties of minerals not all minerals fluoresce? Depends on illumination power (!) if hyperspectral data are needed, their acquisition (VIS-range) can be more cost efficient than the one of hyperspectral NIR images 27 30 March 2015
Outlook determination of feasible excitation wavelengths (problemdependent) implementation using more powerful light sources investigate intra-sample variability investigate different mineral species improve classification 28 30 March 2015
Thank you for your attention 29 30 March 2015
average pixel value/counts All samples mean spectra (360 nm exc.) 280 nm exc. 360 nm exc. 350 500 300 400 250 200 300 150 200 100 100 50 magnesite 1 magnesite 2 magnesite 3 magnesite 4 magnesite 5 talc dolomite granite chlorite quartz 500 600 700 500 600 700 wavelength/nm 30 30 March 2015
intensity (a.u.) intensity (a.u.) Dolomite excitation dependence 300 nm exc. 320 nm exc. origin 1 origin 2 origin 3 sample 1 sample 2 0 500 600 700 500 600 700 340 nm exc. 360 nm exc. 0 500 600 700 500 600 700 31 30 March 2015
pixel value/counts Yellow magnesite comparison between samples 160 140 120 sample 1 sample 2 mean sample 1 mean sample 2 100 80 60 40 20 0 450 500 550 600 650 wavelength/nm 700 750 800 32 30 March 2015
relative pixel value Yellow magnesite normalized spectra sample 1 sample 2 mean sample 1 mean sample 2 0 450 500 550 600 650 700 750 800 wavelength/nm 33 30 March 2015