RAPID CLASSIFICATION OF INFRA-RED … · match best (hull shape vs. absorption feature) and which do not ... minerals, unlike hull-shapes) Objective and reproducible result SUMMARY

RAPID CLASSIFICATION OF INFRA-RED HYPERSPECTRAL IMAGERY OF ROCKS WITH DECISION TREES AND WAVELENGTH IMAGES F.J.A. VAN RUITENBEEK, H.M.A VAN DER WERFF,

W.H. BAKKER, F.D. VAN DER MEER, C.H. HECKER, K.A.A. HEIN

Specim hyperspectral camera, sisuchema setup Wavelength range: 1000-2500 nm (short-wavelength infrared) # pixels across: 384 Spatial resolution: 26 µm

INTRODUCTION

Specim.fi

INFRA-RED HYPERSPECTRAL IMAGE ACQUIRED WITH SISUCHEMA

Pixel size: 26 µm

384 pixels

1396 pixels

Reflectance at 1650 nm:1650nm

Silicified, sericitized dacite‐andesite

Provides detailed information on mineralogical composition and microstructure of rocks

Enables characterization of rock type and rock forming processes Spatial scale comparable to “traditional” thin section Input in predictive modeling of rock chemistry, e.g. ore grade, and

other physical-chemical parameters

USE OF HIGH SPATIAL RESOLUTION IR IMAGERY

Image-sizes are typically large (> 1 GB per raw image) Images contain many pixels ( ~ 1 million per image) and bands

(288) Easy generation of large volumes of image data (1 image

acquired in ~5 minutes, incl. sample preparation) Interpretation of imagery is labor intensive and time consuming

(and often subjective)

PROBLEM

Methods are needed for rapid assessment of mineralogical composition

SISUCHEMA VERSUS HYMAP

~512 pixels126 bands450‐2500nm

Hymap (airborne sensor) SisuChema

384 pixels288 bands1000‐2500nm

Atmospheric interference

SISUCHEMA VERSUS ASD

1 point spectrum2151 bands350‐2500nm

ASD SisuChema

384 pixels288 bands1000‐2500nm

Many strategies involve spectral matching of image and reference spectra and thresholding, e.g. Spectral Angle Mapper

Limitations of these methods: A prioiry knowledge of scene is required for the selection of

reference spectra Matching statistics doesn’t show which parts of the spectrum

match best (hull shape vs. absorption feature) and which do not Selection of threshold for a “match” is rather subjective

INTERPRETATION OF HYPERSPECTRAL IMAGERY

MATCHING OF IMAGE AND REFERENCE SPECTRA

Is this a good match?

Molecular bond

Wavelength position of absorption features is used as the dominant spectral charateristic in the interpretation of reflectance spectra

It is directly related to molecular bond in crystal lattice and often specific to (groups) of minerals

This information is extracted from wavelength images calculated from the IR imagery

A decision tree is used for classification of the wavelength imagery

APPROACH IN THIS STUDY

CALCULATION OF WAVELENGTH POSITION

where w(x) is the interpolated reflectance value atposition x; x is the wavelength position in μm; a,b,c are the coefficients of the parabola function.

where wmin is the interpolated wavelength position at minimum reflectance; a,b is the coefficients of the parabola function.

where depth is the interpolated depth of absorption feature.

W1, D1: Wavelength and depth of deepest featureW2, D2: Wavelength and depth of 2nd featureW3, D3: Wavelength and depth of 3rd feature

W1

D1

W3

D3

W2

D2

Wavelength mapW1 fused with D1

2100nm

2400nm

Wavelength image

Only for exploratory analysis -> no classified mineral map Small variation in wavelength positions often not visible Deep absorption features dominate over shallow features

LIMITATION OF WAVELENGTH MAPPING

Silicified, chloritised amygdaloidal (dacite)‐andesite

Amygdale L: 3.2mm PP Amygdale L: 3.2mm XP

classification using decision treesClassified map

D1 > 0.05

W1 > 2225nm

W1 > 2300nm

Scatter plot of wavelength image

2100nm

2400nm

Wavelength image

Based on analysis absorption features in spectra of USGS spectral library and other spectra (total of 400+ spectra)

DESIGN OF DECISION TREE

Decision tree 2100‐2400nm (Al‐OH, Fe‐OH, Mg‐OH & carbonate features):

W1 2200‐2210W2 2340‐2400

W1 2210‐2220W2 2340‐2400

W1 2200‐2210W2 2160‐2280

CLASSIFICATION WITH DECISION TREE

chalcedony_cu91‐6a.4502.aschydrogrossular_nmnh120555.10236.ascillite_gds4.10903.ascillite_imt1.10982.ascillite_imt1.11041.ascmontmorillonite_saz1.14498.ascmontmorillonite_sca2.14557.ascmuscovite_gds116.15173.ascmuscovite_gds118.15287.ascmuscovite_hs24.15512.ascmuscovite_il107.15566.ascvesuvianite_hs446.23527.asc

dickite_nmnh46967.6913.ascendellite_gds16.7379.aschalloysite_cm13.8921.asckaolinite_cm3.11788.asckaolinite_cm5.11846.asckaolinite_cm7.11904.asckaolinite_cm9.11962.asckaolinite_gds11.12060.asckaolinite_kga1.12117.asckaolinite_kl502.12272.asckaolinite_pfn1_kga2.12176.asc

goethite_mpcma2b.8351.ascillite_il105.10969.asclepidolite_nmnh105541.12766.asclepidolite_nmnh88526‐1.12832.ascmargarite_gds106.13344.ascmontmorillonite_cm20.14324.ascmontmorillonite_cm26.14382.ascmuscovite_gds107.14887.ascmuscovite_gds114.15116.ascmuscovite_gds117.15230.ascmuscovite_gds119.15344.ascmuscovite_gds120.15401.ascmuscovite_hs146.15457.ascnanohematite_br93‐34b2.15589.ascorthoclase_hs13.17283.ascroscoelite_en124.19682.ascspodumene_hs210.21114.asctourmaline_hs282.22996.asc

Short‐list of candidate spectra

Albedo – R1650

Rock sample: Silicified, sericitised amydaloidal andesite

Classified

W1 2200‐2210W2 2340‐2400

W1 2210‐2220W2 2340‐2400

W1 2200‐2210W2 2160‐2280



Short‐list of candidate spectraRock sample: Silicified, sericitised amydaloidal andesite

Albedo – R1650 Classified

W1 2200‐2210W2 2340‐2400

W1 2210‐2220W2 2340‐2400

W1 2200‐2210W2 2160‐2280



Short‐list of candidate spectraRock sample: Silicified, sericitised amydaloidal andesite

Amygdale 2.5mm

Albedo – R1650 Classified

Hydrothermally altered rocks Associated with VMS Cu-Zn deposits Pervasive alteration of volcanic rock Archean (3.2 Ga) submarine setting

CASE STUDY

Albedo – reflectance at 1650nm

Weakly sericite altered and silicified muddy chert

Silicified, seriticized xenocrystic‐phenocrystic (dacite)‐andesite

Silicified, sericitized phenocrystic (dacite)‐andesite

Silicified, sericitized weakly phenocrystic dacite

Silicified, sericitized weakly phenocrystic quenched dacite


Silicified, sericitized amygdaloidal andesite

Silicified, sericitized weakly amydaloidal titanium‐rich andesite

Ferruginous, chloritised basalt

Ferruginous, chloritized (pyroxene‐bearing) andesite

Silicified, and chloritised amygdaloidal (dacite)‐andesite

Micrograph thin section

Al‐rich illite‐muscovite <2210nm

Shallow Fe‐OH feature 2260‐

2300nm

Fe‐chlorite

Al‐poor illite‐muscovite >2210nm

kaolinite

Al‐rich illite‐muscovite

< 2203nm – chert

Chlorite

Mg‐Fe chlorite

Interpreted mineralogy:


Classification – general decision tree

Pervasive alteration


Illite‐muscovite alteration

Kaolinite filled amygdales

Illite‐musc rich amygdales

Zonation:• Al‐rich illite‐muscovite • Al‐poor illite‐muscovite • Chorite +/‐ illite‐muscovite

Classification of wavelength images with decision trees provides method for rapid assessment of mineral composition

No a priory information on mineralogy of rock sample is required Focus on mineral absorption features (diagnostic for many

minerals, unlike hull-shapes) Objective and reproducible result

SUMMARY AND CONCLUSIONS

Scene-specific optimization of decision tree Automation of processing steps Extraction of microstructural / textural information

FURTHER WORK

Spare slides:

Objective: To optimize decision tree for specific scene – sample set Enhancement of spectral variation in wavelength images Calculation of summary products, such as illite and kaolinite

crystallinity, ferrous drop, etc Visual-spatial analysis of contrast enhanced images and selection

of additional end member ROIs Analysis of ROIs: Spectra and scatter plots of wavelength

positions and depth of absorption features and summary products Improvement of slicing intervals and update of decision tree

STEP 2 – DETAILED IMAGE ANALYSIS

Albedo: Reflectance at 1650nm

Illite‐musc crystallinity

Summary product:Illite‐musc crystallinity = Depth H20 / Depth Al‐OH

Depth Al‐OH

Depth H20

Classified with decision tree

CC of W1, W2,W3 between 1850‐2100nm

ROIs




3.25

2

phenocrysts

matrix

xenocrysts



Classified with decision tree

Update decision three with crystallinity data

Illite/muscovite matrix

Well‐ordered illite/muscovite phenocryst

Poorly‐ordered illite/muscovite

Interpretation:

Phenocryst Length: 1.4 mm

Xenocryst Length: 2 mm


Weakly sericite altered and silicified muddy chert

Silicified, seriticized xenocrystic‐phenocrystic (dacite)‐andesite

Silicified, sericitized phenocrystic (dacite)‐andesite


Silicified, sericitized weakly phenocrystic quenched dacite


Silicified, sericitized amygdaloidal andesite

Silicified, sericitized weakly amydaloidal titanium‐rich andesite

Ferruginous, chloritised basalt

Ferruginous, chloritized (pyroxene‐bearing) andesite

Silicified, and chloritised amygdaloidal (dacite)‐andesite

Micrograph thin section



2300nm

Fe‐chlorite


kaolinite


< 2203nm – chert

Chlorite

Mg‐Fe chlorite

Ordered illite/muscovite

Disordered illite/muscovite

“Ordered” kaolinite








Illite‐muscovite alteration

Kaolinite filled amygdales

Illite‐musc rich amygdales

Zonation:• Al‐rich illite‐muscovite • Al‐poor illite‐muscovite • Chorite +/‐ illite‐muscovite



Classification –decision tree ‐ sample specific



2300nm

Fe‐chlorite


kaolinite


< 2203nm – chert

Chlorite

Mg‐Fe chlorite



“Ordered” kaolinite




PhenocrystsXenocrystsAmygdales

Volcanic texture!

RAPID CLASSIFICATION OF INFRA-RED … · match best (hull shape vs. absorption feature) and which do not ... minerals, unlike hull-shapes) Objective and reproducible result SUMMARY

Documents