ITS INCORPORATION INTO SPARSE UNMIXING FOR MINERAL IDENTIFICATION - - PowerPoint PPT Presentation

▶
its incorporation into sparse unmixing
SMART_READER_LITE
LIVE PREVIEW

ITS INCORPORATION INTO SPARSE UNMIXING FOR MINERAL IDENTIFICATION - - PowerPoint PPT Presentation

Jan 26, 2023 349 likes •662 views

BUILDING A HYPERSPECTRAL LIBRARY AND ITS INCORPORATION INTO SPARSE UNMIXING FOR MINERAL IDENTIFICATION Thanh Bui 1,2 , Beate Orberger 3,4 , Simon B. Blancher 1 , Ali Mohammad- Djafari 4 , Henry Pilliere 5 , Anne Salaun 1 , Xavier Bourrat 6 ,

slide-1
SLIDE 1

BUILDING A HYPERSPECTRAL LIBRARY AND ITS INCORPORATION INTO SPARSE UNMIXING FOR MINERAL IDENTIFICATION

Thanh Bui1,2, Beate Orberger3,4, Simon B. Blancher1, Ali Mohammad- Djafari4, Henry Pilliere5, Anne Salaun1, Xavier Bourrat6, Nicolas Maubec6, Thomas Lefevre5, Celine Rodriguez1, Antanas Vaitkus7, Saulius Grazulis7, Cedric Duée6, Dominique Harang5, Thomas Wallmach1, Yassine El Mendili8, Daniel Chateigner8, Mike Buxton9, Monique Le Guen10

Thanh M. BUI

1) Eramet Research, Eramet Group, Trappes, France; 2) L2S, CNRS, Centrale Supélec, Université Paris-Saclay, France; 3) GEOPS-Université Paris Sud-Paris Saclay, Orsay, France; 4) Catura Geoprojects, Paris, France; 5) ThermoFisher Scientific (TFS), Artenay, France; 6) BRGM, Orléans, France; 7) Vilnius University Institute of Biotechnology, Vilnius, Lithuania; 8) CRISMAT-CNRS, Normandie Université, Caen, France; 9) Delft University of Technology, Delft, The Netherlands; 10) Eramet Nickel Division, Eramet Group, Trappes, France

slide-2
SLIDE 2

Contents

  • Introduction
  • Hyperspectral library
  • Sparse unmixing techniques
  • Results
  • Conclusions and perspectives

Thanh M. BUI

2

slide-3
SLIDE 3

SOLSA project

Thanh M. BUI

SOLSA ID Analyse & Identification in field and industrial applications

Profilometer, RGB camera, VNIR/SWIR cameras, XRF Localisation of ROIs on drill cores XRD – XRF – Raman on ROIs Data processing SOLSA ID A, measurement SOLSA ID A, processing SOLSA ID B, measurement SOLSA ID B, processing depth Drill core (Drill core ID)

H2020 SOLSA (Sonic Online and Sample Analysis) project aims at constructing an analytical expert system for on-line-on-mine-real-time mineralogical and geochemical analyses on sonic drill cores.

3

slide-4
SLIDE 4

Nickel laterites

4

  • Three nickel laterite ore types, based on the dominant minerals hosting Ni:
  • Ni resources:
  • Sulfide ores
  • Ni laterites
  • Ni laterites
  • Consitute 60 – 70% of the

world’s Ni resources

  • Reach 60% of total Ni

production in 2014

  • Contribute 20 – 30% of the

total Co supply.

http://www.malagpr.com.au/terralog-services.html

Ores Mean grades

  • f Ni

Principle ore minerals % of total Ni laterite resources Position in lateritic profiles Oxide 1.0 – 1.6 wt% Goethite, absolane, lithiophorite 60% Mid to upper saprolite and upwards to the plasmic zone Hydrous Mg silicate 1.44 wt% Serpentine, talc, chlorite, sepiolite 32% Mid to lower saprolite Clay silicate 1.0 – 1.5 wt% Smectite, saponite 8% Mid to upper saprolite

Average chemical variations on the laterite profile:

Butt et al., 2013

Thanh M. BUI

slide-5
SLIDE 5

SOLSA ID A system

Thanh M. BUI

5 Profilometer RGB camera VNIR (400 – 1000 nm) camera SWIR (1000 – 2500 nm) camera

slide-6
SLIDE 6

Software development scheme

6

Thanh M. BUI

Data registration software Data processing software

ROI

Data acquisition software

Profilometer RGB camera XRF SWIR camera VNIR camera Conveyor SSD depth

ROI ROI

ROI ROI

ROI

slide-7
SLIDE 7

Hyperspectral imaging for mineralogy identification

7 GMEX, 2008, Pontual et al. 1997

Molecules Dominant absorption features OH 1400nm (1550nm and 1750- 1850nm in some minerals) Water 1400nm and 1900nm AlOH 2160-2228nm FeOH 2230-2295nm MgOH 2300-2370nm CO3 2300-2370nm (and also at 1870nm, 1990nm and 2155nm)

Crystallinity variations -> shape variations Compositional variations -> wavelength shifts

slide-8
SLIDE 8

Hyperspectral unmixing

m1 m3 m2

245 um 245 um Wavelength Reflectance m1 m2 m3 Reflectance Halogen light 𝛽2𝒏2 8

Thanh M. BUI

slide-9
SLIDE 9

Hyperspectral unmixing

  • Statistical approaches (Debigion et al. 2008 ; Altmann et al.,

2015)

  • The likelihood: data generation models
  • Priors: constraints on the endmembers
  • Geometrical approaches (Nascimento et al., 2005; Bioucas-

Dias et al. 2009)

  • The observed hyperspectral vectors: simplex set whose vertices

correspond to the endmembers.

  • Sparse regression

Thanh M. BUI

9

slide-10
SLIDE 10

Sparse unmixing

𝑍 = 𝐵𝑌

Y L x n A L x m X m x n min

𝑌

𝐵𝑌 − 𝑍 𝐺

2 + 𝜇 𝑌 2,1

subject to: X≥ 0, 1𝑼X = 1

Iordache et al., IEEE Trans, 2014

  • The observed image signatures can be expressed in the form of linear combinations
  • f a number of pure spectral signatures known in advance (spectral library).
  • Unmixing amounts to finding the optimal subset of signatures in a spectral library

that can best model each mixed pixel in the scene.

  • The sparse unmixing exploits the usual very low number of endmembers (maximum
  • f 4, Berman et al., CSIRO, 2017) present in real images, out of a spectral library.

10

Thanh M. BUI

slide-11
SLIDE 11

Hyperspectral library

  • Other libraries (e.g., USGS) may not contain spectra of pure

minerals.

  • We wish to include spectra that are collected with our

instruments used in our operational exploration.

  • Minerals found in Ni laterites in New Caledonia may not be

present in other libraries.

11 Nontronite Sepiolite

Reference spectral libraries: USGS: https://speclab.cr.usgs.gov/ NASA ASTER: https://speclib.jpl.nasa.gov/

Thanh M. BUI

slide-12
SLIDE 12

Hyperspectral library

Thanh M. BUI

  • Rock and mineral samples provided

by BRGM, ERAMET and the National Museum of Natural History, France

  • Spectra extraction: ENVI 5.4 and G-

MEX by taking into account the wavelength positions and the relative intensities of the absorption features.

12

slide-13
SLIDE 13

Talc: Mg3Si4O10(OH)2

Thanh M. BUI

13 2304 2388 2287 1389

slide-14
SLIDE 14

Kaolinite: Al2Si2O5(OH)4

Thanh M. BUI

14 2303 2159 1808 1395, 1412 2380

slide-15
SLIDE 15

Sparse unmixing techniques

min

𝑌

𝐵𝑌 − 𝑍 𝐺

2 + 𝜇 𝑌 2,1

subject to: X ≥ 0, 𝟐𝑼X = 1 CLSUnSAL (Collaborative sparse unmixing by variable splitting and augmented Lagrangian): subject to: X ≥ 0, 𝟐𝑼 X = 1 SUnSAL (Sparse unmixing by variable splitting and augmented Lagrangian): min

𝑌

𝐵𝑌 − 𝑍 𝐺

2

subject to: X ≥ 0, 𝟐𝑼 X = 1 FCLS (Fully contrained least squares): The optimization is based on the alternating direction method of multipliers (ADMM) min

𝑌

𝐵𝑌 − 𝑍 𝐺

2 + 𝜇 𝑌 1,1

Bioucas-Dias et al., 2010 Iordache et al., IEEE Trans, 2014 Afonso et al., IEEE Trans, 2011

15

Thanh M. BUI

slide-16
SLIDE 16

Hyperspectral library

16 37 spectra representing 21 minerals have been collected: ankerite, calcite, dolomite, magnesite lizardite, nepouite, antigorite, chrysotite, saponite, montmorillonite, nontronite, kaolinite, pimelite, talc, sepiolite, alunite, asbolane, chromite, diaspore, enstatite, forsterite

Thanh M. BUI

slide-17
SLIDE 17

Hyperspectral unmixing

K FCLS SUnSAL CLSUnSAL SRE Time SRE time SRE time 2 14.24 0.022 14.94 0.254 16.74 0.228 3 6.41 0.019 7.45 0.259 11.95 0.230 4 5.25 0.022 7.07 0.499 7.16 0.453

Signal to reconstruction error (SRE) ratio:

Simulated data: SNR = 40 dB

FCLS: Fully constrained least squares SUnSAL: Sparse unmixing by variable splitting and augmented Lagrangian CLSUnSAL: Collaborative sparse unmixing by variable splitting and augmented Lagrangian

17

Thanh M. BUI

𝑇𝑆𝐹 = 10 log 𝐹 𝒚 2 𝐹 𝒚 − 𝒚 2

slide-18
SLIDE 18

Hyperspectral unmixing

QEMSCAN results RGB image

1 cm

Data acquired from a serpentinized harzburgite sample 18

Thanh M. BUI

slide-19
SLIDE 19

Hyperspectral unmixing

19 Proportion (abundance) of each mineral:

Thanh M. BUI

Computation time: 4 mins

slide-20
SLIDE 20

Hyperspectral unmixing

QEMSCAN results RGB image

1 cm

Data acquired from a serpentinized harzburgite sample Unmixing results 20

Thanh M. BUI

Computation time: 4 mins

slide-21
SLIDE 21

Conclusions and perspectives

  • Using our hyperspectral library, the

CLSUnSAL provided the highest accuracy.

  • Need to improve the computation time.
  • Incorporate the spatial context to the

unmixing problem

  • The library is constantly extended
  • 257 spectra have been extracted for 49

minerals

  • A graphic user interface is under

development

  • Machine learning classification

approaches have been implemented.

Thanh M. BUI

21

slide-22
SLIDE 22

Thank you for your attention!

Thanh M. BUI

22

slide-23
SLIDE 23

SOLSA project

Thanh M. BUI

SOLSA ID Analyse & Identification in field and industrial applications

Profilometer, RGB camera, VNIR/SWIR cameras, XRF Localisation of ROIs on drill cores XRD – XRF – Raman on ROIs Data processing SOLSA ID A, measurement SOLSA ID A, processing SOLSA ID B, measurement SOLSA ID B, processing depth Drill core (Drill core ID)

H2020 SOLSA (Sonic Online and Sample Analysis) project aims at constructing an analytical expert system for on-line-on-mine-real-time mineralogical and geochemical analyses on sonic drill cores.

23 Gibbsite

Diaspore

Hematite A drill core ROI 1 ROI 2 ROI 3

slide-24
SLIDE 24

SOLSA software

Thanh M. BUI

24 Hyperspectral classification and unmixing techniques are being integrated

slide-25
SLIDE 25

VNIR/SWIR camera parameters

25 Parameters FX10 VNIR SWIR OLES30 Spectral range (nm) 400 - 1000 1000 - 2500 Spectral bands 224 288 Spectral FWHM (nm) 5.5 12 Spatial sampling 1024 384 FOV (degree) 38 17 Maximum frame rate (fps) 330 450 Exposure time range (ms) 0.1 – 20 0.1 – 20 Aperture 1.7 2 Focal length (mm) 15 30 Measurement distance (m) 0.118 0.316 Field of View (mm) 81.26 94.45 Spatial resolution (um) 79.36 245.97 Depth of Field (mm) 1.91 9.64

Thanh M. BUI

slide-26
SLIDE 26

Spectral classification using machine learning techniques

26 Spectra Pre-processing + Feature Extraction Training using SVM Spectrum Pre-processing + Feature Extraction Features Labels Features Trained Classifier Label

Training phase Prediction phase

Thanh M. BUI

slide-27
SLIDE 27

Spectral classification - results

27

Features Training (%) Testing (%) SAM 99.97 ± 0.07 99.56 ± 0.60 Filtered spectra 100 99.64 ± 0.55 Continuum removal 100 98.62 ± 0.84 Filtered spectra + PCA 99.97 ± 0.07 99.29 ± 0.78 Continuum removal + PCA 99.70 ± 0.28 98.02 ± 1.08

540 samples (randomly selecting 360 for training, 180 for testing); C-SVC, C = 100; RBF, γ = 0.999

Iteratively evaluated in 50 times

slide-28
SLIDE 28

Spectral classification - results

28 Hematite 2 Goethite Hematite 1 Magnetite 2 Chromite Chromite Chromite Chromite Chromite Magnetite 1 Hematite 2 Goethite Hematite 1 Olivine, Chromite Chromite Chromite Chromite Chromite Chromite Magnetite 1 66 mm

Thanh M. BUI

slide-29
SLIDE 29

Hyperspectral classification

Thanh M. BUI

2D conv. 2D conv. Fully connected layer

Logistic classification

5x5xL 3x3x64 1x1x64 128 16 (64, 3x3, 1) (64, 3x3, 1)

2D CNN

Dropout Dropout

L 29

slide-30
SLIDE 30

Hyperspectral classification

Thanh M. BUI

Random Forest:

n_estimators=500, max_features=15, bootstrap=False Accuracy: 0.888

Gradient boosting machines:

n_estimators=500, max_features=15 Accuracy: 0.892

2D CNN, Accuracy: 0.953 1D CNN, Accuracy: 0.926

30

Indian Pines scene, 224 bands from 0.4 – 2.5 um