Mineral interpretation results using deep learning with hyperspectral imagery - PowerPoint Presentation

1. Mineral interpretation results using deep learning with hyperspectral imageryAndrés Bell NavasCarlos Roberto del Blanco AdánFernando Jaureguizar NúñezNarciso García SantosMaría José Jurado RodríguezGrupo de Tratamiento de Imágenes (GTI)Universidad Politécnica de Madrid

2. Table of contentsIntroduction.Mineral recognition system for hyperspectral images of drill-core boxes.Results.Using LWIR spectral range.Using VN-SWIR spectral range.Using both spectral ranges.System conclusions.Final conclusions.

3. IntroductionInnolog (European Research Project):Development of innovative borehole geophysical logging tools and mineral interpretation software:Innovation in tools for wider applicability and sustainability.New acquisition sensors.Main contribution: Innovation in interpretation software for an improved in-situ and online recognition and quantification of minerals in the subsurface.Use of machine learning methods. Test and validate efficiency of new tools in mines.Training in the use of the developed technologies.Raise competitiveness of European mining companies.

4. Hyperspectral database of drill-core boxes: Problems for using it with machine learning techniques:Different types of sensors with:Diverse spectral and spatial resolutions.Different spectral ranges.Distortions in the spectral signatures by:Acquisition method.Geometry of the sensing device, illumination and the cores.Scattering properties of the core materials.Presence of regions without cores (holes and wooden box areas).Mineral recognition system for hyperspectral images of drill-core boxes(1)xyz(2)(3)(1)(2)(3)Core

5. System overview: General features:Deep neural network model adapted to hyperspectral imagery.Prediction: segmentation mask showing the minerals present in the input.Mineral recognition system for hyperspectral images of drill-core boxesMachine-learning compatible databasePredictionMineral classificationHyperspectral database of drill-core boxesTrainingDeep neural network model

6. ResultsDatabases features:Three scenarios:LWIR spectral range.VN-SWIR spectral range.LWIR and VN-SWIR spectral ranges.Spectral RangeRange of wavelengths (nm)Number of bandsLWIR7500 - 12000~90VN-SWIR450 - 2500~340

7. ResultsRecognition results (top 5):Different recognition results depending on:Spectral range.Number of samples per mineral category: imbalanced.MineralsF1-scoreAlbite0,90Amphibole0,97Apatite0,70Carbonate0,96Clinopyroxene0,99Epidote0,98Microcline0,92Quartz0,94Quartz-Clay-Feldspar0,96Sulphide-Oxide0,98Average / total0,93Using VN-SWIR spectral rangeUsing LWIR spectral rangeMineralsF1-scoreActinolite0,77Amphibole0,91Biotite-Chlorite0,82Carbonate0,97Epidote0,74Saponite0,90White Mica0,79White Mica-Chlorite0,95Average / total0,86

8. ResultsRecognition results using both (LWIR and VN-SWIR) spectral ranges:Strong imbalance in the number of samples (especially in Biotite for LWIR).Higher recognition performance with higher number of samples.F1-scoreNumber of samples (pixels)BandMineralVN-SWIRLWIRVN-SWIRLWIRAmphibole0,920,951.733.048917.039Biotite0,720,14173.644Carbonate0,950,952.833.2191.347.898Epidote0,710,9983.113329.661Water-Phyllosilicate0,860,87193.488532.467White Mica0,830,95404.32742.830Average (with Biotite)0,830,81Average (without Biotite)0,850,9413.776Very few samples

9. ResultsMineral classification performance (confusion matrix – top 5):LWIR band.BackgroundClass 1AlbiteQuartzApatiteBiotiteMiscellaneousCarbonateEpidoteMicroclineAmphiboleOrthoclaseGarnet (Ugrandite)White MicaClinopyroxeneOlivinePlagioclaseGypsumQuartz-Clay-FeldsparApatite-OtherOther-CarbonateWater-PhyllosilicateFeldspar-QuartzSulphide-OxidePhyllosilicateTourmalinePredicted mineralTrue mineral0,990,000,000,010,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,010,960,000,000,000,000,000,000,000,000,000,000,000,010,000,000,000,000,000,000,000,000,000,000,000,000,010,010,840,000,000,000,010,000,000,010,000,000,000,000,000,000,000,000,000,000,000,010,090,000,010,000,000,000,001,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,990,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,030,020,000,030,010,770,000,040,000,000,020,000,000,080,010,000,000,000,000,000,000,000,000,000,000,000,020,010,010,010,010,000,930,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,010,000,000,010,000,000,010,930,000,000,000,000,000,000,000,010,000,000,000,000,000,010,000,000,010,000,000,000,000,020,000,000,000,000,970,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,020,000,010,020,000,000,000,000,000,860,000,000,000,000,000,000,000,000,040,000,000,030,010,000,000,000,010,000,000,000,000,000,000,000,000,000,980,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,020,000,000,010,000,000,000,000,000,000,020,930,000,000,000,000,000,000,010,000,000,010,000,000,000,000,120,000,000,070,000,000,000,010,000,000,000,000,800,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,980,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,990,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,001,000,000,000,000,000,000,000,000,000,000,000,000,010,000,000,000,000,000,010,000,000,010,010,000,000,000,000,000,960,000,000,000,000,000,000,000,000,000,010,000,000,020,000,000,000,000,000,000,000,000,000,000,000,000,000,960,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,990,000,000,000,000,000,000,000,070,030,010,070,190,010,000,000,010,010,010,000,000,000,000,000,000,000,000,580,000,000,000,010,000,000,020,000,000,020,000,000,000,010,000,000,010,000,000,000,010,000,000,000,000,000,920,000,000,020,000,000,010,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,980,000,000,000,000,010,000,020,000,000,000,000,000,000,010,000,000,000,000,000,000,000,000,000,000,000,020,940,000,000,000,010,000,000,000,000,000,000,000,000,000,010,000,000,000,000,000,000,000,000,000,000,000,000,980,000,000,020,010,000,000,010,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,920,000,190,030,000,140,560,050,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,04

10. ResultsMineral classification performance (confusion matrix – top 5):VN-SWIR band.0,990,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,980,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,010,000,000,000,000,000,000,960,000,010,000,010,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,010,010,970,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,010,000,020,000,910,020,000,000,000,000,000,010,000,000,000,010,000,000,000,000,000,000,000,010,000,000,020,960,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,040,030,050,050,040,030,720,010,010,000,000,010,000,000,000,000,000,010,000,000,000,000,010,000,150,010,020,010,000,700,000,000,000,030,000,000,010,060,000,000,000,000,000,000,030,010,150,040,020,010,050,000,590,000,010,000,000,000,000,000,000,040,010,000,000,000,100,050,010,010,010,020,010,000,070,690,000,000,000,000,000,000,000,010,000,000,000,000,010,030,060,100,010,010,000,050,000,000,690,010,000,000,000,000,000,000,000,020,000,000,020,000,010,020,000,020,000,000,010,000,000,910,000,000,000,000,000,000,000,000,000,000,020,040,000,040,040,370,100,020,040,000,000,020,300,000,000,000,000,000,000,000,000,000,040,000,060,000,000,000,000,000,000,000,000,000,000,820,000,000,000,050,010,000,010,000,050,040,060,000,040,000,000,000,000,000,000,000,000,000,790,000,000,010,000,010,000,000,020,000,170,000,020,000,000,010,000,000,000,000,000,000,000,770,000,000,000,000,000,000,030,000,070,000,000,010,000,000,000,000,010,000,000,030,020,040,760,020,010,000,000,000,020,030,030,000,000,000,000,000,000,000,000,000,000,000,000,010,000,890,000,000,000,000,040,080,090,010,010,000,030,000,000,000,000,000,000,000,020,010,000,000,680,000,010,000,010,000,010,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,970,000,000,060,020,110,050,010,000,010,000,000,000,000,000,000,020,010,000,000,000,010,000,690,000,100,000,130,050,000,000,150,000,130,000,000,000,000,010,010,000,000,000,020,000,010,40BackgroundMiscellaneousAmphiboleCarbonateClass 4White Mica-ChloriteWhite MicaBiotiteWaterWoodEpidoteSaponiteTalcQuartz-WaterClass 16Biotite-ChloriteChlorite-Biotite-SericiteWater-PhyllosilicateActinoliteDark PhyllosilicateTourmalineOther-WaterPredicted mineralTrue mineral

11. ResultsExamples of problematic spectral signatures:LWIR band:VN-SWIR band:Wavelength (nm)ReflectanceClinopyroxeneWavelength (nm)ReflectanceOlivineWavelength (nm)ReflectanceTalcWavelength (nm)ReflectanceWhite Mica-Chlorite100 % classified as Clinopyroxene37 % classified as White Mica-Chlorite

12. System conclusionsStrong imbalance in the number of samples of the mineral categories:Possible measures to deal with the problem:Gather more samples (time consuming and costly).Apply specific machine learning techniques for scarce data.Discard categories with few samples or group together into the same class.Recognition in the LWIR range more accurate than in VN-SWIR:More variety of minerals.Spectral signature patterns may be more distinctive.VN-SWIR band still useful:Recognition of several mineral categories only possible in this band.Spectral signature patterns in this band may be still discriminative enough.Very interesting option for low-budget systems.

13. Final conclusionsAutomatic mineral recognition from hyperspectral imagery is possible:Successful application of innovative machine learning methods (patent in process).Development of a mineral recognition system for hyperspectral images of drill-core boxes:Better recognition capability for those mineral categories with a higher number of samples.Real-time operation:Viable (less than one second per prediction), necessary for recognition during downhole explorations.

14. Mineral interpretation results using deep learning with hyperspectral imageryAndrés Bell NavasCarlos Roberto del Blanco AdánFernando Jaureguizar NúñezNarciso García SantosMaría José Jurado RodríguezGrupo de Tratamiento de Imágenes (GTI)Universidad Politécnica de Madrid