Data Science for Planetary Science

Transcription

1 Data Science for Planetary Science Paris-Saclay Center for Data Science Frédéric Schmidt

2 Planetary Science Planetary formation From disk to (exo)planets Meteorites, comets

3 Planetary Science Planetary bodies Interior Surface Atmosphere Ionosphere

4 Planetary Science Planetary bodies Interior Surface Atmosphere Ionosphere Remote sensing : imagery

5 Big scientific questions Geologic evolution (tectonic, volcanic,...) Climate evolution (climate change, escape,...) Habitability (origin of life, human exploration)

6 Big scientific questions Geologic evolution (tectonic, volcanic,...) Climate evolution (climate change, escape,...) Habitability (origin of life, human exploration)??

7 Raw data Mars Express (ESA, launched in 2003) : ~50 Tb Mars Reconnaissance Orbiter (NASA, launched in 2005) : ~200 Tb Calibrated data Increase factor : ~10

8 Huge amount of data How to treat the data? How to represent the scientific results (in a global map)? Franquin, Gaston Lagaffe

9 Large volume products High resolution spectra High resolution images Hyperpectral images Multi-angular hyperspectral images

10 Large volume products High resolution spectra High resolution images Hyperpectral images Multi-angular hyperspectral images

11 Imaging techniques Usual Camera Pushbroom system up to pixels spacecraft motion

12 Examples of high resolution images datasets MOC (Mars Global Surveyor, NASA) HRSC (Mars Express, ESA) Neukum et al., 2004 ISS (Cassini, NASA) Porco et al., 2004 HiRISE (Mars Reconnaissance Orbiter, NASA)... Malin and Edgett, 2000 McEwen et al., 2007

13 Data Science Challenges for images Tools for large dataset treatment: 1. global scale mosaic visualisation 2. stereoscopy to create DEM 3. change detection (crater, dust devils, dune,...) 4. automatic feature identification

14 1. Scientific data visualisation Google Mars MapAPlanet JMars, Limitations: no scientific data not complete Slow

15 1. Data visualisation project Web based approach 3D and GIS oriented C. Marmo (GEOPS/IAS/ OSUPS)

16 2. Digital Elevation Model Stereoscopy based on image correlation Limitations: Very slow Uncertainties?

17 3. Change detection HiRISE (august 2011) Flow Summer (~30 S) Liquid water? McEwen et al., 2011

18 3. Change detection HiRISE (august 2011) Flow Summer (~30 S) Liquid water? McEwen et al., 2011

19 4. Feature detection Automatic crater counting on images on DEM Limitations: very slow accuracy Urbach et al., 2009 Stepinski et al., 2009

20 Data Science Challenges for images Data treatment (DEM) Data mining (change detection, feature identification) How to represent the data (global map, time)?

21 Large volume products High resolution spectra High resolution images Hyperpectral images Multi-angular hyperspectral images

22 Reflectance Reflectance Visible and Near-IR signal Contribution: atmosphere surface Atmospheric features Typical soil Wavelength (microns) Ices Wavelength (microns) CO2 H2O

23 Imaging spectrometer Hyperspectral image Maps surface/atmosphere properties > 100 wavelengths CO2 H2O

24 Examples of hyperspectral datasets OMEGA (Mars Express, ESA) Bibring et al., 2004 VIRTIS (Venus Express, ESA) Drossart et al., 2007 VIMS (Cassini, NASA) Brown et al., 2004 CRISM (Mars Reconnaissance Orbiter, NASA)... Murchie et al., 2007

25 Data Science Challenges for hyperspectral images Detection band ratios, wavelets, linear unmixing Quantification radiative transfer inversion

26 Data Science Challenges for hyperspectral images Detection band ratios, wavelets, linear unmixing Quantification radiative transfer inversion

27 Detection using Band ratio Very Fast Limitations: superposition of bands angular effects Pyroxene global map Bibring et al., 2006 Pelkey et al., 2007 Carter et al., 2014 Absorption depth Bibring et al., 2006

28 Detection using Wavelets WAVANGLET correlation in a wavelet coefficient subspace Schmidt et al., IEEE TGRS 2007 Fast and efficient to remove angular effect Limitations: ~10 endmembers Schmidt et al., Icarus 2009

29 Detection using Wavelets WAVANGLET correlation in a wavelet coefficient subspace Schmidt et al., IEEE TGRS 2007 Fast and efficient to remove angular effect Limitations: ~10 endmembers Schmidt et al., Icarus 2009

30 arried Marsresolution Express Thus, based on this model and using the ge bit, has by a spatial ture assumption, the radiance factor at loca Thus, based on this model and shas instrument has three chana spatial resolution at wavelenght λ satisfies the following obser near infrared channels. We ture assumption, the radiance fa trument has three chann the near infrared channels at wavelenght λ satisfies the foll L(x, y, λ) = ar infrared channels. We major chemicals can be dis#! P " e near nge. The infrared analysis ischannels focused (λ) + Φ(λ) α (x, y) ρ (λ) co ρ L(x, y, λ) = a p p Linear unmixing ngle hyperspectral data cube r chemicals can be dis! Combe et al., 2008 min α p.ρp L p=1 outh Polar Cap of Mars in P " The analysis is focused min αp.ρp L, αp > 0, αp = 1 under constraints ce, water ice and dust were where Φ(λ) isρathe spectral atmospheric Legendre et al., 2013 (λ) + Φ(λ) α (x, y p hyperspectral data cube θ(x, y) the angle between the solar directio data cube is made up with compensate non linearities p=1 s from 0.93 µm to 2.73 µm h Polar Cap of Mars inschmidt al., 2014 (solar incidence angle), P t face etnormal nd 128 spectral in the region of coordinates (x water ice andplanes dustfrom were endmembers where Φ(λ) is the spectral a ution of 0.020µm. After cal(x, y) spectrum of the p-th endmember, α p ta cube is made up with θ(x, y) the angle between the so parallel algorithm (GPU) hysical Highly unit used to express the mixture and ρa (λ) the radiation that om 0.93is the µmratio to between 2.73 µm directly, which facefrom normal incidence the area(solar under view. This mixa Limitations: el toward the sensor and from the 128 spectral planes alsoendmembers be written as: in the region of co.ninteractions between phoof 0.020µm. After calp ~50 endmembers spectrum of the p-th endmemb " the planet Mars, through its!! α (x, y) ρ (λ) + E( L(x, y, λ) = cal unit used to express p p the mixture and ρa (λ) the rad s us to identify the different p=1 hich Those is the compounds ratio between anet. are directly from the area under view d Detectioneusing Linear unmixing d t e p t e c p c e A c c OMEGA ORB422_4 Hypersthene PYX0.2H>250u (%) Oxide; Goethite (%) Diopside CPX CRISM (%) Ferrihydrite (%) Olivine Fayalite CRISM (%) Phyll; Chlorite (%) Olivine Forsterite CRISM (%) RMS Phyll; Clay Illite (%) Sum (%) emical species can be identi- where Figure 8: Detection of 8 minerals over 44 spectra on OMEGA image ORB422_4 of Syrtis Major using IPLS in the hue-saturation-value color system. The hue (color) represents the mixing coefficient. The saturation (color or b/w) represents the error. The value (intensity of color or b/w) represents the rms. Spectral mixing coefficient map are shown with following

31 Data Science Challenges for hyperspectral images Detection band ratios, wavelets, linear unmixing Quantification radiative transfer inversion

32 Spectral shape = physical state Grain size Free mean path 100 µm 10 mm Douté, et al, JGR, 1998 Schmitt, et al, Solar System Ice, 1998

33 Inversion using Least square Minimisation technique Surface Poulet et al., 2009 Atmosphere Wolff et al., 2009 Limitations Slow Multiples solutions

34 Inversion using Linear Subspace Look up table Douté et al., LPSC, 2007 GRSIR Bernard-Michel et al., Statistic and computing, 2009 Bernard-Michel et al., JGR, 2009 Projection into a linear subspace Very fast Limitations: Non linearities Multiple solutions CO2 ice grain size

35 Bayesian Inversion Monte Carlo inversion on photometry Limitations: Computation time Ceamanos et al., 2013 Fernando, J. et al., 2013 Maximum likelihood inversion Andrieu, F. et al., in preparation

36 Data Science Challenges for hyperspectral images Radiative transfer inversion (bayesian technique) estimation of surface/atmospheric properties How to represent the data (global map, wavelength, time)?

37 Conclusion Planetary Science (and Geoscience) needs Data Science revolution Data Mining Data visualisation Massive data treatment Virtual Observatory