9. High-level processing (astronomical data analysis)lsiit-miv.u-strasbg.fr/paseo/slides/astroproc_9_high.pdf · High-level processing (astronomical data analysis) ... Master ISTI

Introduction to Astronomical Image Processing

9. High-level processing(astronomical data analysis)

André JalobeanuLSIIT / MIV / PASEO group

Jan. 2006

lsiit-miv.u-strasbg.fr/paseo

Master ISTI / PARI / IV

PASEO

High-level processing:astronomical data analysis

Supervised analysisPixel classification/segmentationSearch for particular elements or structures

Object classificationStar classification, Galaxy classification

Physical parameter estimationPhysical parameter field estimationVelocities, density, ionization, element abundances

Object modelsPN shape and expansion rates, density models

Data mining Star mapping & astrometric reductionSky surveys, unknown object detection

Unsupervised pixel classification/segmentationUnsupervised object classification

Supervised analysis

Recall the principles of supervised data classification and image segmentation

See some examples of object classification (galaxies, stars)

Understand the differences between spatial and spectral classification

Focus on labeling rather than physical interpretation

Supervised classification and segmentation

๏ Data classificationRandom samples (no spatial organization)Data space → feature space → classes

‣ Hard classification: 1 label/element, binary‣ Soft classification: N labels/element, binary‣ Fuzzy classification: N membership fcts/element ∈ [0,1]

๏ Image segmentationDecompose the image into connected regions(common properties: color, texture...)Classification with spatial consistency

N classes,known parameters or models (training)

Segmentation boundaries(multiband nebula image)

Fuzz

yH

ard

Feature space

Probabilistic vs. variational approachesCommon framework: optimization

Spectral element identification

๏ Hyperspectral images (IFS)

‣ Use spectral signatures of elements:catalog of known signatures (line positions)‣ Process each spatial sample (spaxel) independently as a spectrum

๏ Multispectral images

‣ Transform known signatures (integration):look for particular “colors” (e.g. H→red, O→green, dust→blue)

Maximum or minimum finding (1D signal)Statistical fitting (regression): more accurate

Non-orthogonal change of basis (projections)

Pixel classification/segmentation

๏ Training step

‣ Use catalog of spectral signatures‣ Use expert knowledge to manually assignlabels to test images‣ Find optimal feature spaces

๏ Classification / segmentation

‣ Classify or segment, depending onspatial dependencies and regularization

Training step for supervisedimage classification in rockcomposition analysis (X. Shi)A

stro

bio

logy

at

ASU

Object classification (morphological)

๏ Extract sources

๏ Use morphological characteristics to classify objects

‣ PSF-shaped objects: stars‣ Bigger, elliptical or round objects: galaxies (different types)‣ Elongated tracks: cosmic rays, asteroids, space debris‣ Diffuse objects: gas nebulae or galaxies‣ ...

Examples of morphological indicators in galaxy classification:surface brightness and concentration index [Odewahn 95]

Feature extraction(math. morphology andparameter estimation)

Classifier training needed

Object classification (spectral)

๏ Use spectral informationSpatial information might be insufficient or difficult to useUse color indices extracted from bands or spectra(e.g. B-V, U-V, R-I where I=infrared, R=red, V=green&yellow, B=blue, U=ultraviolet)‣ Predefined spectral types for stars‣ Loose correspondence between morphological types and color

indices in galaxy classification

Example of spectral indicatorin galaxy classification:

color (O-E bands) [Odewahn 95]

Star classes,spectral typesand luminosity(HR diagram)

measure L(photometry),measure T(spectrum fittingor color index X-Yfor multispectral)

assign a class

Physical parameter estimation

Grasp the basic ideas of physical parameter measurement

Focus on quantitative physical interpretation of astronomical data

Discover the advantages of 3D object modeling

Physical parameter estimation from spectra

Statistical fitting (regression)

e.g. Voigt profile

continuum:Star surfacetemperature

line profile:Gas pressure

line position:element IDshifted line position:

object velocity

line intensity:gas density

synchrotronradiation,

etc.line fitting, spectrum model fitting

Physical parameter fieldsspatial distribution from IFS or data cubes

๏ For each pixel (spectrum):estimate physical parameters independently

๏ Spatial consistency: Use spatial priors (e.g. smoothness)Constrain physical parameter estimation to stabilize the solution

Regularization: variational, probabilistic (MRF)Optimization methods

Velocity field of NGC 4486 (F. Eisenhauer)

Metallicity index (NGC 3377)

Mean velocity(M32)

(P.T. de Zeeuw)

NGC 4254 velocity fieldusing multiscale regularization

via Hidden Markov Trees(M. Petremand)

Physical parameter fields Kinematics and line strength distributions examples

Galaxy formation observed with Sauron (M. Cappellari)

3D/2D object models

๏ 3D object modeling

‣ Physical models of astronomical objects (e.g. planetary nebulae)density, shape, expansion rate...‣ Stronger constraints than 2D modelssome parameters can be correctly determined only using 3D models: estimation from both image and spectrum

๏ Parameter estimation

‣ Directly in 3D: difficult‣ Multigrid optimization: nested 2D models to initialize 3D optimization

Planetary Nebula parameter estimation,ambiguity removal using long-slit spectra

(A.R. Hajian et al.)

Data mining

See some modern approaches to source extraction & measurement

Grasp some ideas about automated classification and segmentation

Get acquainted to information discovery from large data sets

Astrometric reduction for star-like objects

๏ Measure star positions and magnitudes (build maps)

‣ First extract the stars

•Existing software, e.g. Sextractor

•Extraction algorithm: morphological cleaning & segmentation,or template matching with PSF, maxima finding (recursive: CLEAN)‣ Measure the position

•Compute the centroid or fit an analytic profile‣ Compute the magnitude

•Integrate over 2 circles (star and sky background) or fit analytic profile‣ Compare with existing catalogs for identification

“Sextractor”: source extraction

๏ Automatic source extraction and analysis

‣ Very low level corrections‣ Filtering and segmentation‣ De-blending (spatial source separation)‣ Photometry and astrometry,basic shape parameter computation‣ Cross-identification using catalogs

De-blending is necessaryfor spatially overlapping sources(multiple maxima for one region)

[Bertin 95]

Sextractor shape parameters and layout

Examples of shape parameters and momentscomputed for each detected source

Unsupervised pixel classification/segmentation

๏ Find class parameters & number of classes

‣ Learning process: the most difficult part!‣ Sensitive to the number of bands (curse of dimensionality)

๏ Assign a label to each pixel

๏ Use spatial consistency priors(region-based segmentation)

Multiscale segmentation of the SMC using Hidden Markov Trees

and taking into account missing data

(Flitti et al.)

Unsupervised object classification

๏ Example: AutoClass (Hanson-Stutz-Cheeseman, NASA Ames)

‣ Bayesian learning theory:

•automatic parameter estimation (e.g. class parameters)

•automatic model selection (e.g. number of classes)

Autoclass discovery in the IRAS star atlas(subtle differences btw. infrared spectra) → 2 subgroups

Bayesian re-classificationof the IRAS LRS Atlas [Goebel 89]

Probability theory & optimization

Content-based image retrieval

๏ Process huge catalogs (>50M galaxies for SLOAN catalog)

‣ Goal: facilitate search operations‣ Indexing is needed: feature extractioneach object is located in a multidimensional feature spaceAvoid processing images for each query!

Galaxy retrieval system[Calleja et al.]

Preprocessing for indexing:crop, center, PCA

[Calleja et al.]

ZCAT redshift catalog(distribution on the sky)

Further reading

๏ Galaxies and the universe (course notes)http://www.astr.ua.edu/keel/galaxies/index.html

๏ Supervised classification (remote sensing)http://rst.gsfc.nasa.gov/Sect1/Sect1_17.html

๏ SAURON IFS - instrument and image exampleshttp://www.strw.leidenuniv.nl/sauron/

๏ Astrometric reduction from the IRIS tutorialhttp://www.astrosurf.org/buil/iris/tutorial13/doc31_us.htm

๏ Sextractorhttp://terapix.iap.fr/rubrique.php?id_rubrique=91/

๏ Astrostatistics Tutorials SAMSI 2006http://www.astrostatistics.psu.edu/samsi06/index.html

๏ Astrostatistics Course Notes SU 2005http://www.astrostatistics.psu.edu/su05/astrostat_courses.html

๏ Selected articles in statistics for astronomy & physicshttp://www.astrostatistics.psu.edu/castbib/Bib_arts.html