EO signal models & feature [email protected]
Just an entree…
• Objective: getting the discussion started
• Because of this, presented in possibly more controversial / provocative terms than I really consider the topic to be
• More in–depth lesson by Mihai Datcu in the next days
Approaches to visionD. Marr 1982 “Vision”, Freeman NY:
- Feature-based
- Bottom-up
- Building representations out of signals
D.Ballard 1991 “Animate Vision” in “Artificial intelligence” 48, pp. 57-86
- Objective-based
- Well-posed “localized” problem
- Makes use of prior information
Mining as fusion
Approaches for IE
SRTM Baltimore, 25m
RSAT LasVegas, 15m Intermap Maastricht, 0.5m
XSAR Bern, 30m
Primitive features for optical dataA diverse gamut of content-based primitive feature extractors:
* pixel-based descriptors — colorimetry in multiple spaces
* texture [Franklin et al. Computers & Geosciences 1996]
* morphology, shapes [Benediktsson et al. IEEE TGRS 2003]
* object-oriented descriptors [Blaschke et al. ISPRS JPRS 2010]
* combinations thereof [Blume et al. AeroSense'97]
OPT: A 2012 “golden standard”For metric resolution classification [Chauffert IGARSS 2012]:
• HSV color-based descriptor
• Histograms of Oriented Gradients (HOG) [Pohl et al. IJRS 1998]
• Local Binary Patterns [Molinier IEEE TGRS 2007]
• Line Segment Detector [Molinier IEEE TGRS 2007]
• edge density [Inglada IGARSS 2012]
• SIFT [Perrotton Computer Vision Systems 2008]
Extensions• Morphological image analysis models [Soille1999]
• Synthetic Aperture Radar:
• Physical simulation models [Franceschetti2003]
• Bayesian structural models [Quartulli2004]
Bayesian model-based IE
– Existing models or explicit assumptions for p(S) – Hierarchies of models
( )( )Dp
SpSDpDpDSpDSp )()|()(),(| ==
Updated description
Existing description
New information
SAR: Bayesian information extractionGMRF +
Space Variant Γ(.) Likelihood
Backscatter intensity
Despeckled
backscatter
intensity
Backscatter
Intensity texture
normSRTM Baltimore, 25m
DEM: Bayesian information extractionGMRF + Space Variant Gaussian Likelihood
DEM elevation map
Elevation
texture
norm
Clean
elevation
map
SRTM Baltimore, 25m
Feature fusion for building reconstructionby resolving high-resolution detail
Baltimore, USA