unsupervised oil spill detection in sar imagery through an...

29
Remote Sensing Lab (RSLab) Dept. of Signal Theory & Communications Mariv Mariv í í Tello Tello , Carlos , Carlos L L ó ó pez pez - - Mart Mart í í nez nez , Jordi J. , Jordi J. Mallorqu Mallorqu í í . . Remote Sensing Laboratory ( Remote Sensing Laboratory ( RSLab RSLab ) ) Signal Theory and Telecommunications Department Universitat Politècnica de Catalunya Barcelona, SPAIN Unsupervised Oil Spill Detection in SAR Unsupervised Oil Spill Detection in SAR Imagery through an Estimator of Local Imagery through an Estimator of Local Regularity Regularity Alessandro Alessandro Danisi Danisi , Gerardo Di Martino, , Gerardo Di Martino, Antonio Antonio Iodice Iodice , Giuseppe , Giuseppe Ruello Ruello , Daniele , Daniele Riccio Riccio . . Department of Electronic and Department of Electronic and Telecommunication Engineering Telecommunication Engineering Università di Napoli “Federico II” Napoli, ITALY SEASAR SEASAR Workshop Workshop 2008 2008

Upload: others

Post on 26-Sep-2020

3 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Unsupervised Oil Spill Detection in SAR Imagery through an …earth.esa.int/.../participants/112/pres_112_tello.pdf · 2018. 5. 15. · Mariví Tello, Carlos López-Martínez, Jordi

Remote Sensing Lab (RSLab)Dept. of Signal Theory & Communications

MarivMarivíí TelloTello, Carlos , Carlos LLóópezpez--MartMartííneznez, Jordi J. , Jordi J. MallorquMallorquíí. .

Remote Sensing Laboratory (Remote Sensing Laboratory (RSLabRSLab))Signal Theory and Telecommunications Department

Universitat Politècnica de CatalunyaBarcelona, SPAIN

Unsupervised Oil Spill Detection in SAR Unsupervised Oil Spill Detection in SAR Imagery through an Estimator of Local Imagery through an Estimator of Local

RegularityRegularity

Alessandro Alessandro DanisiDanisi, Gerardo Di Martino, , Gerardo Di Martino, Antonio Antonio IodiceIodice, Giuseppe , Giuseppe RuelloRuello, Daniele , Daniele

RiccioRiccio. .

Department of Electronic and Department of Electronic and Telecommunication EngineeringTelecommunication Engineering

Università di Napoli “Federico II”Napoli, ITALY

SEASAR SEASAR WorkshopWorkshop 20082008

Page 2: Unsupervised Oil Spill Detection in SAR Imagery through an …earth.esa.int/.../participants/112/pres_112_tello.pdf · 2018. 5. 15. · Mariví Tello, Carlos López-Martínez, Jordi

Remote Sensing Lab (RSLab)Dept. of Signal Theory & Communications

IntroductionIntroductionMost often, interpretation of SAR images is performed manually: slow, unpractical and hardly reproducible procedure: computerised schemes are desirable.

- Our objective is to develop methods, specifically conceived to deal with the characteristic properties of SAR imagery

- Extract as much information as possible, automatically, from every single image, single polarization, with no use of external auxiliary data

Barcelona

Barcelona, ERS image, PRI, March 97.

SEASAR SEASAR WorkshopWorkshop 20082008

Page 3: Unsupervised Oil Spill Detection in SAR Imagery through an …earth.esa.int/.../participants/112/pres_112_tello.pdf · 2018. 5. 15. · Mariví Tello, Carlos López-Martínez, Jordi

Remote Sensing Lab (RSLab)Dept. of Signal Theory & Communications

• unlike optical imagery, interpretation of radar images is not consistent with common visual perception

• most of the tools of image processing are conceived from an “optical” point of view

SomeSome preliminarypreliminary considerationsconsiderations (I)(I)

Our purpose is to establish a specific framework for the automatic exploitation of

SAR imagery.

Due to speckle, a SAR image is one realization

of an underlying stochastic non-homogeneous

process.

SEASAR SEASAR WorkshopWorkshop 20082008

Page 4: Unsupervised Oil Spill Detection in SAR Imagery through an …earth.esa.int/.../participants/112/pres_112_tello.pdf · 2018. 5. 15. · Mariví Tello, Carlos López-Martínez, Jordi

Remote Sensing Lab (RSLab)Dept. of Signal Theory & Communications

( ) ( ) ( ) ( ) ( )0 0, , , , ,ii

u x r x r u x r x r u x rγ γ= ⊗ = ⊗∑

The SAR image can be modelled as the convolution of the local complex reflectivity of the observed area with the impulse response of the SAR system.

Random sum of the contributions of all the scatterers within a resolution cell

(random walk process).

Analysis tools have to be inscribed in a statistical framework, but preserving

contextual information.

SAR images are spiky, with a large dynamic range and they involve non-stationary

processes.

A SAR image is one realization of an underlying stochastic non-stationary process.

SomeSome preliminarypreliminary considerationsconsiderations (II)(II)

SEASAR SEASAR WorkshopWorkshop 20082008

Page 5: Unsupervised Oil Spill Detection in SAR Imagery through an …earth.esa.int/.../participants/112/pres_112_tello.pdf · 2018. 5. 15. · Mariví Tello, Carlos López-Martínez, Jordi

Remote Sensing Lab (RSLab)Dept. of Signal Theory & Communications

The brain attaches high information content to vertices, linear structures and edges.

Inspired on the human vision system, our workplan to achieve a specific framework for the automatic interpretation of SAR imagery is:

• Spot detection, directly applied to ship detection.

• Extraction of linear features, directly applied to coastline extraction.

• Texture analysis, applied to oil spill detection.

SEASAR SEASAR WorkshopWorkshop 20082008

Page 6: Unsupervised Oil Spill Detection in SAR Imagery through an …earth.esa.int/.../participants/112/pres_112_tello.pdf · 2018. 5. 15. · Mariví Tello, Carlos López-Martínez, Jordi

Remote Sensing Lab (RSLab)Dept. of Signal Theory & Communications

The brain attaches high information content to vertices, linear structures and edges.

Inspired on the human vision system, our workplan to achieve a specific framework for the automatic interpretation of SAR imagery is:

• Spot detection, directly applied to ship detection.

• Extraction of linear features, directly applied to coastline extraction.

• Texture analysis, applied to oil spill detection.

SEASAR SEASAR WorkshopWorkshop 20082008

Page 7: Unsupervised Oil Spill Detection in SAR Imagery through an …earth.esa.int/.../participants/112/pres_112_tello.pdf · 2018. 5. 15. · Mariví Tello, Carlos López-Martínez, Jordi

Remote Sensing Lab (RSLab)Dept. of Signal Theory & Communications

The proposed algorithm faces the detection not only taking exclusively into account the intensity characteristics of the image but also studying its very localized statistical behaviour.

* Dire

ct re

sult,

no th

resh

old a

pplie

d

target

OCWT XX

Input image Output image *- background noise reduced because the OCWT is sparse- discontinuities target – background enhanced in each direction separately

Situation not resolvable by a CFAR approach !

Horiz

onta

l pro

file

Hist

ogra

m

Horiz

onta

l pro

file

Situation solved by the proposed algorithm !

Hist

ogra

m target

Region in which a threshold would provide a correct detection (target detected with no false alarms). As a consequence, larger coloured region represents a higher detectability rate.

AutomaticAutomatic Spot Spot DetectionDetection

SEASAR SEASAR WorkshopWorkshop 20082008

Page 8: Unsupervised Oil Spill Detection in SAR Imagery through an …earth.esa.int/.../participants/112/pres_112_tello.pdf · 2018. 5. 15. · Mariví Tello, Carlos López-Martínez, Jordi

Remote Sensing Lab (RSLab)Dept. of Signal Theory & Communications

The brain attaches high information content to vertices, linear structures and edges.

Inspired on the human vision system, our workplan to achieve a specific framework for the automatic interpretation of SAR imagery is:

• Spot detection, directly applied to ship detection.

• Extraction of linear features, directly applied to coastline extraction.

• Texture analysis, applied to oil spill detection.

SEASAR SEASAR WorkshopWorkshop 20082008

Page 9: Unsupervised Oil Spill Detection in SAR Imagery through an …earth.esa.int/.../participants/112/pres_112_tello.pdf · 2018. 5. 15. · Mariví Tello, Carlos López-Martínez, Jordi

Remote Sensing Lab (RSLab)Dept. of Signal Theory & Communications

AutomaticAutomatic extractionextraction ofof linear linear featuresfeaturesENVISAT image Sobel filter result Proposed algorithm *

RADARSAT image

* Dire

ctre

sult

–no

tresh

old

appl

ied

SEASAR SEASAR WorkshopWorkshop 20082008

Page 10: Unsupervised Oil Spill Detection in SAR Imagery through an …earth.esa.int/.../participants/112/pres_112_tello.pdf · 2018. 5. 15. · Mariví Tello, Carlos López-Martínez, Jordi

Remote Sensing Lab (RSLab)Dept. of Signal Theory & Communications

AutomaticAutomatic extractionextraction ofof linear linear featuresfeatures

Rivers and inland waters. Oil spills.

ENVISAT image ENVISAT imageResult * Result *

* Dire

ctre

sult

–no

tresh

old

appl

ied

SEASAR SEASAR WorkshopWorkshop 20082008

Page 11: Unsupervised Oil Spill Detection in SAR Imagery through an …earth.esa.int/.../participants/112/pres_112_tello.pdf · 2018. 5. 15. · Mariví Tello, Carlos López-Martínez, Jordi

Remote Sensing Lab (RSLab)Dept. of Signal Theory & Communications

The brain attaches high information content to vertices, linear structures and edges.

Inspired on the human vision system, our workplan to achieve a specific framework for the automatic interpretation of SAR imagery is:

• Spot detection, directly applied to ship detection.

• Extraction of linear features, directly applied to coastline extraction.

• Texture analysis, applied to oil spill detection.

SEASAR SEASAR WorkshopWorkshop 20082008

Page 12: Unsupervised Oil Spill Detection in SAR Imagery through an …earth.esa.int/.../participants/112/pres_112_tello.pdf · 2018. 5. 15. · Mariví Tello, Carlos López-Martínez, Jordi

Remote Sensing Lab (RSLab)Dept. of Signal Theory & Communications

Why an estimator local roughness in SAR images?Why an estimator local roughness in SAR images?- The appearance of oil spills is subject to a great diversity: assumption of a priori models is not efficient, training of algorithms based on neural networks are time consuming, algorithms exclusively based on morphological features are not robust…- Oil spills and look-alikes can present remarkable similarities.

Examples of oil spills Examples of look-alikes

SEASAR SEASAR WorkshopWorkshop 20082008

Page 13: Unsupervised Oil Spill Detection in SAR Imagery through an …earth.esa.int/.../participants/112/pres_112_tello.pdf · 2018. 5. 15. · Mariví Tello, Carlos López-Martínez, Jordi

Remote Sensing Lab (RSLab)Dept. of Signal Theory & Communications

The roughness of the sea is the local distribution of its height.

The roughness in the SAR image is the spatial variability of grey level in the neighbourhood of each pixel.

Our objective is to provide a quantitative measure as local as possible of the regularity of the SAR signal.

SEASAR SEASAR WorkshopWorkshop 20082008

Why an estimator local roughness in SAR images?Why an estimator local roughness in SAR images?

Page 14: Unsupervised Oil Spill Detection in SAR Imagery through an …earth.esa.int/.../participants/112/pres_112_tello.pdf · 2018. 5. 15. · Mariví Tello, Carlos López-Martínez, Jordi

Remote Sensing Lab (RSLab)Dept. of Signal Theory & Communications

The roughness of the sea is the local distribution of its height.

The roughness in the SAR image is the spatial variability of grey level in the neighbourhood of each pixel.

Our objective is to provide a quantitative measure as local as possible of the regularity of the SAR signal.

SEASAR SEASAR WorkshopWorkshop 20082008

Why an estimator local roughness in SAR images?Why an estimator local roughness in SAR images?

Page 15: Unsupervised Oil Spill Detection in SAR Imagery through an …earth.esa.int/.../participants/112/pres_112_tello.pdf · 2018. 5. 15. · Mariví Tello, Carlos López-Martínez, Jordi

Remote Sensing Lab (RSLab)Dept. of Signal Theory & Communications

The roughness of the sea is the local distribution of its height.

The roughness in the SAR image is the spatial variability of grey level in the neighbourhood of each pixel.

Our objective is to provide a quantitative measure as local as possible of the regularity of the SAR signal.

SEASAR SEASAR WorkshopWorkshop 20082008

Why an estimator local roughness in SAR images?Why an estimator local roughness in SAR images?

Page 16: Unsupervised Oil Spill Detection in SAR Imagery through an …earth.esa.int/.../participants/112/pres_112_tello.pdf · 2018. 5. 15. · Mariví Tello, Carlos López-Martínez, Jordi

Remote Sensing Lab (RSLab)Dept. of Signal Theory & Communications

The regularity of a single isolated pixel is nonsense. It depends on its intensity value relative to that of its neighbours.

Multiscale concept

SEASAR SEASAR WorkshopWorkshop 20082008

How to estimate local roughness in SAR images?How to estimate local roughness in SAR images?

Page 17: Unsupervised Oil Spill Detection in SAR Imagery through an …earth.esa.int/.../participants/112/pres_112_tello.pdf · 2018. 5. 15. · Mariví Tello, Carlos López-Martínez, Jordi

Remote Sensing Lab (RSLab)Dept. of Signal Theory & Communications

The regularity of a single isolated pixel is nonsense. It depends on its intensity value relative to that of its neighbours.

Multiscale concept

SEASAR SEASAR WorkshopWorkshop 20082008

How to estimate local roughness in SAR images?How to estimate local roughness in SAR images?

Page 18: Unsupervised Oil Spill Detection in SAR Imagery through an …earth.esa.int/.../participants/112/pres_112_tello.pdf · 2018. 5. 15. · Mariví Tello, Carlos López-Martínez, Jordi

Remote Sensing Lab (RSLab)Dept. of Signal Theory & Communications

By following the scale to scale energyvariations the local regularity can be infered.

The regularity of a single isolated pixel is nonsense. It depends on its intensity value relative to that of its neighbours.

Multiscale concept

SEASAR SEASAR WorkshopWorkshop 20082008

How to estimate local roughness in SAR images?How to estimate local roughness in SAR images?

Page 19: Unsupervised Oil Spill Detection in SAR Imagery through an …earth.esa.int/.../participants/112/pres_112_tello.pdf · 2018. 5. 15. · Mariví Tello, Carlos López-Martínez, Jordi

Remote Sensing Lab (RSLab)Dept. of Signal Theory & Communications

The decay of the Wavelet Transform amplitude across scales is related to the uniform and pointwise Lipschitz regularity of the signal.

Projection of the pointwise evolution across scales (obtained from a WT) of a cut of a homogeneous sea area.

Homogeneous decay

Projection of the pointwise evolution across scales (obtained from a WT) of a cut intercepting an oil spill.

The Lipschitz or Hölder exponent at a point is the maximum slope of log2|Wf(u,s)| as a function of log2s along the maxima lines converging to that point.

Different decays

SEASAR SEASAR WorkshopWorkshop 20082008

How to estimate local roughness in SAR images?How to estimate local roughness in SAR images?

Page 20: Unsupervised Oil Spill Detection in SAR Imagery through an …earth.esa.int/.../participants/112/pres_112_tello.pdf · 2018. 5. 15. · Mariví Tello, Carlos López-Martínez, Jordi

Remote Sensing Lab (RSLab)Dept. of Signal Theory & Communications

Combined Estimation

Horizontal roughness

Vertical roughness

Diagonalroughness

SAR image DWT2D Vertical components (HL)

Horizontal components (LH)

Diagonal components (LL)

Hα Vα Dα

How to estimate local roughness in SAR images?How to estimate local roughness in SAR images?

Flowchart of the proposed algorithm

Page 21: Unsupervised Oil Spill Detection in SAR Imagery through an …earth.esa.int/.../participants/112/pres_112_tello.pdf · 2018. 5. 15. · Mariví Tello, Carlos López-Martínez, Jordi

Remote Sensing Lab (RSLab)Dept. of Signal Theory & Communications

In order to validate the algorithm, it is first run on a simulated surface.Hö

ldere

xpon

ent

Multif

racti

onal

Brow

nian m

otion Hölder exponent retrieved

SEASAR SEASAR WorkshopWorkshop 20082008

ResultsResults

Page 22: Unsupervised Oil Spill Detection in SAR Imagery through an …earth.esa.int/.../participants/112/pres_112_tello.pdf · 2018. 5. 15. · Mariví Tello, Carlos López-Martínez, Jordi

Remote Sensing Lab (RSLab)Dept. of Signal Theory & Communications

The algorithm provides an estimate of the local regularity which is independent from the mean value.Si

mula

ted

spec

kleim

age

A

A*5 Loca

l reg

ularit

yret

rieve

d Despite thedifference of mean

intensity of theinput, the output matrix is exactly

the same.

SEASAR SEASAR WorkshopWorkshop 20082008

ResultsResults

Page 23: Unsupervised Oil Spill Detection in SAR Imagery through an …earth.esa.int/.../participants/112/pres_112_tello.pdf · 2018. 5. 15. · Mariví Tello, Carlos López-Martínez, Jordi

Remote Sensing Lab (RSLab)Dept. of Signal Theory & Communications

Tests on a simulated SAR image. Two dark patches with the same mean damping, one simulated with the parameters corresponding to an artificial oil spill, the other one simulated with those corresponding to a low wind area.

Artificial oil spill

Low wind area

The 2 patches can’t be discriminated through thresholding.

The 2 patches can be discriminated through thresholding.

SEASAR SEASAR WorkshopWorkshop 20082008

ResultsResults

Page 24: Unsupervised Oil Spill Detection in SAR Imagery through an …earth.esa.int/.../participants/112/pres_112_tello.pdf · 2018. 5. 15. · Mariví Tello, Carlos López-Martínez, Jordi

Remote Sensing Lab (RSLab)Dept. of Signal Theory & Communications

SAR

imag

e with

an oi

l spil

lLo

cal e

stima

tion o

f the H

ölder

expo

nents

Egypt, ERS1 pri image, august 92.Egypt, ERS1 pri image, august 92.Egypt, ERS1 pri image, august 92.

SEASAR SEASAR WorkshopWorkshop 20082008

ResultsResults

Page 25: Unsupervised Oil Spill Detection in SAR Imagery through an …earth.esa.int/.../participants/112/pres_112_tello.pdf · 2018. 5. 15. · Mariví Tello, Carlos López-Martínez, Jordi

Remote Sensing Lab (RSLab)Dept. of Signal Theory & Communications

SAR

imag

e with

an oi

l spil

lLo

cal e

stima

tion o

f the H

ölder

expo

nents

Egypt, ERS1 pri image, august 92.Egypt, ERS1 pri image, august 92.Egypt, ERS1 pri image, august 92.

SEASAR SEASAR WorkshopWorkshop 20082008

MAR VERTIDOα α<

ResultsResults

Page 26: Unsupervised Oil Spill Detection in SAR Imagery through an …earth.esa.int/.../participants/112/pres_112_tello.pdf · 2018. 5. 15. · Mariví Tello, Carlos López-Martínez, Jordi

Remote Sensing Lab (RSLab)Dept. of Signal Theory & Communications

Analysis of textureAnalysis of textureSA

R im

age

Loca

l esti

matio

n of th

e Höld

erex

pone

nts

SEASAR SEASAR WorkshopWorkshop 20082008

Page 27: Unsupervised Oil Spill Detection in SAR Imagery through an …earth.esa.int/.../participants/112/pres_112_tello.pdf · 2018. 5. 15. · Mariví Tello, Carlos López-Martínez, Jordi

Remote Sensing Lab (RSLab)Dept. of Signal Theory & Communications

ConclusionsConclusions

An algorithm based on the Hölder exponent has been introduced for automatic detection of oil spills candidates in the sea surface in SAR imagery.

Completely unsupervised. No training is required. No previous filtering (no degradation of the resolution, nor blurring)Multiscale capability

Tests on simulated images have proven its capacity to discriminate between oil spills andlook alikes with the same damping.

SEASAR SEASAR WorkshopWorkshop 20082008

Page 28: Unsupervised Oil Spill Detection in SAR Imagery through an …earth.esa.int/.../participants/112/pres_112_tello.pdf · 2018. 5. 15. · Mariví Tello, Carlos López-Martínez, Jordi

Remote Sensing Lab (RSLab)Dept. of Signal Theory & Communications

IntegrationIntegration ofof thethe algorithmsalgorithmsFrom a computational implementation point of view, the algorithms presented rely on thesame principle:

WTCombination of

waveletcoefficients

Input SAR image Output

Spot detection

From the point of view of the applications, they are closely linked by mutual contributions:

Contour detection

Texture analysis

- Ship detection - Coastline extraction

- Oil spill characterization

- Extraction of oil spills contour- False alarms in oil spill detection discarded- Mask of oil spills candidates

SEASAR SEASAR WorkshopWorkshop 20082008

Page 29: Unsupervised Oil Spill Detection in SAR Imagery through an …earth.esa.int/.../participants/112/pres_112_tello.pdf · 2018. 5. 15. · Mariví Tello, Carlos López-Martínez, Jordi

Remote Sensing Lab (RSLab)Dept. of Signal Theory & Communications

D= 1.13