Effective methods for detecting interesting patterns...

02/03/2017, WS ASI, Rome Marco DIANI

Effective methods for detecting interesting patterns in hyperspectral data

Marco DianiRemote Sensing and Image Processing Group

Department of Information Engineering, University of Pisa

56122 Pisa, Italy

[email protected]


RS&IP Group: main projects

PROJECT DESCRIPTION Funded by/users-customers PRIME

HIGHSENSE2013-2015 – Very high spatial and spectral resolution remote sensing: a novel integrated data analysis system

Italian Ministry of University and Research -MIUR

University of Genoa

SAP4PRISMA2010-2014 – Development of algorithms and products for applications in agriculture and land monitoring to support the PRISMA mission

Italian Space Agency - ASI CNR IMAA

DUCAS2009 – 2013 – Detection in Urban scenario using Combined Airborne imaging Sensors

European Defence Agency - EDA FOI

SULA2010-2012 – Advanced Sensor for Underwater Laser 3D Analysis and Detection

Italian Ministry of Defence DII

COSMOSkyMed

2010-2012 – Development and validation of multitemporal image analysis methodologies for multirisk monitoring of critical structures and infrastructures

Italian Space Agency - ASI University of Genoa

ECOMOS2014- The European Computer Model For Optronic System Performance Prediction (ECOMOS)"

European Defence Agency - EDA DLR

HIPOD2006 - Hyperspectral Imaging Program fOrDefense

European Defence Agency - EDA FOI

Projects funded by or in cooperation with:

Italian Space Agency (ASI)

Ministry of Defence

EDA (ONERA, TNO, RMA, FFI, FGAN etc.)

MIUR

Tuscany Region

National Industry SELEX-ES

Local companies (IDS, FlyBy, etc.)

• Permanent staff– Prof. Marco Diani (Italian Naval Academy)– Prof. Giovanni Corsini (University of Pisa)– Dr. Nicola Acito (Italian Naval Academy)– Dr. Stefania Matteoli (CNR-IEIIT)

• PhD students– Matteo Moscadelli– Dr. Zingoni Andrea


– Object/material detection in HSI

– Object/material detection taxonomy

• Single image analysis

– Anomaly detection

– Spectral matching

• Multitemporal image analysis

– Change detection

– Object relocation

– The “Viareggio 2013” trial

– Conclusions

Outline


The detection problem

Objective: generate a gray scale image (black/white after thresholding) where intensity measures

the degree of interest of the pixels (black=uninteresting/background, white=object/material of

interest).

Hypothesis: Object pixels cover a small fraction of the image.

Detection approach:

Detection vs classification:

Training: in general, only one object reference spectrum is available. Background class must

be learned from the data themselves. Background includes most of the image pixels and is

made up of different classes.

Decision strategy: Bayesian approach based on minimization of the average error probability

does not fit. Neyman-Pearson criterion is invoked. CFAR property is desired.

Dimensionality reduction: methods must preserve rare objects (PCA cannot be used).

Real-time or near real-time is often required.

Detection vs unmixing

not estimated as physical endmembers with corresponding abundances.

0

1

: target material not present On the basis of decide:

: target material present

H

H

X

Noise and lack of fit1 1

Target materials signatures Background signatures

T B

k n

n n

t b t k b n

k n

a a

X Sa Ba w s b w

, 1, 2, ,n B

n nb

Structured background model

(subpixel)


HSI processing for object detection

Anomaly Detection algorithm

Background Characterization

AD MAP

ANOMALY DETECTION

0 0.5 1 1.5 2 2.50

0.1

0.2

0.3

0.4

0.5

0.6

0.7

[m]

Reflecta

nce

Sensor-measured Radiance image

Object Spectral Signature

(Reflectance in VNIR-SWIR)

Object Detection algorithm

Background Characterization

OD MAP

MATERIAL/OBJECT DETECTION

SSB OD

Atmospheric Compensation

5

2.7

2.8

2.9

3

3.1

3.2

3.3

3.4

100 200 300 400 500 600 700

50

100

150

200

T1T2T3T4T5

T6


SSB OD: FM-TD scheme

0 0.5 1 1.5 2 2.50

2000

4000

6000

8000

10000

12000

14000

16000

18000

[m]

Radia

nce [

W

/ c

m2 s

r

m]

0 0.5 1 1.5 2 2.50

2000

4000

6000

8000

10000

12000

14000

16000

18000

[m]

Radia

nce [

W

/ c

m2 s

r

m]

0 0.5 1 1.5 2 2.50

2000

4000

6000

8000

10000

12000

14000

16000

18000

[m]

Radia

nce [

W

/ c

m2 s

r

m]

Set of at-sensor radiance converted

spectra

FORWARD MODELING

MODTRAN C

m

m

sL

1

)(~

reflectance (spectralsignature)

OBJECT SUBSPACE ESTIMATION

N

TTTT ,...,,21

Object subspace basis

SUBSPACE-BASED OBJECT

DETECTION ALGORITHM

ATMOSPHERICCONDITIONS;

VIEWINGCONDITIONS.

sL

At-sensor measured radiance

OFF-LINE


VIEWINGCONDITIONS.


VIEWINGCONDITIONS.

detection statistic


Unsupervised multitemporal analysis: anomalous change detection

50 100 150 200 250 300 350 400

50

100

150

200

250

300

350

400

450

500

50 100 150 200 250 300 350 400

50

100

150

200

250

300

350

400

450

50050 100 150 200 250 300 350 400

50

100

150

200

250

300

350

400

450

500

Radiometric and

atmospheric

equalization

Change

detection

algorithm*

Flight #1

Flight #2

Red spots are changes documented in the ground truth

*CD strategy

Takes into account

coregistration error


Radiometric and

atmospheric

equalization

Spectral

Matching

Flight #1

Flight #2 (wider area)

Spectral

signature

selection

Supervised multitemporal analysis: object rediscovery

ACE GLOBAL

OBJ SUBSPACE

3

4

21


Combining search criteria reduces # of detections

SSB Object detectionMap of pixels whose

spectrum resembles the

reference one

ACD mapMap of pixels where an object

has entered the scene

ACE GLOBAL

OBJ SUBSPACE

ACD

Covariance eq.

Robust to coreg. errors


Combining search criteria reduces # of detections

Pixels whose spectrum

resembles the reference

one

Pixels where an object has

entered the scene .AND.


«Viareggio 2013» trial

N. Acito, S. Matteoli, A. Rossi, M. Diani, G. Corsini, Hyperspectral Airborne «Viareggio 2013 Trial» Data Collection for

Detection Algorithm Assessment, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing,

Vol. 9, Issue: 6, June 2016.

Free data and ground truth available online at http://rsipg.dii.unipi.it/

Campagna di misura

condotta da

CISAM – Pisa

Leonardo (Finmeccanica)

Università di Pisa


Conclusion

Object/material detection is a method for the automatic analysis of hyperspectral images

which is of potential interest in different applications (detection and tracking of pollutants,

detection of areas of vegetation stress, geology, wide area surveillance, search and

rescue - SAR) and for different final users (civil protection, coast guard, agencies for

defense and security, etc.).

Object detection aim to drive the operator’s attention to few ROIs that deserve further

analysis. This reduces the time required to explore wide areas searching for specific

objects and materials.

Hyperspectral sensors complement the information carried by other instruments

(broadband imagery, SAR, etc.) and improve the capability of detecting objects/materials

in a monitored area using the information carried by the spectrum of the radiance

emitted/reflected from an object.

Custom advanced multidimensional signal processing is needed to detect small dim

objects.

Research is still ongoing on many aspects of the HSI detection problem.

Detectability in a subpixel framework.

Technological and processing aspects related to the exploitation of such spectral

regions as SWIR, MWIR, LWIR. This is especially true for MWIR and LWIR regions

that are necessary to guarantee day/night operability.

Benefits from big data analysis?

Effective methods for detecting interesting patterns...

Documents

Transcript of Effective methods for detecting interesting patterns...