Fast Target Detection Framework for Onboard ... - HyspIRI - NASA

Applied Research LLC

Fast Target Detection Framework for Onboard Processing of Multispectral and Hyperspectral Images B. Ayhan and C. Kwan June 3, 2015 Applied Research LLC 9605 Medical Center Dr., Rockville, MD20850

Research supported by NASA SBIR Program

1


1. Contents

1. Contents

2

2. Research Objectives

3

3. Technical Approach

4-8

4. Results

9-20

2


2. Research Objectives •

Develop a robust, automated, and real-time target detection system under varying illumination, atmospheric conditions and target/sensor viewing geometry.

•

Demonstrate the feasibility of the system using actual and/or simulated data.

3


3. Technical Approach Fast target detection framework •

Conventional target detection is done in the reflectance domain: a lot of computations due to atmospheric compensation, not suitable for onboard processing, difficult to change mission goals during mission.

•

Our approach is done in the radiance domain. Only a few target signatures (reflectance) need to be transformed to the radiance domain. This is very suitable for onboard processing such as search and rescue missions.

•

JHU/APL developed a similar approach that uses MODTRAN and AFWA MM5. Our approach was motivated by [1], which is a hybrid framework that uses MODTRAN and a nonlinear analytical model.

[*1] “Hyperspectral material identification on radiance data using single-atmosphere or multiple-atmosphere modeling," Adrian V. Mariano ; John M. Grossmann, J. Appl. Remote Sens. 4(1), 043563 (November 23, 2010).

4


3. Technical Approach • Radiance equation model parameter estimation using MODTRAN outputs [*1]

Bρ A Aρ L= + + P − α Dρ 1− ρ AS 1 − ρ AS L: Radiance

ρ : Material reflectance ρ A : Adjacent region reflectance

S

: Spherical albedo

A,B : Coefficients that depend on atmospheric, geometric and solar illumination conditions P : Path radiance D : Radiance due to direct solar illumination

α : Amount of solar occlusion


5


3. Technical Approach • Radiance equation model parameter estimation using MODTRAN outputs [*1] Atmospheric, geometric and solar illumination conditions

ρ1 = 0.05

ρ2 = 0.6

MODTRAN Simulation (1)

MODTRAN Simulation (2)

DRCT_RFLT: Direct Reflectance (MODTRAN output) GRND_RFLT: Ground Reflectance (MODTRAN output) SOL_SCAT: Solar Multiple Scattering (MODTRAN output)

MODTRAN Output "DRCT_REFL(1)" MODTRAN Output " GRND_RFLT(1) " MODTRAN Output " SOL_SCAT (1) "

MODTRAN Output "DRCT_REFL(2)"

Estimation of radiance equation parameters (A,B, D,P and S) L=

Bρ A Aρ + + P − α Dρ 1 − ρ AS 1 − ρ AS

MODTRAN Output " GRND_RFLT(2)" MODTRAN Output " SOL_SCAT(2) "


6


3. Technical Approach • Radiance equation model parameter estimation using MODTRAN outputs [*1] Bρ A Aρ L= + + P − α Dρ 1 − ρ AS 1 − ρ AS Cρ1 = SOL _ SCAT (1) and Gρ1 = GRND _ RFLT (1) , and

Suppose

Cρ 2 = SOL _ SCAT (2) and Gρ 2 = GRND _ RFLT (2) Then,

Gρ 2 =

Aρ 2 B ρ2 , Cρ 2 = +P 1 − ρ2 S 1 − ρ2 S

and

Gρ1 =

Aρ1 B ρ1 , Cρ1 = +P 1 − ρ1S 1 − ρ1S

The radiance model parameters can then be found as: D = DRCT_RFLT(1) / ρ1 = DRCT_RFLT ( 2) / ρ2

S=

A=

Gρ 2 / ρ 2 − Gρ1 / ρ1 Gρ 2 − Gρ1 Gρ 2

ρ2

− Gρ 2 S

B = (Cρ1 − P )( P=

1

ρ1

− S)

S (Cρ1 − Cρ 2 ) + Cρ 2 / ρ 2 − Cρ1 / ρ1 1 / ρ 2 − 1 / ρ1


7


3. Technical Approach Advantages of the proposed system • Eliminates the need of applying atmospheric correction on the whole image cube and instead simulates the variants of the radiance signature of the target of interest and searches for these signatures in the test radiance image cube • The effects of different illumination, atmospheric conditions, occlusion and varying sensor/target viewing geometries are taken into effect during the simulation of the radiance spectral profiles of the target of interest • Allows generation of look-up tables for several radiance signature variants of the target which will reduce computation/processing time for target detection in operations like “search and rescue” that require quick on-board decisions

Proprietary Information - ARLLC

8


4. Results • Demonstration of radiance model parameter estimation and simulating radiance profiles with the model parameter estimates Atmospheric, solar illumination and geometric location parameters used in two MODTRAN runs to estimate model parameters (S, A, P, D, B)

Model

Tropical

IHAZE

Rural

VIS

5 km

H2OSTR

0.5

Altitude

1 km

Approximate observer position

Latitude: 39.3305º (N), Longitude: 76.2879 (W)

Date

August 29, 1995

Time data collected

18:37 UTC

0.4

S

14

0.3

12

0.25

10

0.2

8

0.15

6

0.1

4

0.05

2

0

0

500

1000

1500 Frequency (nm)

2000

2500

0

3000

0

500

1000

1500 Frequency (nm)

7

3.5

P

A

A Amplitude

Amplitude

Plots of estimated model parameters (S, A, P, D, B)

16

S 0.35

P 3

2500

3000

18 D

D

2000

6

B

B

16 14

2.5

5

2

4

1.5

1

Amplitude

Amplitude

Amplitude

12

3

10 8 6

2

4

0.5

1

2 0 0

500

1000

1500 Frequency (nm)

2000

2500

3000

0 0

500

1000

1500 Frequency (nm)

2000

2500

3000

0

0

500

1000

1500 Frequency (nm)

2000

2500

3000

9


4. Results • Comparing the simulated radiance profiles (from the model parameter estimates) with the MODTRAN results 25

MODTRAN generated radiance Radiance estimated with model parameters

Case 1 (fixed reflectance of 0.6) 20

1

0.8

0.4

10

Radiance model with estimated parameters

0.2

0 0.4

15 Radiance

Reflectance

MODTRAN 0.6

0.6

0.8

1

1.2 1.4 1.6 Wavelength (µm)

1.8

2

2.2

5

2.4

0

0

500

1000

12

Case 2 (reflectance of a green tree)

1500 2000 Wavelength (nm)

2500

3000

MODTRAN generated radiance Radiance estimated with model parameters

Reflectance Data

10

1 tree1000010x2Easd0x2Eref 0.5

8

MODTRAN Radiance

R e f le c t a n c e

0.4

0.3

0.2

Radiance model with estimated parameters

0.1

0

0.4

0.6

0.8

1

1.2

1.4 1.6 Wavelength (µm)

1.8

2

2.2

2.4

6

4

2

0

The results are almost identical which shows that radiance model parameter estimation is successful !

0

500

1000


2500

3000

10


4. Results • AVIRIS data for Los Angeles Station Fire (Aug 2009) to detect burnscar The Station Fire took place between August 26 and October 16, and a total of 160,577 acres (251 sq mi; 650 km2) were affected, 209 structures had been destroyed, including 89 homes [*2]. It first started in the Angeles National Forest near the U.S. Forest Service ranger station on the Angeles Crest Highway (State Highway 2) [*2].

AVIRIS images acquired on October 6, 2009 (fire is mostly over)

[*2] http://en.wikipedia.org/wiki/2009_California_wildfires

11


4. Results • Getting groundtruth of burned locations for AVIRIS data using MODIS MCD45A1 product MODIS H8-V5 tile (Los Angeles region falls into this tile)

MCD45A1 burned area product for this tile for Sep 2009 (red pixels indicate burned areas)

Zoom into the region for the LA Station Fire in MCD45A1 product 10 20

500

30 40 1000

50 60 1500

70 80 2000

90

500

1000

1500

2000

20

40

60

80

100

120

Based on AVIRIS image data coverage for each strip the groundtruth maps for burned area are extracted

12


4. Results • Burnscar pixel selections from AVIRIS image strips Based on extracted groundtruth maps and also visual examination, burnscar pixels are selected from each AVIRIS image data (pixels with direct sunlight and under shadow with some occlusion) Direct sunlight:

With occlusion:

r09

r10

r11

r12

r13

r14

r15 13


4. Results • Radiance and reflectance signatures of selected burnscar pixels r09 (Sample:389, Line:1929) r09 (Sample:575, Line:1955) r10 (Sample:3437, Line:2887) r10 (Sample:665 Line:2778) r11 (Sample:498, Line:2043) r11 (Sample:532, Line:2298) r12 (Sample:377, Line:1997) r12 (Sample:398, Line:1953) r13 (Sample:182, Line:2197) r13 (Sample:163, Line:2235) r14 (Sample:519, Line:3232) r14 (Sample:537, Line:3280) r15 (Sample:619, Line:4013)

2000 1800 1600

1200 1000

0.7 0.6 0.5

0.3

400

0.2

200

0.1 1000

1500 Wavelength (nm)

2000

2500

0

3000

0

r09 (Sample:389, Line:1929) r09 (Sample:575, Line:1955) r10 (Sample:3437, Line:2887) r10 (Sample:665 Line:2778) r11 (Sample:498, Line:2043) r11 (Sample:532, Line:2298) r12 (Sample:377, Line:1997) r12 (Sample:398, Line:1953) r13 (Sample:182, Line:2197) r13 (Sample:163, Line:2235) r14 (Sample:519, Line:3232) r14 (Sample:537, Line:3280) r15 (Sample:619, Line:4013)

2000

Radiance (7 MODIS bands)

0.8

600

500


0.9

0.4

0

Actual radiance signatures of selected pixels (in DN) for MODIS bands only

1

800

1500

1000

Reflectance signatures of selected pixels after QUAC atmospheric correction (MODIS bands only)

Reflectance

Radiance

1400

Reflectance signatures of selected pixels (after QUAC atmospheric correction) Reflectance

2200

Actual radiance signatures of selected pixels (in DN)

0.7

500


2000

2500


0.6

0.5

0.4

0.3

0.2 500

0.1

2

3

4 MODIS band index

5

6

7

0

1

2

3

4 Wavelength (nm)

5

MODIS bands: 648 nm, 858 nm, 470 nm, 555 nm, 1240 nm, 1640 nm, and 2130 nm (In target detection investigations in this work, we used the 7 bands that are closest to these MODIS bands )

6

7

14


4. Results • Atmospheric, illumination and geometric location conditions used in radiance model parameter estimation and simulation of radiance signatures for the selected reflectance profiles Geometric locations of the AVIRIS image strips and corresponding UTC values

AVIRIS Image Strip

Min Lat

Min Long

r09

34.1436

-118.376

34.5639

-118.276

18:51

r10

34.1488

-118.31

34.6222

-118.208

19:01

r11

34.152

-118.236

34.5625

-118.133

19:11

r12

34.1571

-118.165

34.5788

-118.066

19:21

r13

34.1603

-118.091

34.5626

-117.99

19:29

r14

34.1575

-118.022

34.5984

-117.924

19:39

r15

34.1634

-117.943

34.563

-117.847

19:48

Max Lat

Max Long

UTC

Ground Elevation (km) 0.67 0.93

1.11

Variants of selected burnscar reflectance profiles (based on QUAC)

1.22

0.7 r09 r09 r10 r10 r11 r11 r12 r12 r13 r13 r14 r14 r15

0.6

Reflectance

0.5

0.4

0.3

(Sample:389, Line:1929) (Sample:575, Line:1955) (Sample:3437, Line:2887) (Sample:665 Line:2778) (Sample:498, Line:2043) (Sample:532, Line:2298) (Sample:377, Line:1997) (Sample:398, Line:1953) (Sample:182, Line:2197) (Sample:163, Line:2235) (Sample:519, Line:3232) (Sample:537, Line:3280) (Sample:619, Line:4013)

1.17 1.18

Atmospheric and time/date conditions

0.2

0.1

0

0

500


0.98

2000

2500

Model

Mid Latitude Summer (MODEL = 2)

VIS (Visibility) H2OSTR (Water vapor)

20 km 0.270

IHAZE

Rural (IHAZE =1)

Altitude Observer position

14 km (see above table for observer positions)

Data collection day Data collection time

Oct 6, 2009 (see above table for UTC times) 15


4. Phase 1 Results • Simulated radiance signatures with estimated radiance model parameters and target detection results Simulated radiance signatures for the selected pixels


3000

2500

Radiance

2000

1500

1000

500

0 0

500


2000

2500

MODTRAN-generated radiance signatures (MODIS bands only)

Radiance

Simulated radiance signatures for the selected pixels (MODIS bands only)


3500

3000

2500

2000

1500

Target detection score images with M-SAM for "r09" using simulated radiance signatures, using actual radiance signatures and using actual reflectance signatures with QUAC (groundtruth map for “r09” image is also shown). In target detection, we used the 7 MODIS bands only.

 si , r j   SAM (s i , r j ) = cos−1   s r   i j 

1000

M-SAM(si , burnscar) = min( SAM (si , r1 ), SAM (si , r2 ),..., SAM (si , rK )) 500

1

2

3

4 MODIS band index no

5

6

7

where burnscar = {r1 , r2 ,..., rK } si = test pixel spectral profile r j = burnscar signature variant 16


4. Results • ROC curves with M-SAM detection technique for AVIRIS images r10

r09

1

1

r10

r09

0.9

0.9

0.8 Actual radiance signatures with M-SAM Actual reflectance signatures with M-SAM MODTRAN-simulated radiance signatures with M-SAM

0.7 0.6

Probability of detection


0.8

0.5 0.4 0.3

0.5 0.4 0.3 0.2

0.1

0.1 0.1

0.2

0.3

0.4 0.5 0.6 False alarm rate r11

0.7

0.8

0.9

0 0

1

0.1

0.2

0.3

0.4 0.5 0.6 False alarm rate r12

0.7

0.8

0.9

1

1

1

r11

0.6

0.2

0 0

Actual radiance signatures with M-SAM Actual reflectance signatures with M-SAM MODTRAN-simulated radiance signatures with M-SAM

0.7

0.9

0.9

0.8

0.8 Actual radiance signatures with M-SAM Actual reflectance signatures with M-SAM MODTRAN-simulated radiance signatures with M-SAM

0.7 0.6



r12

0.5 0.4 0.3 0.2


0.6 0.5 0.4 0.3 0.2

0.1 0 0

0.7

0.1 0.1

0.2

0.3

0.4 0.5 0.6 False alarm rate

0.7

0.8

0.9

1

0 0

0.1

0.2

0.3


0.7

0.8

0.9

1

17


4. Results • ROC curves with M-SAM detection technique r13

r14

1

r14

1 0.9

0.8

0.8

0.7


0.9


0.6 0.5 0.4 0.3

0.7 0.6

0.4 0.3 0.2

0.1

0.1

0 0

0.1

0.2

0.3


0.7

0.8

0.9

0 0

1


0.5

0.2

0.1

0.2

0.3


0.7

0.8

0.9

1

r15 1

r15 0.9 0.8 Probability of detection


r13

0.7 0.6 0.5


0.4 0.3 0.2 0.1 0 0

0.1

0.2

0.3


0.7

0.8

0.9

1

18


4. Results • AUC (Area under curve) measures from ROC curves

r09

AUC(overall) AUC(partial)

Using actual radiance signatures (target detection in radiance domain) 0.8958 0.0761

r10


0.9395 0.0655

0.9183 0.0612

0.9400 0.0653

r11


0.9028 0.0427

0.8814 0.0472

0.9087 0.0471

r12


0.8696 0.0473

0.8487 0.0544

0.8783 0.0514

r13


0.8772 0.0691

0.8601 0.0679

0.8771 0.0676

r14


0.8370 0.0648

0.8055 0.0631

0.8415 0.0659

r15


0.8536 0.0676

0.7794 0.0585

0.8713 0.0687

AVIRIS Filename Analysis Type

Using QUAC reflectance signatures (target detection in reflectance domain) 0.8856 0.0750

Using simulated radiance signatures (reflectances are from QUAC) (target detection in radiance domain ) 0.8935 0.0745

Overall: Whole ROC curve is used in AUC computation Partial: ROC curve up to a false alarm rate of 0.1 is used in AUC computation

19


4. Results •

Observations - With the simulated radiance signature library using estimated radiance models, we observed detection performances close to the detection performance obtained with using actual radiance signatures that are retrieved from actual images and in some cases slightly better than those. - It is possible that if a better set of MODTRAN parameters has been selected in the radiance model parameter estimation that fully complies with the atmospheric conditions during the actual data acquisition, or a more realistic set of reflectance signature variations for burnscar has been used, the detection performances could perhaps be further improved. - These results are found highly promising in demonstrating the feasibility and effectiveness of the proposed fast target detection idea.

20

Fast Target Detection Framework for Onboard ... - HyspIRI - NASA

Fast Target Detection Framework for Onboard ... - HyspIRI - NASA

Suggest Documents

Fast Target Detection Framework for Onboard ... - HyspIRI - NASA

Onboard Processing for Low-latency Science for the HyspIRI Mission

HyspIRI - MDPI

An introduction to the NASA Hyperspectral InfraRed Imager (HyspIRI ...

Predictive Onboard Flow Control for - NASA Technical Reports Server ...

fast circlet based framework for optic disk detection

A Fast Framework for Abrupt Change Detection Based on Binary

fast onboard texture analysis for planetary ... - Semantic Scholar

12.4 INFLIGHT ICING DETECTION WITH ONBOARD MICROWAVE ...

Fast ignitor target studies for HiPER

Rapid spectral cloud screening onboard aircraft and ... - aviris-ng - NASA

improved target recognition and target detection

improved target recognition and target detection

Unsupervised subpixel target detection for hyperspectral ... - UMBC

Sensor Deployment Strategy for Target Detection - CiteSeerX

Matched filters for multispectral point target detection

Recovery of Target Response Function for Centre-of ... - cddis - NASA

Fast stixel computation for fast pedestrian detection - Rodrigo Benenson

Fast, specific, ultrasensitive detection

Fast Collision Detection Me

fast immunohistochemical biomarker detection device for cancer ...

Fast Obstacle Detection for Urban Traffic Situations

Fast Collision Detection for Haptic Displays Using

Fast Community Detection For Dynamic Complex Networks