Automated Assessment of NIIRS and GRD of High Resolution Satellite Images through Edge Profile Analysis of Natural Targets

Taejung Kim, Jae-In Kim

Image Engineering Lab Dept. of Geoinformatic Engineering Inha University

Backgrounds

Various ways of describing image quality
From engineering side, there are many technical parameters Ground sampling distance, Modulation Transfer Function(MTF) (ratio @ sampling freq.), Signal to Noise Ratio (SNR), Relative Edge Response (RER), etc.
Tech. parameters may not represent image quality for user Image users may be more interest in other parameters mapping accuracy, interpretability, etc.
Image quality regarding interpretability NIIRS (National Image Interpretability Rating Scales) GRD (Ground Resolvable Distance)
Image quality assessed mostly by Artificial Targets
Usually for calibration / validation purpose Specially manufactured artificial targets are used Special arrangements (target size, orientation) are required Images around targets are analyzed for RER and SNR Edge profiles are transformed to MTF through curve fitting

(Helder et al., 2004)

Research Purpose
Automated image quality assessment from natural targets
artificial targets natural targets manual edge selection automated selection RER, MTF, SNR GRD, NIIRS reliability of image quality parameters
Operational image quality assessment of all remote sensing images without extra costs

NIIRS

(National Image Interpretability Rating Scales)
Originally used for intelligence/military images In 1996, published by IRARS (Imagery Resolution Assessments and Reporting Standards) For each rating, identifiable targets are defined Separate rating scales exist for military targets and civil/natural targets and for panchromatic, multispectral, radar images NIIRS values are assessed visually by certified image analysts NIIRS values are provided within the satellite metadata

GRD (m)

over 9.0 4.5 9.0

Visible NIIRS

Interpretability of the imagery is precluded by obscuration, degradation, or very poor resolution Detect a medium-sized port facility and/or distinguish between taxiways and runways at a large airfield. Detect large hangars at airfields. Detect large static radars (e.g., AN/FPS-85, COBRA DANE, PECHORA, HENHOUSE).

2.5 4.5

1.2 2.5 0.75 1.2
Identify the wing configuration (e.g., straight, swept, delta) of all large aircraft (e.g., 707, CONCORD, BEAR, BLACKJACK).
Identify all large fighters by type (e.g., FENCER, FOXBAT, F-15, F-14). Detect the presence of large individual radar antennas (e.g., TALL KING). Distinguish between a MIDAS and a CANDID by the presence of refueling equipment (e.g., pedestal and wing pod). Identify radar as vehicle-mounted or trailer-mounted.

0.40 - 0.75

0.20 0.40 0.10 0.20 less than 0.10
Distinguish between models of small/medium helicopters (e.g., HELIX A from HELIX B from HELIX C, HIND D from HIND E, HAZE A from HAZE B from HAZE C).
Identify fitments and fairings on a fighter-sized aircraft (e.g., FULCRUM, FOXHOUND). Identify the rivet lines on bomber aircraft. Detect horn-shaped and Wshaped antennas mounted atop BACKTRAP and BACKNET radars. Differentiate cross-slot from single slot heads on aircraft skin panel fasteners. Identify small light-toned ceramic insulators that connect wires of an antenna canopy.

NIIRS assessment by GIQE

General Image quality Equation Proposed by regression analysis between NIIRS, GSD, MTF and SNR values of images Enables assessment of NIIRS from tech. parameters determined by edge analysis
NIIRS = a - b* log(GSDGM) + c* log(RERGM) (d*H) (e*G/SNR)
RERGM: Geometric means of Relative Edge Response in x and y direction H: Geometric means of Overshoot height G: Noise gain due to Edge sharpening, Kernel Value of MTF Correction GSD: Ground Sampling Distance SNR: Signal to Noise Ratio
(Ground Resolvable Distance)
The minimum distance between two objects to be identified as separate objects Inverse of Line pairs per mm (lp/mm) GRD is assessed by image analysts

GRD assessment

GRD can be assessed from PSF (Point Spread Function)
H : Flying height f : Focal length R : Half peak width of PSF

Proposed procedures

Select initial edge points from artificial vs. natural targets manually vs. automatically Determine edge orientation and generate edge profiles Calculate normalized edge profile and edge center Check the criteria for accepting edge profiles Calculate RER, H, SNR and NIIRS Generate point spread function and calculate GRD Repeat the process for other edge points (usually > 50) Determine NIIRS and GRD for the whole scene
Validation of GRD/NIIRS Assessment
Orientation-invariant edge analysis
GIQE uses RER in only x- and y-directions For natural targets, we have to use edges of arbitrary orientation We need to extract edge profiles perpendicular to edge orientation Test image: bar patterns with orientation changed incrementally by 15 by different cameras
Camera Tested: SONY Siber-Shot DSC-S950 SONY 550(DSLR) Cannon Exsus 900 Ti Samsung Kenox S500

1.00 0.80 0.60

0.40 0.20 0.00
1.00 0.80 0.60 RER 0.40 0.20

SONY Siber-Shot DSC-

SONY 550(DSLR)

0.80 0.60

0.40 0.20 0.75 90

Cannon Exsus 900 Ti

60 Samsung Kenox S90
GRD estimation from in-door scenes

Test image:

Camera spec.:
Model CCD size Focal Length Image Size CCD Cell size
EOS 450D 22.2mm 14.8mm 55mm 0.005197mm
Imaging distance (Flying height): 981mm, 1232mm,1454mm, 2090mm, 3132mm
Edge analysis for quality assessment
From bar pattern, extract edge profiles, PSF and GRD GRD values assessed by 7 researcher were averaged as reference GRD values
GRD estimation from edge analysis
GRD values from edge analysis were almost identical to reference (RMSE: 0.01mm)
GRD Imaging Distance 3132mm 2090mm 1454mm 1232mm 981mm Reference GRD 0.7081mm 0.4753mm 0.3288mm 0.3001mm 0.2324mm 2 * GSD 0.5863mm 0.3912mm 0.2722mm 0.2306mm 0.1836mm Average of GRD of Individual Average GRDs Edge Profile 0.7094mm 0.7057mm 0.4747mm 0.4665mm 0.3305mm 0.3230mm 0.3058mm 0.3063mm 0.2127mm 0.2112mm
GRD estimation from an out-door scene

Test data

(Bruce Mathews and Theodore Zwicker, 1999)
- Tri bar pattern with varying sizes - Reference GRD is estimated by checking minimum identifiable bar pattern
Bar Group Horiz. 151.25 75.60 67.40 60.00 53.50 47.60 Size(inches) Vert. 30.25 15.13 13.47 12.00 10.69 9.53 GRD (in) 60.50 30.25 26.95 24.01 21.39 19.06 Bar Group Horiz. 16.84 8.42 7.50 6.68 5.96 5.31 Size(inches) Vert. 3.37 1.68 1.50 1.34 1.19 1.06 GRD (in) 6.74 3.37 3.00 2.67 2.38 2.12

27th Group

100 Edge locations were selected Reference inches Pixel 2.8350 2.0693 GSD*2 2.7400 2.0000
Extracted Edge Profile and Point Spread Function GRD Average of Individual GRDs 2.7784 2.0280 GRD of Average Edge Profile 2.6552 1.9381
GRD estimation from simulated images
from each ref. images, 3 simulated images were generated
(a)scene1, distance 3132mm
(b)scene2, distance 2090mm
(c)scene3, distance 1454mm
(d)scene4, distance 1232mm
(f)scene5, distance 981mm
Ref. Image * PSF (Gaussian with GRD 1,2,3) 5 refs 3 PSFs = 15 simulated images
reference GRDs were estimated by visual inspection theoretic GRDs were also calculated mathematically
For each image, edge profiles at 200 locations were extracted

Scene 1

Conv PSFs GRD 1.0Pixel 2.0Pixel 3.0Pixel 0001

Scene 2

0002 0.0177 0.2727 0.0261 -0.2207

Scene 3

0003 0.0227 -0.0007 -0.1624 -0.1842 0004

Scene 4

0005 0.0309 0.0027 0.0589 -0.1369

Scene 5

Total RMSE(Pixel) 0.0992 0.1542 0.1034 0.1690
Difference between theoretically driven GRDs vs estimated GRDs
0.0756 0.0493 0.0327 -0.0223 -0.2043 -0.2051 -0.1480 -0.2021
NIIRS estimation from simulated images
generated images with NIIRS by changing GSDs check the minimum identifiable font size for each image blur the reference by Gaussian filter to make the same minimum font size as the images with different GSDs

Reference

NIIRS Minimum font size (pt) Image size 0.1000 x 1500

1.4145 x GSD0

-0.707 x 1060

2.0000 x GSD0

-1.500 x 750

2.8302 x GSD0

-1.353 x 530

GSD0, Reference

Estimated NIIRS were very close to the true values
Reference RER SNR H GRD NIIRS True NIIRS Estimated NIIRS |Error| 0.6376 47.5078 1.0664 1.5861 3.9320 Min. pt = 5 0.3993 162.0340 0.9692 2.3732 3.4427 -0.5000 -0.4893 0.0107 Min pt = 7 0.2347 92.2690 0.7727 3.8407 2.9258 -1.0000 -1.0062 0.0062 Min pt = 11 0.1560 21.9131 0.6898 5.5866 2.4688 -1.5000 -1.4632 0.0368
Validation of the use of natural targets
Artificial targets vs. natural targets
Test images: Komspat-2 images with tarps

Area GSDx GSDy

Taejeon 0.979 0.994

Kimje 1.000 1.000

Jinju 0.980 0.996

Hamyang 1.092 1.048 21

Using natural targets, similar quality parameters were assessed Differences in NIIRS are within the error range of GIQE (1=0.30) Degradation in SNRs from natural targets We need more test with other dataset
Tarp 10 0.2967 59.10 0.8353 2.37 3.53 Natural 2069 0.3028 49.48 0.8834 2.68 3.48 Natural 730 0.2898 42.94 0.8529 2.86 3.46 Kimje Points RER SNR H GRD(m) NIIRS Hamyang Points RER SNR H GRD(m) NIIRS Tarp 10 0.2238 38.18 0.7987 2.89 3.15 Tarp 10 0.2413 48.68 0.7704 3.34 3.21 Natural 976 0.2768 36.84 0.8324 2.87 3.40 Natural 707 0.2736 44.04 0.8357 3.26 3.29 22

Daejeon Points RER SNR H GRD(m) NIIRS Jinju Points RER SNR H GRD(m) NIIRS
Tarp 10 0.3065 59.52 0.8058 2.56 3.62
Validation of automated edge selection

Automated edge selection

apply line detection algorithm check line length (10 pixels) Extract edge profiles edge profile selection criteria are same as manual selection
Tests with Kompsat-2 images
Quality degradation for automated edge selection (in particular in GRD) better edge selection criteria required Differences in NIIRS are within the error range of GIQE (1=0.30)
Tarp 10 0.2967 59.10 0.8353 2.37 3.53 Tarp 10 0.3065 59.52 0.8058 2.56 3.62
Natural Manual Natural Auto
Kimje Points RER SNR H GRD(m) NIIRS
Tarp 10 0.2238 38.18 0.7987 2.89 3.15

Natural

2069 0.3028 49.48 0.8834 2.68 3.48 Natural 730 0.2898 42.94 0.8529 2.86 3.46
55806 0.2837 39.38 0.8474 3.03 3.44

Natural Auto

976 0.2768 36.84 0.8324 2.87 3.40
55806 0.2707 34.49 0.8198 3.11 3.39

Hamyang

Points RER SNR H GRD(m) NIIRS
10 0.2413 48.68 0.7704 3.34 3.21
707 0.2736 44.04 0.8357 3.26 3.29
17858 0.2768 38.15 0.8445 3.31 3.40
36101 0.2716 36.02 0.8314 3.44 3.28 24
GRD/NIIRS estimation from sat. images
Acquisition date Area Image size GSD X(m) GSD Y(m) G(Noise Gain)

QB001 IK001

2002/2/7/2/34 Daejeon 1100411004 0.90 0.96 4.16
2007/2/23/01/49 Damyang 1500015500 1.086 1.039 2.34
2005/1/15/2/27 Daejeon 2504427552 0.793 0.711 4.16
- using natural targets - manual or automatic selection - Published NIIRS : Value in Metadata (QB) or in literature (IK) - Slight quality degradation for automated selection (but not big) - Differences in NIIRS are within the error range of GIQE (1=0.30)
type QuickBird IKONOS Kompsat-2 edge selection manual auto manual auto manual auto points RER SNR H GRD(m) NIIRS Published NIIRS 4.5000 4.3000 -

372 11749

0.6389 0.6128 0.5354 0.5334 0.3705 0.3336
42.89 38.55 38.04 36.67 36.41 34.29
1.037 1.043 1.012 1.023 0.957 0.932
1.11 1.15 1.46 1.49 2.68 2.99
4.65 4.57 4.11 4.09 3.51 3.39
Automated NIIRS estimation for images along the same strip (Komspat-2 strip)

ID Points RER SNR H GRD(m) NIIRS 0.3841 36.34 1.051 2.92 3.2054 0.3894 34.77 1.056 2.93 3.2493 0.3966 35.19 1.057 2.89 3.1995 0.3901 35.24 1.055 2.91 3.3059 0.4026 34.80 1.059 2.87 3.3876 0.4186 34.10 1.060 2.79 3.4336 0.4205 34.41 1.058 2.78 3.3683 0.4160 35.17 1.064 2.81 3.3497 0.4190 35.32 1.066 2.82 3.2310 0.4201 35.32 1.071 2.82 3.45
4.00 3.80 3.60 3.40 3.20 3.00 2.80 2.60 2.40 2.20 2.00
GRD distribution Mean: 2.86m, Stdev: 0.06m NIIRS distribution Mean: 3.40, Stdev: 0.04
All images on the same strip showed very constant GRD/NIIRS values. NIIRS values are within the error range of GIQE (1=0.30)!
Automated NIIRS estimation for images along the same strip (IKONOS strip)
ID Points RER SNR H GRD(m) NIIRS 0.5107 39.90 1.033 1.69 3.1960 0.4998 38.91 1.020 1.71 3.1503 0.4932 41.17 1.021 1.72 3.1523 0.5089 41.77 1.017 1.67 3.5230 0.5206 46.30 1.029 1.60 4.6618 0.5209 46.72 1.034 1.60 4.4574 0.5160 46.82 1.031 1.61 4.4487 0.5219 45.01 1.032 1.63 4.4099 0.5225 42.60 1.036 1.65 4.3712 0.5137 42.11 1.028 1.67 3.99
5.00 4.80 4.60 4.40 4.20 4.00 3.80 3.60 3.40 3.20 3.00 I-1 I-2 I-3 I-4 I-5 I-6 I-7 I-8 I-9 I-10
GRD distribution Mean: 1.65m, Stdev: 0.04m NIIRS distribution Mean: 3.99, Stdev: 0.04 All images on the same strip showed very constant GRD/NIIRS values. NIIRS values are within the error range of GIQE (1=0.30)!

Conclusions

Conclusions
GRD/NIIRS estimation through edge analysis Feasible but tests with ref. NIIRS are required. The use of natural target Feasible but tests with more dataset are required. Automated image quality assessment is feasible But, more rigorous selection criteria is required Can the proposed method be used for image quality assessment for operational basis?