Coefficient of Thermal Expansion and their Importance.pptx
Classification Accuracy for RISAT-1 Hybrid Polarimetric Data
1. Varsha Turkar1, Shaunak De1, G. G. Ponnurangam1,
Rinki Deo1, Y.S. Rao1 and Anup Das2
1 CSRE, Indian Institute of Technology Bombay
2 Space Application Center, ISRO
APSAR 2013 Tsukuba Japan - Sep. 23 - 27, 2013.
2. Introduction
• Studies on compact and hybrid polarimetric SAR data is
currently in focus.
• Primary reasons: [1]
• Wider swath than Full-Pol mode
• Low PRF requirement – less demanding on hardware
• Higher incidence angle range coverage
• Studies demonstrated with compact-pol: [2]
• Crop classification
• Soil moisture estimation
• Ship detection and sea-ice classification
[1] R.K. Raney, “Hybrid-polarity SAR architecture”, IEEE Trans. Geosci. Remote Sens., 45(11): 3397 –3404,
Nov. 2007
[2] F.J. Charbonneau, B. Brisco, R.K. Raney, H. McNarin, P.W.Vachon, J.Shang, R. DeAbreu, C. Champagne,
A. Merzouki and Geldsetzer, “Compact polarimetry overview and application assessment”, Can. J. Remote
Sens., vol. 36, 2, pp. s298-s315, 2010.
APSAR 2013 Tsukuba Japan - Sep. 23 - 27, 2013.
3. RISAT-1 – The first compact pol SAR
• The work carried out so far has been
based on simulated hybrid-pol data
• RISAT-1 – first spaceborne hybrid
PolSAR system
• Indigenously developed
• C-band (5.35 GHz) hybrid polarimetric
SAR
• Multi-polarisation and multi-resolution
• 50m – 1m spatial resolution
• RH/RV, HH/HV modes supported
• Right circular transmit and coherent linear
receive mode (CTLR)
Courtesy: ISRO
APSAR 2013 Tsukuba Japan - Sep. 23 - 27, 2013.
4. Backscatter ( 0 ) Calculation
• SLC data is supplied as 16 bit
integers
• Converted to complex floating
point
• Radiometric correction of data
• The calibration constant (KdB) is
supplied
𝐷𝑁 = 𝐼2 + 𝑄2
𝝈 𝟎 𝒅𝑩 = 𝟐𝟎 𝒍𝒐𝒈 𝟏𝟎 𝑫𝑵 − 𝑲 𝒅𝒃 + 𝟏𝟎 𝒍𝒐𝒈 𝟏𝟎
𝒔𝒊𝒏 𝜽𝒊
𝒔𝒊𝒏 𝜽 𝒄
Here: 𝐷𝑁 = 𝐷𝑖𝑔𝑖𝑡𝑎𝑙 𝑁𝑢𝑚𝑏𝑒𝑟
𝜃𝑖 = 𝐿𝑜𝑐𝑎𝑙 𝐼𝑛𝑐𝑖𝑑𝑒𝑛𝑐𝑒 𝐴𝑛𝑔𝑙𝑒
𝜃𝑐 = 𝐶𝑒𝑛𝑡𝑒𝑟 𝐼𝑛𝑐𝑖𝑑𝑒𝑛𝑐𝑒 𝐴𝑛𝑔𝑙𝑒RH 0 (db) - Mumbai
5. Export to C2 Matrix
𝐶11 = (𝑅𝐻 𝑅𝑒𝑎𝑙 × 𝑅𝐻 𝑅𝑒𝑎𝑙) + (𝑅𝐻𝐼𝑚𝑎𝑔𝑖𝑛𝑎𝑟𝑦× 𝑅𝐻𝐼𝑚𝑎𝑔)
𝐶12 𝑟𝑒𝑎𝑙 = (𝑅𝐻 𝑅𝑒𝑎𝑙 × 𝑅𝑉𝑅𝑒𝑎𝑙) + (𝑅𝐻𝐼𝑚𝑎𝑔× 𝑅𝑉𝐼𝑚𝑎𝑔)
𝐶12 𝑖𝑚𝑎𝑔 = (𝑅𝐻𝐼𝑚𝑎𝑔 × 𝑅𝑉𝑅𝑒𝑎𝑙) − (𝑅𝐻 𝑅𝑒𝑎𝑙× 𝑅𝑉𝐼𝑚𝑎𝑔)
𝐶22 = (𝑅𝑉𝑅𝑒𝑎𝑙 × 𝑅𝑉𝑅𝑒𝑎𝑙) + (𝑅𝑉𝐼𝑚𝑎𝑔× 𝑅𝑉𝐼𝑚𝑎𝑔)
𝐶 =
𝐸 𝑅𝐻 𝐸 𝑅𝐻
∗
𝐸 𝑅𝐻 𝐸 𝑅𝑉
∗
𝐸 𝑅𝑉 𝐸 𝑅𝐻
∗
𝐸 𝑅𝑉 𝐸 𝑅𝑉
∗
• Two channel data – i.e. RH and RV
• Supplied as SLC data(complex) 16
bit integer values
• After conversion to float C2 is
calculated:
6. Classification Methods
APSAR 2013 Tsukuba Japan - Sep. 23 - 27, 2013.
• Wishart supervised classifier
• Compute the mean covariance matrix (C2) over the training areas
𝐶 𝑚 = E Z 𝑚 ]
This is the mean covariance matrix for class 𝑚 .
• The complex Wishart distrubution is given by:
• The distance dm is computed for each pixel, for each class
• The pixel is assigned to the class with the minimum distance
7. Classification Methods
APSAR 2013 Tsukuba Japan - Sep. 23 - 27, 2013.
• SVM (Support Vector Machine)
• It is based on search of optimal hyperplane which can separate the classes.
• The SVM makes the use of non-linear function which transforms the data
from input space to higher dimension feature space so that the data can be
linearly separable.
• Various kernels may be used:
• Linear
• RBF
• Polynomial
8. Objectives of Study
• Backscattering coefficient (σ0) for discrimination of
various land features using both linear and hybrid
polarimetric RISAT-1 data
• Compared classification accuracy using RADARSAT-
2 simulated hybrid and RISAT-1 compact pol data
APSAR 2013 Tsukuba Japan - Sep. 23 - 27, 2013.
9. Study Area: Mumbai, India.
Scene Center
Longitude:72.930005
Latitude :19.220882
RISAT-1
RH/RV – FRS
15th Nov 2012
APSAR 2013 Tsukuba Japan - Sep. 23 - 27, 2013.
10. Study Area: Mumbai, India.
RISAT-1
RH/RV – FRS
15th Nov 2012
Test site chosen is the metropolis of Mumbai, India.
The area consists of:
• Built-up dense urban settlements
• Moderately dense deciduous forest
• Mangroves
• Wetlands
• Bare soil
• Water
• Grasslands
Urban Areas
Courtesy: indianexpress.com
Forest Mangroves
APSAR 2013 Tsukuba Japan - Sep. 23 - 27, 2013.
11. Study Area: Mumbai, India.
RISAT-1
RH/RV – FRS
15th Nov 2012
Wetland / Saltpan
Bareland Water
Grassland
APSAR 2013 Tsukuba Japan - Sep. 23 - 27, 2013.
12. Data sets and Field data collection
RISAT-1 data has been acquired on two successive days over Mumbai.
Satellite Mode Date of Acquisition Incidence angle
RISAT-1 HH/HV 14th Nov 2012 49.3
RH/RV 15th Nov 2012 35.9
RADARSAT-2 Full Pol. 16th Feb 2011 41.73
Ground-truth parameters in terms of soil moisture, vegetation height and
biomass, etc. were collected synchronous with the satellite passes.
APSAR 2013 Tsukuba Japan - Sep. 23 - 27, 2013.
14. Methodology
RISAT-1 Data
• cFRS-1 [RH/RV]
• FRS-1 [HH/HV]
Pre-processing
• Data extraction
• Calibration
Multilook
• 3:3 in Range: Azimuth
Compute statistics for 6 test areas
Plot Backscattering Coefficient
APSAR 2013 Tsukuba Japan - Sep. 23 - 27, 2013.
15. RISAT-1 0 ԁB Analysis
Class HV HH RV RH
Grass-land -12.48 -4.13 -7.47 -3.35
Bare-land -14.67 -5.88 -10.56 -6.47
Water -17.62 -11.53 -15.12 -11.64
Mangroves -11.30 -3.28 -6.37 -2.79
Forest -12.76 -4.20 -7.18 -4.27
Urban -13.15 -1.87 -5.87 -0.28
Wetland -16.99 -10.57 -10.69 -9.54
AVERAGE 0 ԀB VALUES FOR LINEAR AND HYBRID MODE DATA
APSAR 2013 Tsukuba Japan - Sep. 23 - 27, 2013.
16. RISAT-1 0 ԁB Analysis (Cont.)
Mean and standard deviation of σ0 dB of RISAT-1 linear and hybrid mode data for various classes.
APSAR 2013 Tsukuba Japan - Sep. 23 - 27, 2013.
17. RISAT-1 0 ԁB Analysis (Cont.)
• There is a clear separation of mean 0 values of various classes
• Yet, we can not classify data on backscatter alone
• Standard deviation of features is high
• Overlaps with mean values of other features
• Example: forest and mangrove class overlap
• The standard deviation from mean is consistent in all classes
• Value ranging from 2.18 in water to 2.73 in the forest class
• Exception: urban class - higher standard deviation of 4.32
• There is a 13.4o difference in the incidence angle between the
RH/RV and HH/HV datasets from RISAT-1 : This may be the
reason for the difference in mean σ0
APSAR 2013 Tsukuba Japan - Sep. 23 - 27, 2013.
23. Combining CPR, SPAN and m-δ / m-
APSAR 2013 Tsukuba Japan - Sep. 23 - 27, 2013.
Histogram of CPR for test areas
• The CPR, SPAN and individual
components of the m-δ / m-
decompositions (Vol, Dbl, Surface)
are normalized and used as input
bands to the SVM classifier.
• CPR helps discriminate between
mangroves and forest areas (see
histogram)
• SPAN helps discriminate urban
areas from background.
24. Classification Results – Test Area
Class
RISAT-1 Hybrid Pol
RISAT-1
Dual Pol
RADARSAT-2 Simulated Hybrid Pol
Wishart
m-,
CPR,SPAN
(SVM)
m-χ,
CPR,SPAN
(SVM)
Wishart Wishart
m-,CPR-
SPAN (SVM)
m-χ,
CPR,SPAN
(SVM)
Water % 100.00 100.00 100.00 65.15 100.00 100.00 100.00
Mangroves % 73.87 76.78 77.41 37.21 67.84 60.29 56.96
Urban % 78.64 91.56 97.85 69.95 75.25 71.03 73.43
Forest % 86.52 99.34 99.57 82.74 45.58 45.01 41.23
Wetland % 91.57 94.45 94.77 48.67 98.83 97.97 98.93
Grassland % 78.16 84.66 85.44 32.46 45.00 57.09 60.96
Overall User
Acc. %
84.67 91.61 92.84 58.57 68.45 68.17 67.69
Classification accuracy for various land covers using test areas
APSAR 2013 Tsukuba Japan - Sep. 23 - 27, 2013.
25. Results and Analysis
• Wishart supervised classifier:
• RISAT-1 – (RH/RV)
• 80.62% for training areas
• 84.67% for test areas.
• RADARSAT-2 (Simulated RH/RV)
• 72.83% for training areas
• 68.45% for test areas
• The classification accuracy increases by 7% after combining the three
components of m-χ or m-δ with the CPR and SPAN [3] for RISAT-1.
• RISAT-1 hybrid polarimetric data performs better than RADARSAT-2
simulated hybrid polarimetric data for all three combinations.
• The lowest classification accuracy of 32.46% for the grassland class is due
to its confusion with forest class.
[3] V. Turkar, Shaunak De, Y. S. Rao, A. Bhattacharya and A. Das, “Comparative Analysis Of
classification Accuracy For RISAT-1 Hybrid Pol. Data”, Proc. IEEE IGARSS 2013, Melbourne.
APSAR 2013 Tsukuba Japan - Sep. 23 - 27, 2013.
26. Classified Images of RISAT-1
Classified image of RISAT-1 C-band Hybrid polarimetrc Mumbai data (a) Wishart supervised (b)SVM classified (m-χ + CPR + SPAN)
APSAR 2013 Tsukuba Japan - Sep. 23 - 27, 2013.
Wishart Supervised SVM (m-χ + CPR + SPAN)
Legend
Water
Mangroves
Forest
Urban
Wetland
Grassland
28. EFFECT OF TRAINING AREA
SELECTION
Comparison of classification
RISAT-1 (RH/RV)
APSAR 2013 Tsukuba Japan - Sep. 23 - 27, 2013.
29. Classification – with different training areas
APSAR 2013 Tsukuba Japan - Sep. 23 - 27, 2013.
Classification accuracy for various land covers using test areas
Class
RISAT-1 Hybrid Pol
Large Training Area
RISAT-1 Hybrid Pol
Small Training Area
Wishart Wishart
Water % 100.00 100.00
Mangroves % 73.87 84.67
Urban % 78.64 91.66
Forest % 86.52 76.47
Wetland % 91.57 93.16
Grassland % 78.16 88.79
Overall User Acc. % 84.67 89.12
31. Confusion Matrix –Small training areas
APSAR 2013 Tsukuba Japan - Sep. 23 - 27, 2013.
RISAT-1 RH/RV Wishart
Class Urban Forest Mangroves Water Wetland
Urban 97.95 0.34 0.73 0 0
Forest 0.28 88.33 11.2 0 1.59
Mangroves 1.78 9.9 88.06 0 0
Water 0 0 0 100 1.59
Wetland 0 1.43 0 0 96.81
Overall Accuracy: 94.23
32. Confusion Matrix –Large training areas
APSAR 2013 Tsukuba Japan - Sep. 23 - 27, 2013.
RISAT-1 RH/RV Wishart
Class Urban Forest Mangroves Water Wetland
Urban 85.97 0 0 0.63 13.4
Forest 0 92.59 0.71 6.7 0
Mangroves 0 13.42 85.58 1 0
Water 0 8.12 0.09 91.44 0.35
Wetland 12.78 0 0 2.3 84.92
Overall Accuracy: 88.10
33. Conclusion
• Mean and standard deviation values follow the same trend for both the
imaging modes: linear and hybrid
• Urban class exhibits higher standard deviation from mean
• The horizontally polarized receive components, HH and RH are higher than
their respective vertically polarized receive components, HV and RV
• The performance of hybrid polarimetric (RH,RV) data in terms of classification
accuracy is better than dual polarization (HH, HV) data
• The classification accuracy increases by combining three components
(surface, double and volume) of m-χ or m-δ along with CPR and SPAN for
RISAT-1 and RADARSAT-2 hybrid polarimetric data.
• RISAT-1 hybrid polarimetric data classification accuracy is better than
simulated hybrid polarimetric data from RADATSAT-2.
APSAR 2013 Tsukuba Japan - Sep. 23 - 27, 2013.