Download - POLSAR CHANGE DETECTION
LAND COVER CHANGE DETECTION USING RADARSAT-2 POLARIMETRIC SAR IMAGES
Zhixin Qi and Anthony Gar-On YehThe University of Hong Kong, Hong Kong, China
Introduction1
Study area and data2
Methodology3
Results and discussion4
Conclusions5
Outline
Background
There are many illegal land developments in some of China’s rapidly developing regions, such as the Pearl River Delta (PRD).
RADAR vs. Optical remote sensing
Radar remote sensing, which is not affected by cloud conditions, is promising for monitoring short-term land cover changes.
Multi-polarization vs. single-polarization
Single-polarization SAR
HH
Polarimetric SAR (PolSAR)
VH
VVHH
HV
polarization
The polarization information contained in the waves backscattered from a given medium is highly related to:
• its geometrical structure reflectivity, shape and orientation• its geophysical properties such as humidity, roughness, …
H
V or H
H
H
Completely Polarised Scattering Partially Polarised Scattering
Study Objective
Research questions
• Unsupervised methods
– They cannot determine types of changes.
• Post-classification comparison (PCC)
– Poor accuracy of PolSAR image classification caused by the limited spectral information and speckle noise
• Pixel-based methods
– They may cause false alarms due to the speckle effect.
Study objective
• This study aims to develop a new method that integrates change vector analysis (CVA) and post-classification comparison (PCC) with object-oriented image analysis (OOIA) to detect land cover changes from RADARSAT-2 polarimetric SAR (PolSAR) images.
Study Area
Land Cover Classes in the Study Area
Study Data
• RADARSAT-2 Fine Quad-Pol images (Single Look Complex).
• Full polarization: HH, HV, VH and VV.
• Incidence angle: 31.50°.
Field Work
Field Work
Class Sub-class Plots Pixels
Change Barren land to crop/natural vegetation 68 28,995
Water to crop/natural vegetation 47 19,089
Crop/natural vegetation to water 75 17,738
Barren land to built-up areas 51 14,257
Barren land to water 41 9,417
Water to lawns 7 4,089
Crop/natural vegetation to barren land 10 3,392
Water to barren land 9 3,380
Total 308 100,357
No change Banana 107 41,656
Barren land 84 23,743
Forests 118 36,437
Lawns 98 34,159
Crop/natural vegetation 202 93,196
Built-up areas 224 65,939
Water 130 66,130
Total 963 361,260
Number of change and no-change samples selected for the verification of land cover change detection results
Methodology
Land cover change detection using two PolSAR images acquired over the same area at different times
• Polarimetric decomposition can be used to support the classification of PolSAR data. It is aimed at extracting polarimetric parameters related to the physical scattering mechanisms of the observed objects.
• In this study, PolSARPro_4.03 software package was used to implement polarimetric decomposition.
Polarimetric Decomposition
Polarimetric Decomposition
Polarimetric Decomposition
Decomposition Methods
• Pauli (Cloude & Pottier, 1996)• Barnes (Barnes, 1988)• Huynen (Huynen, 1970)• Cloude (Cloude, 1985)• Holm (Holm & Barnes, 1988)• H/A/Alpha (Cloude & Pottier, 1997)• Freeman 2 Components (Freeman, 2007)
• Freeman 3 Components (Freeman & Durden, 1998)
• Van Zyl (Van Zyl, 1993)• Neumann (Neumann et al., 2009)• Krogager (Krogager, 1990)• Yamaguchi (Yamaguchi et al., 2005)• Touzi (Touzi, 2007) methods
Image Segmentation
Determining the optimal scale for the segmentation of the Pauli RGB composition image of RADARSAT-2 PolSAR data.
Separate Segmentation
Hierarchical Segmentation
Hierarchical segmentation for delineating image objects from two successive RADARSAT-2 PolSAR images
A feature is an attribute that represents certain information concerning objects of interest, such as color, shape, and texture.
Change Vector Analysis (CVA)
Two images, image (t1) and image (t2), are acquired over the same area at different times t1 and t2. If k features are extracted from an image object, the feature vectors of the image object in the two images are given by X = (x1, x2, …, xk)T and Y = (y1, y2, …, yk)T respectively, the feature change vectors are defined as
where G includes all the change information between the two images for a given image object, and the change magnitude is computed with
The higher the is, the more likely that changes take place. Unsupervised classifiers or threshold methods are commonly applied on the change magnitude to identify changes.
kk yx
yx
yx
YXG
22
11
G
2222
211 )()()( kk yxyxyxG
G
(1)
(2)
Change Vector Analysis (CVA)
March 21, 2009 September 29,2009
Change magnitude Changed areas
PolSAR image classification
Methodology of land cover classification using RADARSAT-2 PolSAR images
Proposed Method Vs. WSC
Classification Results
Land Cover Change Results
Land cover change detection results (a)Proposed method (CVA, PCC, and OOIA)(b)WSC-based PCC(c)PCC and OOIA (without CVA)(d)CVA and PCC (without OOIA)(e)CVA and OOIA (without PCC)
e
Accuracy AssessmentChange type Accuracy
statisticsProposed method(CVA, PCC, and OOIA)
WSC-based PCC CVA and OOIA (without PCC)
PCC and OOIA(without CVA)
CVA and PCC(without OOIA)
All the types DA (%) 86.71 94.94 90.51 93.15 88.13
FAR (%) 3.35 38.10 7.57 17.87 10.25
OER (%) 5.51 30.92 7.99 15.48 10.60
BL-CN DA (%) 46.50 33.23 NA 58.08 31.23
FAR (%) 0.11 0.50 NA 0.18 0.30
OER (%) 3.46 4.66 NA 2.80 4.59
BL-BU DA (%) 46.59 32.98 NA 48.93 45.46
FAR (%) 0.09 0.33 NA 0.09 0.87
OER (%) 1.73 2.39 NA 1.66 2.53
BL-W DA (%) 47.97 41.89 NA 48.32 44.88
FAR (%) 0.09 0.79 NA 0.20 0.40
OER (%) 1.15 1.96 NA 1.25 1.52
CN-BL DA (%) 73.88 34.08 NA 76.62 38.41
FAR (%) 0.85 0.85 NA 0.94 0.60
OER (%) 1.04 1.58 NA 1.11 1.05
CN-W DA (%) 52.06 37.38 NA 52.06 40.27
FAR (%) 0.05 0.27 NA 0.05 0.31
OER (%) 1.89 2.67 NA 1.89 2.59
W-BL DA (%) 68.76 51.30 NA 68.91 68.05
FAR (%) 0.86 1.33 NA 1.26 1.62
OER (%) 1.09 1.68 NA 1.48 1.85
W-L DA (%) 89.90 71.14 NA 89.90 47.79
FAR (%) 0.51 1.05 NA 0.51 0.60
OER (%) 0.59 1.29 NA 0.59 1.05
W-CN DA (%) 63.87 39.22 NA 63.87 38.40
FAR (%) 0.11 0.20 NA 0.11 0.16
OER (%) 1.60 2.70 NA 1.60 2.70
Conclusions
• The proposed method performs much better than WSC-based PCC in term of land cover change detection using RADARSAT-2 PolSAR images.
• The use of CVA before PCC can significantly reduce false alarms caused by the error of the classification of PolSAR images.
• Using PCC after CVA can reduce false alarms caused by environmental changes, such as seasonal vegetation growth and moisture variation. PCC that is based on the proposed classification approach, which integrates polarimetric decomposition, decision tree algorithms, and SVMs, achieves much higher accuracy than WSC-based PCC.
• OOIA reduces false alarms caused by speckles in PolSAR images and improves the accuracy of change type determination.
• Further investigation will be conducted to examine the effect of seasonal vegetation growth on the monitoring of human-induced land cover changes as well as how to distinguish between human-induced land cover changes and changes caused by seasonal vegetation growth.
Conclusions
• The proposed method performs much better than WSC-based PCC in term of land cover change detection using RADARSAT-2 PolSAR images.
• The use of CVA before PCC can significantly reduce false alarms caused by the error of the classification of PolSAR images.
• Using PCC after CVA can reduce false alarms caused by environmental changes, such as seasonal vegetation growth and moisture variation. PCC that is based on the proposed classification approach, which integrates polarimetric decomposition, decision tree algorithms, and SVMs, achieves much higher accuracy than WSC-based PCC.
• OOIA reduces false alarms caused by speckles in PolSAR images and improves the accuracy of change type determination.
• Further investigation will be conducted to examine the effect of seasonal vegetation growth on the monitoring of human-induced land cover changes as well as how to distinguish between human-induced land cover changes and changes caused by seasonal vegetation growth.