Alos PALSAR polarimetric data for land land cover classification in Amazon.

Size: px

Start display at page:

Download "Alos PALSAR polarimetric data for land land cover classification in Amazon."

Garey Montgomery
5 years ago
Views:

1 Alos PALSAR polarimetric data for land land cover classification in Amazon. An emphasis on phase information. Investigators: Luciano Vieira Dutra Corina C. Freitas Associate Res: Graziela B Scofield MS Students: Rogério G. Negri Sumaia Aboud-Neta Daniel Andrade.

2 Objectives Assessment of the Full Polarimetric data and/or derived features for land cover classification, specially the following: Polarimetric coherence magnitude and phase; Interferometric coherence magnitude; Complex multi-look data; Phase itself and fusioned; Several types of classifiers; Performance of standard and specialized; Discrimination of certain classes: recent deforestation. degraded forest, agriculture.

3 Study area: Tapajós National Forest. south of Santarém in the Brazilian Amazon. Pará State. Data: Level 1.1 PLR images (March 8 *, 2007 and April 23, 2007) Landsat-5 (June 21, 2007) * No classification

4 Field Work 2005 (updated by field reports and visual interpretation) Primary forest. degraded forest (2 years ago). forest under exploitation. recent deforestation(2 years ago). clean pasture. bare soil.

5 Pixel Based Classifiers Standard Classifiers. Standard (Gaussian) ML/ICM with equal a prioris for 3 or more channels cases; SVM Support Vector Machines. Radial Basis Functions Kernel Special pdf cases for SAR data ML and ICM; Two intensity polarimetric channels; From Wishart Distribution => Bi-variate Gamma (Lee et al, 1994) Phase Difference using coherence matrix; Also from Wishart Distribution (homogenous area) S r = SHH 2SHV S Wishart, K or G0 Distributions (Homogeneity dependent) Full Polarimetric. [ ] T VV

6 Palsar color composite: HH-HV-VV (RGB) and training classes. March 8, 2007 April 23, 2007

7 Overview of Polarimetric Coherence Maps March 8, 2007c April 23, 2007 June 21, 2007 Coherence HH1_VV1 Coherence HH2_VV2 LANDSAT 5 (5)R(4)B(3)G

8 HH-VV Coherence and Phase Information March 8, 2007 March April 0.56 Pasture Soya HH,HV,VV (R,G,B) 0.40 Secondary Forest 0.40

9 HH-VV Phase and Coherence Information (continued) April 23, 2007 March April 0.38 Primary Forest Soya HH,HV,VV (R,G,B) 0.63 Bare Soil 0.46

10 Fusion of Total Power, HH-VV Phase Diff and PolCoh info.

11 Two color compositions compared

12 Multi-date Polarimetric Coherence Composition Composition: R (5),G(4),B(3) Composition : R (HH1_VV1),G (HH2_VV2), B(HH1_VV1) Images : (1) March 8, 2007; (2) April 23, 2007

13 Overview of Interferometric Coherence Maps Coherence HH1_HH2 Coherence VV1_VV2 Average Coherence (HV1_HV2 + VH1_VH2) / 2 R (HH1_HH2), G (VV1_VV2) B ((HV1_HV2 + VH1_VH2) / 2)

Main SVM results 4 best from 28 classifications Type of Data Channels (all

First-others 2 Intensities; InCoh HH,HV,HH_HH 70.57 0.6549 Soya1 (89.

0) < 10-7 3 Intensities HH,HV,VV 64.82 0.5875 Soya1( 93.31) Second (39.

14 Main SVM results 4 best from 28 classifications Type of Data Channels (all pre-filtered) Overall Accuracy (%) Kappa Best User Ac Worst User Ac P value First-others 2 Intensities; InCoh HH,HV,HH_HH Soya1 (89.5) Second (45.8) 2 Intensities HH,HV Soya1 (93.3) Second (39.0) < Intensities HH,HV,VV Soya1( 93.31) Second (39.1) < Int, InCoh, PolCoh HH,HV, HH_HH,HH_VV Soya2 (80.4) Second (45.9) < 10-9

Main ML/ICM results 4 best* from 18 classifications Type of

Ac Worst User Ac P-value First-others 2 Intens,InCoh

0) 2 Intensities InCoh, PolCoh HH,HV, HH_HH,HH_VV 79.23 0.

3 Full Pol Multilook matrix S r mlc 78.78 0.

15 Main ML/ICM results 4 best* from 18 classifications Type of Data Channels (all ICM) Overall Accuracy (%) Kappa Best User Ac Worst User Ac P-value First-others 2 Intens,InCoh HH,HV,HH_HH Soya1 (99.7) Second (52.0) 2 Intensities InCoh, PolCoh HH,HV, HH_HH,HH_VV Soya1 (98.1) Second (44.4) 0.3 Full Pol Multilook matrix S r mlc Bare Soil (100) Degraded (50.0) Intensities HH,HV Bare Soil (99.8) Second (38.5) < *All ICM

16 Conclusions These conclusions must be considered for this seven classes scenario. Use of Interferometric Cohrence alone presented the best results for both classifiers addition of PolCoh worsened the accuracy indication that intrinsic dimensionality is 3 with InSAR. When using only Intensity channels it is better to use the proper distributions. HH-HV is a good choice with no InSAR products. SVM which is not a contextual classifier delivered worse results but consistent with ICM results. Phase presents discriminatory info specially for Soya, and it is taking in account when using Multi-look complex data. Two years degraded forest left to regenerate began to confuse with regeneration.

17 Acknowledgements Special thanks to JAXA and ASF for providing imagery AO-108 Project.

18 Melhor resultado SVM ROIs 12 SVM_comfiltro_INTHH_IntHV_coeintHH1_HH2_class P F1 SE AG1 AG2 FD RG user ac P , pasto F , floresta primaria Kappa SE , solo AG , Var Kappa AG , FD ,65 floresta degradada Overall RG , sec pior

19 Melhor resultado do ICM ROIs 14 HHHVInt(HH_HH)-conf-ICM P F1 SE AG1 AG2 FD RG CLASSES USUARIO P P 89,94 F F1 68,04 Kappa SE SE 96,12 AG AG1 99,74 Var Kappa AG AG2 92,42 FD FD 58,23 Overall RG RG 51,

20 SVM Results (no filter) Filtering (Y/N) Data Channels Overall Accuracy (%) Kappa N i,i,icoh(1_2) HH,HV,HH_HH N i,i,icoh(1_2), PolCoh2 HH,HV, HH_HH,HH_VV N i,i HH,HV N i,i,i HH,HV,VV

21 Maxver Results (no context) Filtering (Y/N) Data Channels Overall Accuracy (%) Kappa N i,i,icoh(1_2), PolCoh2 HH,HV, HH_HH,HH_VV N i,i,icoh(1_2) HH,HV,HH_HH N i,i,i HH,HV,VV N i,i HH,HV

HH polarization 1=Primary Forest, 2= Recent deforestation, 3= Clean pasture, 4= Selective Logging Gray levels 50000 45000 40000 35000 30000 25000 20000 15000 10000 5000 0 1 2 3 4 Class RGB = HH-VV-HV

22 HH polarization 1=Primary Forest, 2= Recent deforestation, 3= Clean pasture, 4= Selective Logging Gray levels Class RGB = HH-VV-HV VV polarization return. 1=Primary Forest, 2= Recent deforestation, 3= Clean Pasture; 4= Selective Logging Gray levels Class HV polarization return. 1=Primary Forest, 2= Recent deforestation, 3= Clean Pasture, 4= Selective Logging Gray Levels Class

23 Alos PALSAR polarimetric data for land land cover classification in Amazon. An emphasis on phase information. Luciano Vieira Dutra. Graziela Balda Scofield Rogério Galante Negri Sumaia Resegue Aboud Neta Daniel Andrade Corina da Costa Freitas

Review. Guoqing Sun Department of Geography, University of Maryland ABrief

Review. Guoqing Sun Department of Geography, University of Maryland ABrief Review Guoqing Sun Department of Geography, University of Maryland gsun@glue.umd.edu ABrief Introduction Scattering Mechanisms and Radar Image Characteristics Data Availability Example of Applications