Article Information
- JunZhan Wang, JianJun Qu, WeiMin Zhang, KeCun Zhang . 2016.
- Classification of full-polarization ALOS-PALSAR imagery using SVM in arid area of Dunhuang
- Sciences in Cold and Arid Regions, 8(3): 0263-0267
- http://dx.doi.org/10.3724/SP.J.1226.2016.00263
Article History
- Received: February 12, 2016
- Accepted: April 22, 2016
Remote-sensing technology has been an important method for studying land use/land cover and land use/land cover change (LUCC), and classification is always a focus of field research. Classification for optical imagery has achieved tremendous success (Bastin,1997; Giacinto et al.,2000; Melgani and Bruzzone,2004; Emerson et al.,2005), with higher accuracy achieved because of optical imagery with rich spectrum and texture information. The application of optical imagery has some limitations due to cloud, rain, fog or other inclement weather conditions. The development of microwave remote sensing is an improvement to optical remote sensing, in particular, active microwave remote sensing provides unique advantages from new data sources for extracting land use/land cover information (Ferro Famil et al.,2001; Macrì Pellizzeri,2003; Wen et al.,2009; Triloki et al.,2010). In recent years, radar remote sensing has made great progress with successful launches such as ENVISAT-ASAR, ALOS-PALSAR, and RADARSAT-2. Also, the synthetic aperture radar (SAR) data mode has developed from the single-polarization and single-angle to multi-polarization and multi-angle. As an advanced instrument for remote sensing, polarimetric SAR has been applied widely in many fields, such as ecology, environmental monitoring, geological exploration, and vegetation investigation. Lee et al.(2001)compared classification results amongst single polarization, the standard (HH, VH) and (VV, HV) dual-pol modes, and quad-pol SAR imagery for P-, L-, and C-band frequencies. Their results show that various applications allows for optimally selecting the frequency and the combination of polarization. Ainsworth et al.(2009)compared classification results amongst standard dual-pol modes, compact polarimetric modes, and pseudo-quad-pol data imagery. The overall classification accuracy of the pseudo-quad-pol data is essential the same as the classification accuracy obtained directly employing the underlying dual-pol imagery. Recently, some theorems about polarimetric decomposition have been introduced (Cloude and Pottier,1996; Dong et al.,1998), which aim at establishing a correspondence between physical characteristics of the considered areas and the observed scattering mechanisms, where the results of the decomposition agree with the general understanding of radar backscatter. Moreover, classification techniques for agricultural areas have been developed, based on the decomposition results (Cloude and Pottier,1997). Compared with single-polarization and dual-polarization SAR data, the full-polarization SAR data includes four polarization bands, where new characteristic information can be extracted based on the full-polarization SAR data. The purpose of this paper is to explore the ability of the full-polarization SAR data and the new characteristic information in classification.
Support vector machine (SVM) based on statistical learning theory, proposed by Cortes and Vapnik (1995), is an effective supervised classifier. It is used widely in face recognition, hand writing identification, and automatic target recognition, which can achieved good classification performance with small training data sets. SVM has been a new focus in the field of machine learning. Several researchers have tried to use SVM for classifying SAR images, and obtained promising results (Georgios,2009; Wen et al.,2009). In this paper, the radial basis function (RBF) kernel was used for constructing the SVM classifier.
Compared with single- and dual-polarization SAR data, to what extent full-polarization SAR data can improve in classification is important. New characteristic information can be extracted by the difference operation, ratio operation, and principal component transform based on the full-polarization (HH, HV (or VH), VV) modes SAR data. It is necessary to explore the new characteristic information in improving classification accuracy. In this paper, full polarization L band PALSAR data was obtained, and classification performance of full-polarization and new characteristic information versus full-, dual-, single- polarization is compared qualitatively and quantitatively with SVM taken as the classifier.
2 Study area and data processingThe study area is part of Dunhuang city, western Gansu corridor, northwest China (Figure 1). Dunhuang city falls within an arid climatic zone with an annual average rainfall of 39.9 mm, but the annual mean amount of evaporation reaches to 2,486 mm. ALOS (Advanced Land Observing Satellite) was successfully launched by the Japan Aerospace Exploration Agency's (JAXA) on January 24,2006. ALOS carries three sensors:(1) the Panchromatic Remote-Sensing Instrument for Stereo Mapping (PRISM) for digital elevation mapping,(2) the Advanced Visible and Near Infrared Radiometer type 2(AVNIR-2) for land cover characterization, and (3) the Phased Array type L-band Synthetic Aperture Radar (PALSAR) for day-and-night and all-weather observation. PALSAR can operate at four primary modes with diverse polarizations and offnadir angles:(a) high-resolution single-polarization (FBS) mode,(b) high-resolution, dual-polarization (FBD) mode,(c) fully-polarimetric (PLR) mode, and (d) ScanSAR mode. The center frequency of PALSAR is 1,270 MHz, resulting in a wavelength of 23.62 cm. In this study, the fully-polarimetric mode data was obtained in 2007-05-13, this mode data has four polarization bands, which are HH, HV, VH and VV polarization, and the incidence angle is 8°-30°.
The PALSAR data is level 1.1 data, with the data pre-processing process presented in Figure 2. Finally, the backscattering coefficients imagery was obtained, and the resolution of the imagery is about 24 m. Figure 1b presents the combination of HH, HV and VV polarization.
Generally, because the mono station radar satisfies the reciprocity theorem, the backscattering coefficients of VH polarization are equal to HV polarization. The statistical characteristics of the backscattering coefficients of VH and HV polarization are presented in Table 1, and the three bands used for classification are HH, HV (or VH) and VV polarization. New characteristic information was extracted based on HH, HV (or VH) and VV backscattering data. Three new features were extracted by difference operations which are expressed as HH-HV, HH-VV, HV-VV. Three new features were extracted by ratio operation, which are expressed as HH/HV, HH/VV, HV/VV. Also, one new feature was extracted by principal component transform, where the first principal component (PC1) included the most information, thus PC1 was selected as another feature for classification. Next, the single-, dual-, full-polarization SAR data and new SAR characteristic information will be used to analyze quantitatively the classification accuracy based on SVM.
In this paper, the study area includes six classes as follow: farm land, building, water, Gobi, orchard and unused land. The same sample points of each class were used for each classification. The classification results were evaluated using overall accuracy and Kappa coefficient, for each class, the number of sample points as input in the SVM classification is presented in Table 2. The sample points are distributed in the study area as uniformly as possible, are manually extracted by visual interpretation, and used to calculate the overall accuracy. The RBF kernel used of SVM classifier in the ENVI software has two important parameters that need to be set, which are the kernel parameter γ and penalty parameter C. In this paper, in order to compare the classification results each other,γ was set at 0.1 and C was set at 100.
The imagery was classified based on the three kinds of PALSAR polarization data, which are HH, HV (or VH) and VV polarization, using SVM classifier. The sample points were used for accuracy evaluation. Results are presented in Table 3. This shows that VV-polarization data has better accuracy than HH and HV data.
Based on the HH, HV (or VH) and VV polarization data, three group dual-polarization data are generated, which are HH & HV, HH & VV and HV & VV. Using the same sample points, dual-polarization data was classified using SVM. Results are presented in Table 4. This shows that dual-polarization data has a much better accuracy than single-polarization, because it contains more backscatter information and texture information. The HV & VV data has the highest accuracy, the overall accuracy reached to 74.77%, kappa coefficient is 0.69.
Item | HH & HV | HH & VV | HV & VV |
Overall accuracy | 0.6253 | 0.6785 | 0.7477 |
Kappa coefficient | 0.53 | 0.59 | 0.69 |
Based on the HH, HV (or VH) and VV polarization data, three new features were extracted by difference operations, which are HH-HV, HH-VV, HV-VV. Three new features were extracted by ratio operation, which are HH/HV, HH/VV, HV/VV, and one new feature was extracted by principal component transform, which is PC1. Then, the full-polarization, full-polarization combination of new characteristic information was classified using SVM. Results are presented in Table 5. Compared with the classification results of single-, dual-polarization data, full-polarization PALSAR data and full-polarization PALSAR combination of the new features data can greatly improve classification accuracy. Compared with the classification result of full-polarization PALSAR data, new features can improve classification accuracy. The full-polarization PALSAR data combination of all the new features has the highest accuracy, the overall accuracy reached to 90.01%, kappa coefficient is 0.89. Thus, it is helpful for the classification of SAR data with rich polarization information or new information by bands math.
Item | Full-polarization | Full-polarization and difference | Full-polarization and ratio | Full-polarization and PC1 | Full-polarization, difference, ratio and PC1 |
Overall accuracy | 0.8021 | 0.833 | 0.8359 | 0.8435 | 0.9101 |
Kappa coefficient | 0.74 | 0.79 | 0.81 | 0.82 | 0.89 |
In this paper, classification performance of full-polarization and new characteristic information versus full-, dual-, single- polarization is compared qualitatively and quantitatively with SVM taken as the classifier. It is shown that single- polarization SAR data is poor in land use/cover classification because of the limited backscatter and texture information. Though VV polarization data has the highest accuracy, compared with dual- polarization (HV/VV) data, it is less than 18.24% in overall accuracy. Therefore, the dual- polarization data of HV & VV is the suitable choice for classification without full- polarization SAR data. Full- polarization PALSAR data has a better classification result, especially when adding new features by bands math or PCA.
Acknowledge:This work was supported by the National Natural Science Foundation of China (41401408,41371027). The authors would like to thank all the experts and editors.
Ainsworth TL, Kelly JP, Lee JS, 2009. Classification comparisons between dual-pol, compact polarimetric and quad-pol SAR imagery.. ISPRS Journal of Photogrammetry and Remote Sensing, 64(5): 464–471. doi: 10.1016/j.isprsjprs.2008.12.008 |
Bastin L, 1997. Comparison of fuzzy c-means classification, linear mixture modeling and MLC probabilities as tools for unmixing coarse pixels. International Journal of Remote Sensing, 18(17): 3629–3648. doi: 10.1080/014311697216847 |
Cloude SR, Pottier E, 1996. A review of target decomposition theorems in radar polarimetry. IEEE Transactions on Geoscience and Remote Sensing, 34(2): 498–518. doi: 10.1109/36.485127 |
Cloude SR, Pottier E, 1997. An entropy based classification scheme for land applications of polarimetric SAR. IEEE Transactions on Geoscience and Remote Sensing, 35(1): 68–78. doi: 10.1109/36.551935 |
Cortes C, Vapnik VN, 1995. Support vector networks. Machine Learning, 20(3): 273–297. |
Dong Y, Forster B, Ticehurst C, 1998. A new decomposition of radar polarization signatures. IEEE Transactions on Geoscience and Remote Sensing, 36(3): 933–939. doi: 10.1109/36.673684 |
Emerson CW, Lam NS, Quattrochi DA, 2005. A comparison of local variance, fractal dimension, and Moran's I as aids to multispectral image classification. Int. J. Remote Sens., 26(8): 1575-1588. |
Ferro-Famil L, Pottier E, Lee JS, 2001. Unsupervised classification of multifrequency and fully polarimetric SAR images based on the H/A/Alpha-Wishart classifier. IEEE Transactions on Geoscience and Remote Sensing, 39(11): 2332–2342. doi: 10.1109/36.964969 |
Georgios CA, 2009. SVM-based target recognition from synthetic aperture radar images using target region outline descriptors. Nonlinear Analysis, 71(12): e2934–e2939. |
Giacinto G, Roli F, Bruzzone L, 2000. Combination of neural and statistical algorithms for supervised classification of remote-sensing images. Pattern Recognition Letters, 21(5): 385–397. |
Lee JS, Grunes MR, Pottier E, 2001. Quantitative comparison of classification capability: Fully polarimetric versus dual-and single-polarization SAR. IEEE Transactions on Geoscience and Remote Sensing, 39(11): 2343–2351. doi: 10.1109/36.964970 |
Macrì Pellizzeri T, 2003. Classification of polarimetric SAR images of suburban areas using joint annealed segmentation and "H/A/a" polarimetric decomposition". ISPRS Journal of Photogrammetry & Remote Sensing, 58(1-2): 55–70. doi: 10.1016/S0924-2716(03)00017-0 |
Melgani F, Bruzzone L, 2004. Classification of hyperspectral remote sensing images with support vector machines. IEEE Transactions on Geoscience and Remote Sensing, 42(8): 1778–1790. doi: 10.1109/TGRS.2004.831865 |
Triloki P, Dharmendra S, Tanuja S, 2010. Advanced fractal approach for unsupervised classification of SAR images. Advances in Space Research, 45(1): 1338–1349. |
Vapnik VN, 1995. The Nature of Statistical Learning Theory. Springer Verlag. New York,: 1–50. |
Wen X, Zhang H, Zhang J, et al, 2009. Multi-scale modeling for classification of SAR imagery using hybrid EM algorithm and genetic algorithm. Progress in Natural Science, 19(8): 1033–1036. doi: 10.1016/j.pnsc.2009.01.003 |