基于多种极化分解方法和全极化合成孔径雷达数据的干旱区盐渍化监测

土壤盐渍化不仅严重制约着干旱区农业可持续发展，并且对绿洲生态环境构成了重大威胁。该文提出将多种极化目标分解方法、绕封模型（wrapper feature selector, WFS）特征子集选择方法和支持向量机（support vector machine, SVM）结合起来（简称WFS-SVM），利用全极化合成孔径雷达（polarimetric synthetic aperture radar, PolSAR）数据实现对土壤盐渍化监测。以新疆干旱区典型绿洲--渭干河-库车河三角洲绿洲（渭-库绿洲）为研究区，对研究区四极化Radarsat-2数据进行多种极化目标分解处理，得到相应的特征参数和特征分量。采用WFS方法进行SVM最佳特征子集的选择，并选出最佳适应度的子集对 SVM进行训练。从而构建基于最佳特征子集和最优分类参数的WFS-SVM分类模型，对研究区进行不同程度盐渍地信息（包括重度盐渍地和中-轻度盐渍地）的提取，并结合野外实地考察验证数据，将分类结果与经典的Wishart监督分类方法和一般SVM分类方法进行了对比和验证。结果表明，该方法较大程度地提高了全极化PolSAR影像在干旱区盐渍地信息提取的精度，相比Wishart监督分类，该方法分类总精度和Kappa系数分别提高了14.12个百分点和0.18，证明了该文所提出的监测方法具有有效性和研究潜力。该成果也将促进PolSAR数据在干旱区盐渍化监测中发挥更大的作用。

Monitoring soil salinization in arid area using PolSAR data and polarimetric decomposition method

Abstract: Soil salinization, one of the most widespread soil degradation processes on earth, has become critical problem which constrains long-term sustainable development of agriculture in arid and semi-arid areas, and it poses major threat to ecosystem stability of oases. It is, therefore, indispensable and of utmost importance to real-time monitoring and evaluating of soil salinity in those regions. The main objective of this study is to examine a method which combines polarimetric target decomposition, optimal feature subset selection (wrapper feature selector, WFS) and support vector machine (SVM) algorithms for monitoring soil salinization using fully polarimetric synthetic aperture radar (SAR) data in arid area of Xinjiang. This study has developed and successfully applied methodologies which integrate polarimetric decomposition, WFS and SVM algorithms for classification and extraction of salinized soils using quad-polarized RADARSAT-2 polarimetric SAR (PolSAR) image. A threefold exercise is conducted. Firstly, fully polarimetric Radarsat-2 data in ascending passes are acquired over the study area (on the delta oasis between the Weigan and Kuqa River in northwest of Xinjiang, China) on July 4, 2014, and multiple polarimetric decomposition methods (Pauli, Freeman 2, Freeman 3, Barnes 1, Barnes 2, Holm 1, Holm 2, Cloude, Huynen, Yamaguchi 3, Yamaguchi 4, VanZyl 3, Krogager, Touzi, Neumann 2, H/A/Alpha) are implemented to support the classification of PolSAR data, whose aim is to extract polarimetric parameters and SAR discriminators related to the physically scattering mechanisms of the observed objects. Next, the WFS which utilizes genetic algorithms is adopted to select and provide the best feature subset to implement SVM classification, a cross-validation method is employed to identify the optimum classification parameters and obtain an optimal SVM classification model, and then the SVM classifier is trained with the training samples acquired in the field investigation. Finally, a WFS-SVM classification model is constructed, optimized and implemented based on optimal match of polarimetric features and optimum classification parameters, and the soils with different salinization degrees (strongly and moderately-slightly salinized soils) are extracted; then a comparison between the proposed method, the Wishart supervised classification which is based on the coherency matrix and conventional SVM classification without feature selection is made to test the performance of the proposed method, and the detailed field observations and ground data are used for the validation of the adopted methodology. The results showed that: 1) The optimal match of selected polarimetric features of the Radarsat-2 data over the study area by using WFS algorism was 16 polarimetric parameters including T11, C11, C33, Pauli_a, Fee3_Vol, Free3_Odd, Cloude_T11, Huy_T11, VZ3_Vol, VZ3_Odd, Krog_KS, Krog_KD, Neu2_delta_mod, pH value, RVI (radar vegetation index) and SPAN. 2) Compared with the Wishart classification and the commonly used SVM algorithm without feature selection, a significant improvement on the overall classification accuracy could be observed by the WFS-SVM approach with the best overall classification accuracy of 84.94% (Kappa: 0.80), while those by the Wishart and the SVM were 70.82% (Kappa: 0.618) and 82.87% (Kappa: 0.774) respectively. This represented 14.12% (Kappa: 0.18) and 2.07% (Kappa: 0.027) improvement over that of Wishart and SVM classification, respectively. Furthermore, the extraction accuracy of different salinized soils was also enhanced with the proposed methodology; extraction accuracy of strongly salinized soil was increased to 80.57% while that of Wishart and SVM classification was 58.42% and 77.22%, respectively, and extraction accuracy of moderately-slightly salinized soils was improved from 58.27% and 83.15% to 85.39%. All in all, the classification results demonstrate the potential and effectiveness of the proposed WFS-SVM classification scheme for implementing classification and monitoring of soil salinization, which can play a vital role in monitoring and mapping of soil salinity by using PolSAR data in arid and semi-arid regions.