Dryland Crop Identification And Biological Parameters Estimation Based On Full-polarization SAR Data

Monitoring of the crop cultivated area and growth is a key element for agricultural remote sensing. Along with the development of remote sensing technology, wide-range monitoring of the crop cultivated area and growth has become possible. Monitoring results can provide great reference for national agricultural production and management and food policy making. At present, the Ministry of Agriculture mainly uses optical remote sensing data in its remote sensing monitoring of crops. However, in the monitoring of dryland crops in North China, because optical remote sensing is vulnerable to restrictions of frequent cloudy and rainy weather in the critical growth period of crops, it is difficult to obtain complete and continuous optical remote sensing data, which seriously affects the timeliness and accuracy of monitoring results. Synthetic Aperture Radar(SAR) remote sensing technology has the capacity of providing all-day and all-weather surface information, thus making up for the lack of optical remote sensing data in cloudy and rainy weather. Therefore, SAR has a broad market and huge potential of application. A lot of research has been conducted of the application of SAR in identification and growth monitoring of crops both at home and abroad. However, scholars have been mainly focused on paddy lands, but rarely focused on dryland. In this paper, a study is made of SAR-based identification and growth monitoring of dryland crops.In this research, Shenzhou of Hebei Province in the North China Plain is chosen as the study area. Multi-temporal and full-polarization Radarsat-2 SAR data is used to identify dryland crops and estimation biological parameters. Research is conducted mainly in the following 3 aspects: Firstly, combined with remote sensing and field survey data, backscattering characteristics of typical surface features in the study area are analyzed, indicators for classifying the surface features are established; some classification methods(the decision tree and support vector machine) are chosen for typical surface feature identification and accuracy verification. Moreover, a comparison is made of the crop identification accuracy of different methods. Secondly, the optimal phase of identifying dryland crops and its combination modes are analyzed, various auxiliary variables(polarimetric decomposition and texture) are extracted to identify typical surface features, the classification accuracy is taken as an evaluation index, the random forest method is adopted to evaluate the importance of each variable to the classification accuracy improvement. Thirdly, correlations between backscattering coefficients and crop growth parameters(crop height, leaf area index, the plant dry and wet weight and so on) are analyzed. A model of correlations between backscattering and crop growth parameters is built and used to retrieve crop growth parameters.Research results show that:1. When the SAR is used to identify dryland crops, the time phase of crops in the early growth stage should be highlighted. The optimal phase to identify maize is between the seedling stage and the early jointing stage, and cross-polarization is the optimal polarization method. HV data(June 27) is used to identify maize, with accuracy of over 80%. The optimal phase to identify cotton is between the seedling stage and the later budding stage, and cross-polarization is the optimal polarization method. Two-date HV data(July 21 and June 3) is used to identify cotton, with 73.31% accuracy. In the identification of dryland crops, the SVM is superior to decision tree, and the SVM enjoys obvious advantages in identifying small plots and controlling speckle noise.2. Multi-variable information is used to identify dryland crops in the study area. According to analytical results, in the maize identification in North China, polarimetric information makes a great contribution, polarimetric information brings a 7% increase over backscattering in the accuracy of classification result. Polarimetric information can increase the discrimination of maize from building lands. The introduction of texture information and polarimetric decomposition brings a 3% increase in the classification accuracy of cotton. Through variable selection for maize, it can be seen that 5 variable(VH、Alpha、Yamaguchi4-Odd、Freeman-Vol、Mean)combinations can ensure a higher accuracy of maize identification. However, a combination of all information still has optimal results.Comparisons of multi-variable information combinations show that multi-temporal combination modes have better classification results than mono-temporal combinations; the effect of polarimetric decomposition is better than texture information; a combination of optical and radar data produces optimal results.3. Correlation between radar data and crop growth parameters differs in different stages. In the early and later stage of crop growth stages, some growth parameters have significant correlations with radar data. In the medium stage, radar data is unsuitable for an retrieve of the crop growth parameters. Compared with co-polarization, cross-polarization is more sensitive to the crop growth parameters in the early stage of crop growth. The polarization ratio does not show much advantage in the retrieve biological parameters. An empirical formula is used for back calculation of the plant height, LAI, and the plant wet and dry weight. Moreover, retrieve results are verified and proved to be satisfactory. That is to say, the goal of maize condition monitoring is achieved.