Study On Classification Method Using SAR Images For Vegetated Area Identification

Forest and agricultural vegetation monitoring provides information for management purposes. It is also a very important part in the global climate change and carbon cycle research. Remote sensing is an effective and timely approach to obtain wide-area land cover information. Remote sensing using Synthetic Aperture Radar has become a complementation to optical remote sensing due to its all time, all weather data acquisition capability and offering different information from the optical. With the emergence of many new spaceborne SAR systems, more possibilities are offered to the researchers and users than ever we had as well as the challenges for data processing and applications.With the availability of more and more SAR data, what are the best methods for data or feature selection and data processing to satisfy a specific application? This lies the start point of this research. For SAR image classification research, both the understanding of the physics and the development of the algorithm are important. The data and feature selection should be based on the understanding of the physical mechanisms of the interaction between the electromagenetic wave and the target. Aiming at the identification of the two vegetation types, namely forest and rice, research on land cover classification using SAR images was carried out, exploiting differnt SAR parameters including SAR backscatter intensity, InSAR coherence and the scattering mechanism information provided by SAR polarimetry. The contents of this paper are listed as follows:(1) Theory and principles of SAR remote sensing was introduced in order to understanding the information contained in the data, which is the basis for SAR image classification. The fundamental of SAR remote sensing is the interaction between the microwave and natural targets. The factors to influence the radar backscattering include both parameters of microwave and of targets. In this part, the scattering mechanisms were introduced thgether with the statistics of the backscattering, speckle and InSAR coherence, as well as the theory and rules for data preprocessing.(2)Study on the rule based classification using SAR data was carried out with the aim of forest identification. This kind of classification methods are based on the understanding of the physical processing behind the SAR observations.The multi-temporal, multi-incidence and multi-polarization SAR data at C and L band were analyzed to find the best suitable data and features for forest and non-forest separation. A hierarchy of rules were built based on the analysis. Overall accuracy of 78.36% was achieved with producer's accuracy and user's accuracy 62.02% and 82.12% correspondingly。The accuracy for coastal forest is as high as 93%。The proposed method has good potential for practical applications with these haigh mapping accracies.(3)Rice field backscattering behavior at L band, in particular the Bragg scattering, was analyzed and the corresponding rice mapping method was develped with ALOS PALSAR FBD data. Beginning with a finding of the enhanced HH backscattering of rice field in ALOS PALSAR images, the physical mechanism of this phenomena was analyzed and explained with the reason of the Bragg scattering. Based on the understanding of physics, the data selection and classification method were developed for rice mapping with L band SAR. In the test site, mapping accuracy of 86% was achieved. L band SAR can be used for rice mapping when the HV polarization is available to cope with the Bragg resonance effect, which is a progress of this paper compared to the former studies.(4)An object-oriented classification method for polarimetric SAR data was developed to deal with the speckle noise problem. A Gaussian distribution assumption is usually made for optical remote sensing image processing and analysis. In this paper, a polarimetric base transformation was introduced to change the covariance matric representing by complex numbers to nine intensity parameters, which can be expressed with real number and fit the Gaussian distribution. Combinning the polarimetric transformation technique and multi-resolution image segementation and object-oriented classification method, the object-oriented classification method was developed.