PolSAR Image Terrain Classification Based On Target Decomposition And Machine Learning

Pol SAR(Polarimetric Synthetic Aperture Radar) uses different polarization alternately way to transmit and receive radar signals in order to get a 2 by 2 complex matrix associated with a target. This matrix can reflect amplitude information, phase information and target scattering information. So, Pol SAR becomes an important means for earth observation. At present, Pol SAR applies widely in the field of object recognition and image interpretation. It includes urban planning, resource exploration, disaster monitoring, object detection, vegetation, and other precision strike military targets. In remote sensing image interpretation and information processing, terrain classification has become a popular topic in the world among scholars, but our Country is still at primary stage in interpretation of remote sensing image. This thesis starts work with two important aspects of target classification, including feature extract and classification. The main contents made as follows:Analysis the applicaiton of Pol SAR and the value of Pol SAR systems. Introduce the theory of Pol SAR image classification and history of Pol SAR, and also point out the severe challenge of Pol SAR image classification.A method for Pol SAR image terrain classification based on target decomposition and SVM is proposed. This method can solve a problem which traditional method only use one kind feature cannot get a satisfied result. We use three scattering characteristics by Cloude decomposition and four scattering characteristics by Freeman decomposition. We start train SVM after select correct training data and testing data. When the training is done, We start to predict the whole picture. The results show that this method is simple and efficient. Using a combination of features can get a 6.5 percent higher classification accuracy than using single feature.A method for Pol SAR image terrain classification based on target deposition and ELM is proposed. This method can solve a problem that traditional machine learning has to set a complex parameter and training for a long time. Learning fast and training for less time is the main feature for ELM. At the same time, ELM can get a comfortable accuracy.. Simulation results show that the ELM learning algorithm is indeed much faster than SVM, and accuracy of ELM is only slightly lower than SVM.