PolSAR Image Terrain Classification Based On CNN Features Learning And SVM

Polarimetric Synthetic Aperture Radar(Pol SAR) obtains polarization information of the target by measuring the polarimetric scattering echo of each resolution cell in orientation and range. Compared with the conventional SAR, Pol SAR has the advantage of obtaining more abundant scattering information of targets, which improves the ability of terrain classification and object recognition. Therefore, Pol SAR image terrain classification has played a more and more significant role in the Pol SAR information processing field. Its aim is to make the full use of Pol SAR images' measured data, such as scattering matrix, polarization coherent matrix, etc. to label each pixel in Pol SAR images.Currently, Pol SAR image terrain classification algorithm is regarded as the classification of the characteristics directly through shallow learning. This kind of method only gets the low-level features and not fully combines polarization information and spatial correlation of Pol SAR images. These features are not inadequate to reveal key information of Pol SAR images, and then this would lead to less accurate results when finally the images are classified. Generally, two vital factors infulencing the result of Pol SAR image terrain classification are feature learning and classifier selection. Based on the above, the following work has been done in this thesis:1. A Pol SAR image terrain classification method based on convolutional features whose convolutional layer's parameters learned by Auto Encoder(AE) and Support Vector Machine(SVM) is proposed. This algorithm takes the advantage of image spatial correlation, which ensures the relative position unchanged in the local neighborhood pixels. Each pixel is input to the convolutional neural network in the form of an image, and the convolutional layer's parameters are obtained by training AE network. In this way, it maps the original feature space to the deep feature space. Compared with other shallow or deep feature learning methods, it extracts more advanced and effective features, and then utilizes efficient SVM classifier to classify, thus improving the classification accuracy.2. A Pol SAR image terrain classification method with pooling features pooled by Principal Component Analysis(PCA) and SVM is proposed. Due to the local pooling operation for each convolution map obtained by convolution operation with the first method, deep convolutional network and SVM would take a long training time. To solve this problem, PCA is utilized to do the pooling, so each convolution map can be pooled as a whole. In this way, it can shorten training and test time in experiments. Experiments results have shown the effectiveness and robustness to parameters.3. A Pol SAR image terrain classification method based on Turbo Pixels segmentation and SVM is proposed. From the experiment results of the above two methods, it can be observed that noised points exist obviously, which severely affecting classification accuracy. To solve this problem, we combine SVM with Turbo Pixels algorithm which has low complexity to parameters learning. First, we use Turbo Pixels to segment Pol SAR images, and then pick up some pixels from superpixels as representatives to do CNN features learning. Finally, SVM classifier is used to classify these representatives, and then the results are to label superpixels by voting. The experiment results of three Pol SAR images have shown that this algorithm not only significantly improves the classification accuray, but also greatly shorten experimenting time.