Classification Of Polarimetric SAR Images Based On Deep Belief Networks

Polarimetric synthetic aperture radar(Pol SAR) has been one of important research in the SAR technology and has been applied in many areas such as agricultural production, urban planning and sea ice monitoring. Many domestic and foreign experts have proposed many methods based on supervised or unsupervised for Pol SAR image classification. However, these methods for Pol SAR image classification have many defects. For example, many methods only used physical scattering mechanisms information and scattering models are used alone to classify Pol SAR image. And to some extent it leads to object categories fuzzy. In addition, these methods can not extract features directly from the original Pol SAR data and need to manual carefully design many models and decomposition theory algorithms. These methods are usually time consuming and the calculation is very complex. Some of these methods can only deal with single pixel and can not capture the spatial structure information of the adjacent pixels.In the past 20 years, a large number of machine learning techniques have been applied to the polarimetric SAR image processing and have gained a good performance. Recently, the deep learning technology shows a great advantage on the areas such as speech recognition, natural language processing and natural image classification and many researchers are intersted in the deep learning technology. This paper combines the technology of Deep Belief Networks(DBNs) with scattering mechanisms to discuss the Pol SAR image classification.Firstly, third chapter of this paper, proposes a new Pol SAR image classification method based on DBNs with mutil-features integration. In this method we combine the coherent matrix elements, the parameters from H/a decomposition, the features obtained from gray level co-occurrence matrix with color histogram feature to train a DBNs model which consists with some restricted boltzmann machines(RBM)and this can efficiently overcome the drawbacks that the traditional neural netwoks easily converge to local optima and the computational complexity. Compared with the method based on neural networks(NN) and the method based on Support Vector Machine(SVM), the proposed method can achieve higher classification accuracy in the four-look polarimetric L-band Flevoland data set.In addition,the fourth chapter presents another method based on DBNs with learning features automatically. First, we convert the raw coherency matrix data to a 9-dimensional data; second, a large number of patches are randomly obtained from each dimension in the data, and then these image patches are converted in column-major order to a vector for training a RBM. We can get the structure features in each dimension of this 9-dimensional data and these features contain the relationships between adjacent pixels. Considering the scattering properties, the raw coherency matrix elements are combined with features learned by using the RBM to form the final feature sets to train a 3-layer DBNs for Pol SAR image classification. The proposed method considers the spatial structure information between adjacent pixels, therefore the image edge can be well retained and the proposed method can gain a better classification performance. Experiments are carried on the real Pol SAR data set and the results show that the proposed method has a good performance for Pol SAR image classification.