Deep PPL-SAR Terrain Classification Method Combined With Spatial Information

Polarimetric Synthetic Aperture Radar(POLSAR)has been one of the most advanced technologies in remote sensing.Correspondingly,it is very urgent to interpret the POLSAR data accurately,where POLSAR image classification is an important issue.The traditional ways of POLSAR image classification are efficient to some extent,which make full use of the polarimetric information and classify the POLSAR data pixel-by-pixel.It is noticed that spatial information plays a significant role in natural image processing.In the task of POLSAR image classification,the performance of classification methods would be further improved if the spatial information is taken into consideration.In our work,both polarimetric and spatial information are used together to complete the POLSAR image classification,where multi-layer features are extracted by deep learning and the classification is acomplished by a simple classifier.There are three main contributions,as listed below.1.The Local Binary Pattern(LBP)feature,which contains spatial information,is fined and confused with the polarimetric information.Then the popular Deep Belief Network(DBN)is used to learn multi-layer features,and a one-layer classifier is employed for classification.Unlabeled pixels are fully taken advantage of to reduce the necessary of abundant labeled pixels,and in the meanwhile the classification performance can be guaranteed.2.Even though spatial information is used in the previous item(method),it still is pixel-based and a mass of repeated computation is necessary,which leads to a high cost naturally.To reduce the cost,an efficient method based on Convolutional Neural Network(CNN)and Simple Linear Iteration Clustering(SLIC)is proposed.A preliminary classification result is generated by CNN,where the spatial information is captured by the kernels in CNN.Then SLIC is used to clean the preliminary result caused by noises.3.However,the downsampling operation in CNN gives rise to a loss of details.For example,the edges and noises in the POLSAR image often lead to an entire mistakes of the neighboring pixels.Therefore the pixel-based idea is re-considered for better performance.Specially high-resolution POLSAR image is generated with the help of an Auto Encoder(AE),and then it is classified by the trained DBN,which is obtained in the first method.The final classification result is achieved by a downsampling operation on the high-resolution result.