Remote Sensing Image Classification Based On The Combination Of Deep Learning And Conditional Random Field

Remote sensing image classification is one of the fundamental techniques in remote sensing image processing system.Its accuracy will directly affect the application of subsequent remote sensing analysis and interpretation.Most of the traditional remote sensing image classifiers usually have shallow structures and always obtain a single-layer feature with no hierarchical structure.In contrast,Deep Learning(DL)can get a hierarchical feature representation automatically,which is more conducive to the image classification.In addition,the probabilistic graphical model is also a hot research topic in the field of image analysis.As one of the typical representatives,Conditional Random Field(CRF)model has achieved plentiful results in many image-processing applications.Since CRF model can make full use of the contextual information in images,it has a higher classification accuracy compared with the ordinary pixel-wise classification approaches.This dissertation mainly studies the classification of remote sensing images,which is based on the combination of deep learning and conditional random field.Our chief contributions are described as follows:(1)A remote sensing image classification method based on deep learning is designed.This method applied three different convolutional neural networks for training and prediction,including MyNet model,AlexNet model,and VGG model.At the same time,different scales of convolutional neural network are also fused for classification.Experiments show that the proposed method can obtain essential features and its classification results are more accurate than traditional methods.(2)A remote sensing image classification method based on the combination of deep learning and conditional random field is proposed.Firstly,we pre-classify the whole remote sensing image into certain land-cover types by CNN,using the results of class membership probabilities as the unary potential in the CRF model.Then the pairwise potential of CRF is defined by a linear combination of Gaussian kernels,which forms a fully-connected neighbor structure instead of common 4-neighbor or 8-neighbor structure.Eventually,a highly efficient approximate inference algorithm,namely,mean field inference,is generated for the CRF model.Experimental results demonstrate that this method supresses much classification noise due to considering the spatial contextual information in images,and it greatly improves the accuracy.(3)Regional Restriction(RR)are integrated into the whole CRF model framework.The approach uses Mean-Shift algorithm to obtain super-pixels and corrects the classification results by calculating their average of posterior probabilities in order to encourage the consistency of connected areas.The final model still uses mean field algorithm to achieve the inference.Experiments show that the edges of objects we gained are preserved more accurately and the classification result is more satisfactory.