Multispectral Remote Sensing Image Classification Based On Improved RBM

Remote sensing images provide a wealth of observation data for surface observation. In order to use remote sensing data effectively, effective interpretation and analysis of remote sensing images is needed, transforming spectral information of remote sensing image to the user's category information. The classification of remote sensing image is very important in the analysis and application of remote sensing data and a hotspot in the field of remote sensing. However, the phenomenon that the same substance with different spectrums and different substance with the same spectrums in multispectral remote sensing images increases the difficulty of classification. In addition, access to the ground truth data or experts to label training data is time consuming, and the price is high. Although there are many classification methods of multispectral remote sensing image, there is still no very effective algorithm. Therefore, it is very important to explore more efficient and accurate method of remote sensing image classification, which has very important practical significance and application value.Deep learning is an emerging field of machine learning in recent years. The research content of this field is to study the modeling and learning of artificial neural networks with multiple nodes. This "deep neural network" is more similar to the human brain in information processing, so it is believed that it may solve some complicated problems. The machine learning problems of Restricted Boltzmann machine(RBM) in the field of deep learning is the key problem. RBM has a strong ability of unsupervised learning, can learn complex rules in data, and has been used as a structural unit of deep neural network. It has been widely concerned in recent years and applied successfully in many fields.In this paper, we mainly study how to improve the RBM to achieve higher classification accuracy for the multispectral remote sensing image. Our work and contributions of this paper are as follows:1.A semi-supervised classification algorithm based on error corrected K-RBM and improved RBM-softmax model is proposed. Firstly, the algorithm uses the error corrected K-RBM to cluster the unlabeled samples, separating them into different categories, and obtain the classification data with high level confidence according to the reconstruction error; Based on the a few of labeled samples, the class labels of the improved data with high level confidence are obtained, and the training sample are formed with the data and the very small number of labeled samples. Then we use the training samples to train the improved RBM-softmax classifier model, obtaining the parameters of the model. Finally, the test data is classified by the well-trained RBM-softmax classifier. The error corrected K-RBM model overcomes the shortcomings of the K-RBM model, which can not guarantee the smallest interval within-class and the biggest interval of the inter-class. The error corrected K-RBM model not only uses the feature of the pixel itself, but also the labels of the neighour pixels to determines the class label of the pixel, improving the classification accuracy. The improved RBM-softmax model has the characteristics of primary visual cortex cells, that is sparse and selective. Experiments are carried out on two groups of multispectral remote sensing images, which proves the effectiveness of the improvement.The results show that the semi-supervised classification method proposed in this paper has the advantages of high classification accuracy and fast speed.2. An improved RBM model based on cross entropy regularization term is proposed, and the method to update parameters is derived, in order to obtain more separable multispectral remote sensing image features. Firstly, the feature of multispectral remote sensing image is extracted by the improved RBM, and input the feature to the softmax classifier for classification. Experiments on two sets of multispectral remote sensing images verify the effectiveness of the method.