A Remote Sensing Image Classification Based On Active Deep Learning

With the constant improvement of the remote sensing image resolution, there is an urgent request from remote sensing image to get more useful data and information, the classification of remote sensing image plays an increasingly important role in the social life and economic construction. However, for remote sensing images, it doesn’t exist a large number of tag samples in real life. Making the sample is not only time-consuming but very expensive. But with the constant improvement of the acquisition technology, it is becoming more and more easy to obtain a large number of unmarked samples. So in order to improve the classification performance of remote sensing image, we consider how to make full use of a few of marked samples and a large number of marked samples, active learning can effectively solve this problem.Active learning algorithm as a way to construct effective training set, it through iterative update in turn and looks for samples containing maximum information from unmarked samples, on the premise of limited time and resources, it can improve the efficiency of classification algorithm. At present, active learning has been a hot problem in the field of machine learning research.Deep learning is committed to modeling the multi-layer neural network and studying the learning research question, the concept of deep learning is the result of the artificial neural network research, it through the multi-layer neural network to analysis the data and solve the related optimal problem of the deep structure.In this paper, we use a kind of deep learning algorithms-sparse auto encoder(SAE), it is looking for high-dimensional data feature at the same time by sparse regularization term to make features to have sparsity, through this way, it can not only ensure the extracted features can eliminate redundant, but also has good representation capability. This paper presents a remote sensing image classification method based on deep learning and n EQB.Firstly, deep learning uses training samples to obtain the initial classification, and then active learning is used to choose the most beneficial samples from unmarked samples to be marked by experts, the marked samples will be rejoin in training samples, in this wayto iterative update classifier.In this paper, in order to improve the classification accuracy of SAE, active learning is joined in the framework of SAE. Active learning by increasing the number of training samples to improve the classifier performance. In order to verify the effect of the combination of these two methods, we compare the experiment with SAE with RS and SVM with active learning method respectively, Furthermore, sensitivities to the parameters with the two algorithms are also tested in the comparison experiments.Experimental results show that: in the classification of remote sensing image, with same relative few training samples, SVM is usually slightly better than SAE; active learning schemes can take into good effect in achieving higher classification accuracy for the both methods.