Scene Classification Based On The Neural Network

Technologies in remote sensing domain have been applied in many domains along with the fast development of them. Scene classification takes a great part of them. As the number of the remote sensing sources grows, more and more analytic methods and technologies have been developed and large numbers of new methods and ideas of classification have come forth. The choice of the appropriate classifier and the selection of features with good capability of partition are crucial to decide the performance.This paper first introduces about several traditional unsupervised and supervised algorithms based on statistics briefly. As the traditional methods can not meet the need nowadays, we select the Neural Network algorithm instead, because it is artificial and automatic. As the grays of the same scene always change along with the change of the season, weather and time, we choose the features which are interrelated with the structure and distribution. In this paper, we select the texture features which show the situation of the spatial change of the grays in images. 21 features have been introduced from the 193 candidate features which are composed of several co-occurrence features, gray level-gradient co-occurrence features, gray level-smoothed co-occurrence features and run-length features based on the principle that we should choose the feature which has big square deviation between-class and small square deviation within-class. This paper also expatiates about some problems which will appear during the classification of an image and propose some solutions based on the guarantee of high preciseness of classification. We propose a superposition search method and use the data fusion in level of decision-making to improve the veracity and accuracy during the traversal of a whole image. The class of scattered residential area is an important and special one during the scene classes. Because the accuracy we need of the class of scattered residential area is higher than the needs of other classes, and it has some complicated modes, a method based on the total lightness and superposition search has been proposed, and we reuse the BP Neural Network to distinguish the modes of the class of scattered residential area during the process of classification of a whole image. In the end, the paper summarizes the whole study, recapitulates the research result and affirms that the algorithm is quite efficient.