Cloud Classification Of Satellite Imagery Based On Stacked ELM

Meteorological satellite can observe a full range of the earth, so as to obtain the satellite cloud image which contains rich meteorological information, satellite cloud image can be used to identify systems of different weather, for weather analysis and forecasting and disaster monitoring provided the powerful basis. However, exploration and analysis of satellite images, in most cases still use the artificial visual interpretation method, which not only mixes up people’s subjective consciousness, but also hinders the fully extract and maximize the use of information-rich of satellite images, so achieving automatic classification of satellite images of the computer is a big mainstream of the satellite image information processing. At the same time as the training data of satellite cloud image are large and complex, training will take up a lot of memory, if we use traditional ELM algorithms to complete the training of the data, it is easy to lead the entire program to collapse because of the insufficient memory.Based on this, this paper researches the stacked extreme learning machine(S-ELMs) algorithm and innovatively use of it to the automatic classification of satellite cloud image. S-ELMs is approximately a big ELM that is divided into several small series of ELM, under the condition of the small memory requirements and fixed network size, can also study data in high-dimensional space of ELM, well solve the problem of the large and complex data in the current cloud image processing. This article also innovatively introduce the texture feature of satellite cloud image, respectively with pure spectral characteristics forms a kind of training text, with the combination of texture features and spectral characteristics constitute a kind of training text. Based on these two kinds of training text, using S-ELMs algorithm to achieve the satellite cloud image classification, verified S-ELMs algorithm is validity in the satellite cloud image classification, and respectively contrasting the experiment of S-ELMs and ELM, S-ELMs and SVM, analyze the classification advantages of the same training text and different algorithms, the different training text and the same algorithm, to better study the advantages and disadvantages of stacked extreme learning machine in the satellite cloud image classification, and analyze the influence that the introduction of texture feature has on classification effect.The results show: S-ELMs can be effectively used in the satellite cloud image classification, compared with ELM and SVM, S-ELMs have obvious advantages on memory requirements, can solve the problem that training data is too large and complex, and improves the generalization ability; S-ELMs has obvious advantages in learning speed than SVM, but S-ELMs classification accuracy is lower than SVM; S-ELMs and ELM have no difference in learning speed when their classification accuracy is equal, on the basis of the sacrifice of time cost, S-ELMs classification effect have a little improved than ELM, at the same time the introduction of texture feature can improve the sensitivity of the algorithm for the training sample feature vector, distinguish ground objects that using pure spectral characteristics to discern can’t discern, so as to improve the overall classification accuracy of the experiment.