Cloud Fraction Computing Of Satellite Images Based On Deep Extreme Learning Machine

Cloud classification, cloud detection and cloud fraction computing based on satellite images are the main ways to obtain regional cloud distribution,and also the basis of satellite meteorological applications, however, the current research fails to make foll use of the characteristics of satellite cloud images and satellite optical parameters,causing the result of cloud classification and cloud fraction computing is not good, from the actual meteorological applications, the use of cloud fraction computing technology is not ideal. After reviewing the many domestic and foreign research results, this paper applies the deep extreme learning machine to compute the cloud fraction of satellite images.In recent years, the research about neural network is very active, in which the extreme learning machine has shown good applicability and robustness in a large number of practical applications, with good self-learning ability and fast learning speed, simultaneously, the deep learning based on multi-layer neural network has a stronger learning ability,therefore,this paper uses a method based on the deep extreme learning machine to detect and classify the clouds in the satellite cloud images,and then on this basis to solve the cloud fraction computing problem of satellite cloud images. In order to reflect the advantage of the deep extreme learning machine, another detecting method of cloud images based on convolution neural network is proposed,and the deep extreme learning machine is compared with the convolution neural network. The main work of this paper includes the following parts:Firstly, we extract a large number of fixed pixel-sized satellite cloud images blocks as the training sample of the deep extreme learning machine, and then using the deep extreme learning machine to detect the thick cloud, the thin cloud, the clear sky and the overlap of thin cloud and thick cloud in the different channels of the satellite cloud images. Then the detection result is compared with the detecting cloud images of the traditional threshold method and the extreme learning machine and the convolutional neural network.Secondly, after the cloud detecting, then using the improved "spatial correlation method" to calculate the total cloud fraction, and finally, the total cloud fraction computing results of satellite cloud images are analyzed and compared with the expert database. After continuous improvement of detection algorithm and cloud experiment, this paper will make some contribution to the observation and interpretation of satellite cloud images.