Research On Satellite Cloud Image Segmentation And Recognition Based On Deep Learning

A satellite cloud image is an image of cloud cover and ground surface features on the earth observed by a meteorological satellite from top to bottom.Satellite cloud images can identify different weather systems,determine their location,estimate their strength and development trends,and provide a basis for weather analysis.At the same time,the use of satellite cloud images to accurately identify the types and distribution of clouds is of great significance and value to improving the quality of weather forecasts and the ability to prevent disasters.This paper uses convolutional neural networks and deep convolution generation adversarial network as the basis for an in-depth study of satellite cloud image segmentation and recognition.The main work and innovations are:(1)In the traditional satellite cloud image segmentation method,cumbersome algorithm flow and complex feature extraction method,this paper uses a method based on convolutional neural network,which can automatically extract features,the process is concise and more efficient.At the same time,the number of parameters in the FCN segmentation model is too large after multiple convolutions,which leads to time-consuming and laborious model training.In this paper,the ordinary convolution is improved to a "sparse convolution" method,which greatly reduces the network parameters and calculations,and the training speed has been significantly improved,the average training efficiency has been increased by 26.3%,and the segmentation effect is almost not affected.(2)For training convolutional neural network models,a large amount of data is often required,and some types of satellite cloud images have a small amount of data,and it is difficult to collect a sufficient amount of data.This paper generates training data by deep convolution generation adversarial network to expand training set effectively improves the recognition accuracy of the recognition model.At the same time,because the convolution kernel in the convolutional neural network is too large,resulting in high computational complexity,a convolution reconstruction method is proposed.Multiple small convolutions are superimposed instead of large convolutions,and the same feeling is ensured when the calculation amount is reduced.The training time is reduced;For a single Sigmoid activation function to easily cause gradient dispersion,a variety of activation function fusion methods are proposed to improve the training accuracy.At the same time,the accuracy of model recognition combined with multiple methods reached 88.5%,an increase of 13.2% compared with the original model.