Research On Satellite Cloud Image Recognition Algorithm Based On Multi-granularity Information Fusion

Remote sensing satellite cloud image plays an extremely important role in meteorological research and application,and has important research value in natural disaster prediction and meteorological detection technology.Cloud image classification is the core of remote sensing satellite interpretation.The accuracy and speed of satellite cloud image detection directly affect the atmospheric science research and weather environment prediction.The traditional shallow machine classification learning algorithm is difficult to extract the cloud image features from remote sensing satellites effectively,and there is a bias of artificial experience judgment.As a result,the traditional machine learning algorithm can not classify the cloud image well,and the error detection rate is high and the training time is time-consuming.The traditional deep learning network algorithm can automatically realize the analysis of cloud image features.Although it has achieved a high accuracy in the satellite cloud image classification,the deep learning network is prone to over-fitting when the training data set is small.Generally speaking,the more convolutional layers the deep neural network has,the greater the number of parameters.However,it will undoubtedly prolong the classification time,resulting in too slow detection speed,which cannot meet the requirements of atmospheric science research and weather and meteorological environment prediction.In view of the above problems,this paper proposes an extended multi-scale cascaded forest network and a packet pyramid convolutional residual network,considering the accuracy and speed of the algorithm model for satellite cloud image recognition.The expanded multi-scale cascaded forest algorithm injected the hollow idea of expanded convolution into the multi-scale sliding window,which greatly reduced the large number of parameters brought by the multi-scale mechanism and expanded the reception field of feature extraction information,thus improving the efficiency of the algorithm in the application of satellite cloud image classification.Grouping pyramid residual pyramid residual unit and put forward by the network grouping unit instead of the original residual unit residual,trying to apply pyramid convolution layer to residual branch,the pyramid for standard convolution convolution,it contains different scale and depth of convolution kernels,ensure the multi-scale feature extraction,compared with the original residual unit for the characteristics of information more abundant,and pyramid convolution can be efficiently calculated in a parallel way.