Multi-climate Satellite Cloud Image Detection Based On Dilated Multi-grained Cascade Forest

The Cloud image detection classification is the premise of meteorological research and application.The accuracy of accurate detection of meteorological cloud images directly affects subsequent atmospheric science research and meteorological prediction applications.When the traditional shallow learning classification algorithm is used for satellite cloud image detection classification,the satellite cloud image can not be well characterized.There is human error prior knowledge deviation,the model can not implement effective cloud image detection,and the cloud category false detection rate is very high.While the deep neural network algorithm has achieved high accuracy in cloud image detection,the convolutional neural network is easy to over-fitting when the number of data sets is small,and the convolutional neural network training takes a long time,with the number of network layers.The increase in the amount of calculations also rises rapidly,making it slow to detect and failing to meet meteorological research and application requirements in terms of classification speed.This paper selects two representative methods from the existing cloud image detection methods from the two aspects of model detection speed and accuracy,and compares with our proposed multi-granular cascade forest method.The depth limit learning machine method can not effectively extract and analyze the parameters because it does not need backpropagation in the training process,but the texture detail information cannot be effectively extracted in the cloud map.The multi-scale convolutional neural network can be effective in cloud image detection.The feature texture and spatial information are extracted,and the overall accuracy of the model is higher than other existing methods,but the training takes a long time.Compared with these methods,the hollow multi-granular cascade forest method can well balance the contradiction between model speed and accuracy.The comparison of the satellite cloud image detection experiments shows that the satellite cloud image is detected by the cavity-based multi-granular cascade forest method.Each cavity feature window obtains a large enough receptive field while retaining the internal spatial information of the image.The cloud image is characterized by good generalization performance.The hollow feature window of the model increases the diversity of feature vectors,but reduces the computational complexity and memory consumption of the model.The model training and detection speed is cascaded compared with the original multi-granularity.The forest model has more than doubled,and the memory required by the model has been reduced by about one-half,and the accuracy of the model has improved compared with the original model,which can be distinguished not only in the high-altitude cloud and snow-covered conditions.The difference in texture between the clouds and the snow;and the multi-granular cascade forests in the plains and lakes of the plains can also distinguish the vegetation from the lake.The work of this paper has some help for some subsequent meteorological research and application.