Research On Cloud Type And Cloud Cover Recognition Technology Based On Lightweight Convolution Neural Network

Cloud is an important part of the earth's thermal balance and water vapor cycle.It is of great significance for weather forecast,flight support and climate research to accurately identify cloud shape and cloud amount.Ground cloud observation is an important part of ground meteorological observation.For a long time,ground cloud observation mainly relies on the manual visual observation and empirical analysis of meteorological observers,which is lack of objectivity and poor comparability.On the other hand,the cost of manual observation is too high to meet the actual needs,so it is urgent to realize the automatic observation of ground-based cloud image.In recent years,with the development of ground observation equipment and digital imaging technology,some ground-based observation equipment has been developed at home and abroad to record and save the ground-based cloud image,but the fine interpretation of the image still needs to rely on the meteorological observer.Therefore,how to use the technology of deep learning and computer vision to realize the automatic observation of the ground-based cloud has become the research focus in the field of meteorological observation.In this paper,based on the ground-based cloud image provided by the standard meteorological station,the automatic observation technology of the ground-based cloud is studied from two aspects of cloud type recognition and cloud cover detection.The main research contents and research results are as follows:1.The standard ground-based cloud image data set HBMCD and standard GT ground-based cloud image data set HBM?GT with the uniform quality,large data volume and complete cloud genus are constructed.Standardized cloud image data set is an important foundation for automatic recognition of ground-based cloud.In view of the existing open source cloud image data sets with different data quality,incomplete cloud species and low data volume.Based on the TSI cloud image data provided by the standard meteorological station,after image correction and under the guidance of professional meteorological observers,this paper classifies and forms the basic data set including ten kinds of clouds and one kind of cloudless cloud image,totaling eleven kinds of cloud images,and uses migration learning for auxiliary verification.After data enhancement and other operations,HBMCD is formed.The open source data set SWIMSEG and SWINSEG after data enhancement are used as the initialization data set of UNet.After the initial training of the network model,the HBMCD is segmented by the network,and the network output is combined with the YCC channel separation,threshold segmentation and other operations to remove the interference of the sun's strong light.After manual calibration,the HBM?GT is formed.2.A lightweight cloud recognition network model LCCNet is proposed.In order to solve the problems of limited identification of cloud,high memory consumption and low accuracy of existing cloud recognition methods,this paper first explores the performance of each classical network model in cloud recognition by means of transfer learning method,and then the LCCNet is designed and implemented by using the network model structure with optimal recognition effect and the techniques of reducing model parameters such as depthwise convolution,pointwise convolution,channel shuffle,dilated convolutions,etc.Experiments show that LCCNet can extract the features of each cloud genus better.At the same time,on the premise of ensuring high recognition accuracy,LCCNet greatly reduces the amount of network parameters and computational complexity,which provides the possibility for actual deployment.3.A lightweight cloud segmentation network LCSegNet is proposed.In view of the poor segmentation effect of the existing semantic segmentation network for cloud images and the inability to solve the problems of solar interference and high memory occupation,this paper uses the "Encoder-Decoder" network model framework to design the structure of LCSegNet,and combines the idea of channel splicing to scale the characteristics of encoder part and decoder part with the same number of channels to realize the aggregation of different levels of image features,thus,the boundary loss of the feature is avoided.Finally,the semantic segmentation mask with the same resolution as the original image is output.Experiments show that LCSegNet can not only effectively remove the interference of strong sunlight on cloud image segmentation,so as to achieve accurate segmentation of cloud image,but also has the advantages of low parameters,small calculation and low memory consumption.4.Design and build a cloud element recognition platform.In order to push the network model to practical application,this paper selects Py Qt5 framework and Python language to build the platform,and uses CSS to optimize the page.After obtaining the cloud image data,the platform uses the trained cloud recognition network model LCCNet and cloud segmentation network model LCSegNet to analyze the cloud type and cloud amount.By visiting the platform,users can intuitively obtain the image and recognition results of the visualization part of the cloud image in the analysis process.This paper mainly introduces the design and implementation process of the application platform,including functional modules such as cloud type recognition,cloud segmentation result display,cloud cover detection,and gives the use results and performance analysis.