Research On Segmentation And Recognition Technology Based On Ground Cloud Image And System Construction

Cloud phenomenon is an important role in the field of meteorological observation and prediction.The shape and quantity of cloud reflect different weather characteristics.Accurate observation of cloud amount and cloud shape change is of great significance for weather prediction,flight support and climate research.Ground-based cloud observation is an effective method to obtain cloud amount and cloud shape information in the field of meteorological observation.On the one hand,ground-based cloud observation by manual directly observed ground-based cloud,on the other hand,ground-based cloud image is obtained by ground-based observation equipment,and then refined judgment is made by professional observers.However,both of them have problems of heavy workload and lack of objectivity of observation results.In recent years,many automatic ground-based cloud image observation techniques based on deep learning have been studied at home and abroad,including segmentation technology for cloud amount detection and classification technology for cloud shape recognition,which have greatly promoted the development of meteorological observation technology.However,while the segmentation technology has high accuracy,it has the problems of complex model and low segmentation speed;When obtaining cloud amount and cloud shape,cloud image segmentation and classification technologies are mostly realized in the form of single task,which is difficult to achieve the optimal effect and efficiency.Based on this,taking the ground-based cloud image provided by the Meteorological Bureau as the research object,this thesis studies the segmentation and recognition technology of ground-based cloud image,so as to realize the fast and accurate segmentation and recognition of cloud image.The main research contents and results of this thesis are as follows:1.High-quality,complete category and large-scale with labels ground-based cloud image datasets GBCD and GBCD-GT are constructed and improved.Aiming at the problems of small amount,low image quality,incomplete cloud image categories and no classification labels of cloud image segmentation data,this thesis corrects the ground-based cloud image collected by TSI,and manually labels them under the guidance of professionals to form the original dataset containing 11 types of groundbased cloud image.Then,the classification network is used to assist the secondary classification of cloud images,and finally the initial ground-based cloud image dataset is formed.In this thesis,after training UNet with open source segmentation dataset,UNet is used to segment the initial ground-based cloud image according to the label.Then,YCrCb channel separation,threshold segmentation and image matrix phase are used to remove the sunlight from the cloud image.Finally,after manual calibration,the initial GT ground-based cloud image dataset is formed.The initial ground-based cloud image dataset and initial GT ground-based cloud image dataset are enhanced to form open source GBCD and GBCD-GT datasets,which provide data guarantee and support for subsequent experiments in this thesis and cross-validation of experiments in related fields.2.A ground-based cloud image segmentation model BFSegNet based on singletask learning is proposed.Aiming at the problems that the existing cloud image segmentation methods with high accuracy have complex model and low segmentation speed,this thesis constructs the ground-based cloud image segmentation model BFSegNet by using the bilateral segmentation backbone architecture and the design idea of lightweight convolution module.The model extracts detail and semantic features from detail and semantic branches,and the two features are fused by the feature fusion module.Finally,the segmentation mask is formed by up-sampling.The experimental results show that the model not only has the advantages of small parameters and computation,but also can realize fast and accurate ground cloud image segmentation,which is convenient for the subsequent practical deployment and application of the model.3.The multi-task model BFRSegNet of simultaneously implement segmentation and identification for ground-based cloud image is proposed.Considering that most of the current segmentation and recognition methods of ground-based cloud image are separately implemented in the way of single task,which is difficult to achieve the optimal effect and has low efficiency.In this thesis,the BFRSegNet model is constructed by using the idea of multi-task learning,and the lightweight recognition network is added on the basis of the single task model BFSegNet to accomplish two tasks in one network.Experimental results show that the segmentation accuracy is95.68%,the average intersection ratio is 74.33%,and the recognition accuracy of the model is 95.07%.The segmentation and recognition effects are better than other single task models.Therefore,BFRSegNet can efficiently and accurately realize the two tasks of segmentation and recognition on the premise of adding less parameters and computation,which provides a feasible scheme for the automatic observation of ground-based cloud image.4.Based on the above algorithm research,the intelligent cloud recognition application system of ground-based cloud image is designed and implemented.This thesis uses Springboot + HTML + Javascript + CSS to realize the visual display of model results,historical data query and other functional modules.On the visual display module of model results,after the system obtains the ground-based cloud image,users can intuitively obtain the cloud amount and cloud shape information from the visual presentation of model segmentation and recognition results.The visual results are saved in the system and displayed in the historical data query module for subsequent inspection and verification.The system promotes the multi task model of ground-based cloud image segmentation and recognition for practical applications,and improves the intelligence and informatization of ground-based cloud image observation.