Research And Application Of Weather Phenomenon Recognition Algorithm Based On Deep Learning

Weather conditions are directly related to people's travel planning,clothing,crop growth,disaster prevention and other aspects of production and life.The research on weather phenomenon has a long history.As an important field,its exploration has never stopped.The traditional weather recognition technology based on artificial experience was far from meeting people's demand for meteorological services due to its time-consuming,subjective and narrow coverage.With the development of social modernization,the new generation of meteorological observation instruments and equipment,such as ground meteorological stations,satellites,radars and so on,have made great progress,provided necessary hands for accurate testing and collection of meteorological data.However,how to deal with these weather data efficiently,analyze and use them scientifically,so as to achieve the purpose of accurate weather classification and recognition,is still a very challenging task in the field of meteorology.With the rapid development of computer network and computer technology,especially the rise of big data and artificial intelligence technology in recent years,using computer vision and deep learning methods to study weather phenomena has brought hope for the automatic and accurate classification and recognition of weather phenomena.It can not only reduce more labor costs and realize real-time,comprehensive,continuous and quantitative weather observation,but also provide effective help for weather prediction,agricultural disaster warning,intelligent transportation and other development.Therefore,based on the computer vision and deep learning network models,this paper studies the algorithm of weather phenomenon automatic recognition,and constructs multiple weather dataset to verify the recognition effect.At the same time,a weather phenomenon detection platform was designed and developed to verify the theoretical results and make it more practical.The main contents of this thesis are as follows:(1)A multi weather phenomenon dataset CCW was constructed,which has many kinds of weather,covers a wide area,has a large amount of data and is oriented to any scene.It includes six kinds of weather phenomena: haze,sandstorm,rain,snow,frost and dew.19753 weather phenomenon image samples were collected by means of meteorological observation station equipment,daily shooting and web crawler,including real life scene,images of different time,region and background.Due to the different frequency of different weather,there was a problem of data imbalance between different types of samples in the dataset.In this paper,different weather features were extracted,and data enhancement methods were used to expand the small sample size categories.Finally,a multi category weather dataset with large sample size and data balance was built.This dataset effectively improved the problems of small sample size,fixed scene and single category of weather phenomenon data set in the field of image classification.(2)A lightweight weather phenomenon recognition algorithm based on depth network model was designed and developed.In this paper,firstly,Dense Net model was selected to verify the effect of deep learning method for weather recognition by using transfer learning method.And then the weather characteristics were learned by adjusting model parameters and back propagation.The recognition results show that this method has high accuracy and good feasibility.Furthermore,a lightweight network model was built.The jump connection structure was used to solve the problem of gradient dispersion and model degradation of convolution neural network.The Squeeze-and-Excitation block was used to enhance the learning inhibition of the effective features.The deep separable convolution was used to extract the weather features instead of the traditional convolution,so as to ensure the learning ability and reduce the network parameters and computational complexity.The recognition accuracy of the algorithm was 96.29%,and the memory occupied by the model was32.97 MB.It not only ensured the recognition effect,but also reduced the memory occupation,which has provided a solution for the intelligent deployment of meteorology.(3)A multi-functional weather phenomenon detection system was designed and implemented.The system was based on B/S architecture and SSM(Spring + Spring MVC + Mybatis)framework.It implemented a multi module system platform,including news headlines,weather phenomenon learning,weather identification,urban and weather statistics.Then,the performance analysis and pressure test of the platform were carried out.