Research On Precipitation Phenomenon Recognition Technology Based On Raindrop Spectrum And Machine Learning

Observation of precipitation phenomenon is one of the important contents of meteorological observation.Realizing automatic observation of precipitation can provide better meteorological services,and improving the accuracy of automatic precipitation observation is one of the important contents of its research.This research is based on the Agricultural Meteorological Automated Observation Station project of the Aerospace New Meteorological Technology Co.,Ltd.of the 23 Institute of Aerospace Science and Industry.It uses the precipitation data of nearly 100 stations in Guizhou Province from 2017 to 2018 to analyze and find that the current precipitation phenomenon instrument still exists Errors in some aspects are mainly reflected in: when there is sleet,it is easy to be falsely reported as rain and snow,the recognition accuracy of drizzle and hail is low,and when faced with certain unknown phenomena,false alarms will occur.In response to the above problems,this paper designs and proposes a method based on raindrop spectrograms to automatically identify precipitation phenomena under natural weather conditions.The main research is as follows:1.According to the precipitation phase distribution,the minute raindrop spectrum is designed,the original precipitation data is converted into minute raindrop spectrum image data,and the improvement is made on the basis of the raindrop spectrum,and a 32×32 raindrop spectrum design method is proposed.Fitting the improved GK curve makes the improved raindrop spectrogram more prominent description of the characteristics of different precipitation.2.Using a 32×32 raindrop spectrogram,two methods are proposed to identify precipitation phenomena.One is to use the traditional feature extraction algorithm HOG and SURF to combine two classifiers SVM and random forest to classify precipitation types.HOG descriptor,cross-designed 8 kinds of experimental models to comprehensively compare and analyze.The second is to improve the network structure of VGGNet-16.The 11-layer convolutional neural network RCNet is adopted.While reducing model parameters and accelerating the speed of network convergence,it effectively avoids the over-fitting situation during VGGNet-16 training,making the model It is more stable and has better robustness.3.Through the 32×32 raindrop spectrogram design method,16,500 raindrop spectra were obtained.The obtained raindrop spectra were identified and classified using the above-mentioned method.Three evaluation indicators of total accuracy,class precision,and class recall were used to evaluate the model.Experiments show that the convolutional neural network RCNet designed in this paper is in the precipitation type.The total accuracy rate of recognition reached 91.05%,the highest accuracy rate and recall rate reached 97.69% and 98.09%,respectively.Compared with the traditional precipitation phenomenon instrument,the total accuracy rate was increased by 11.87%,realizing the recognition and classification of precipitation phenomena.,Has a certain engineering application value.