Research On Marine Radar Rainfall Algorithm Based On Convolutional Neural Network

Frequent meteorological disasters in China have caused serious economic losses.Rainfall is one of the important factors.Accurate and quantitative estimation of rainfall is important in preventing floods and reducing secondary disasters caused by short-term sudden rainfall.With the development of meteorological technology,the radar can meet the rainfall observation needs well in the real-time and detection range of rainfall detection.Traditionally,weather radars used for rain measurement have high cost,complicated equipment,and high installation and site selection requirements.In this paper,from the perspective of cost performance,the choice of marine radar,combined with image processing to measure rainfall.The main research contents of this paper include:(1)The experimental platform for measuring the rainfall of marine radar is constructed,and the original rainfall data of the collected radar is processed,including removing the clutter of the ground object and the rain attenuation compensation,obtaining the echo image of the rain,and selecting the obvious and concentrated area of the echo on the image.And save,as the input of the neural network model training and testing.(2)Based on the classical LeNet-5 convolutional neural network model,this paper first designs a single-level convolutional neural network model,and then adds a fine-scale estimation part based on it,and designs a multi-level convolutional neural network model.More image information is obtained at different levels,and in order to avoid the phenomenon of gradient disappearance or gradient explosion,this paper optimizes the multi-level convolutional neural network,adds the residual network module,and designs a multi-level residual convolutional neural network model.The training set was created with the samples of light rain,moderate rain and heavy rain,and input into the single-level,multi-level and multi-level residual convolutional neural network.The three network models were trained,and the loss rate was calculated by the cross entropy loss function.It can quickly get the minimum loss,and use the batch training method to find out that the three network models have the ability to identify and classify the images,and they can achieve a good convergence and stable state.(3)The model that reached the steady state at the end of training was tested,and the sample test set of light rain,moderate rain and heavy rain images was re-established,and the recognition rates of the three models for light rain,moderate rain and heavy rain were respectively counted.Through experimental analysis,the multi-level residual convolutional neural network has the highest recognition rate and the best effect,and it is used in the practical application of marine radar measurement.In this paper,the rain echo image under unknown rain conditions is selected,and the multi-level residual convolutional neural network is used to judge it.The predicted results are compared with the data of Dalian Meteorological Network.The rainfall conditions of the two are consistent and meet the practical application demand.