Precipitaion Nowasting Based On Deep Learning

Precipitation nowcasting refers to high-resolution predictions of precipitation and hydro-meteorological phenomena within the next zero to two hours.Precipitation nowcasting forecast plays a crucial role in mitigating the impact of natural disasters,improving agricultural productivity,and managing transportation and urban areas.Traditional numerical weather prediction methods often fail to achieve satisfactory forecast results due to limitations in model computation time complexity and the ability to extract nonlinear information.With breakthroughs in computer computing power and the development of neural network technology,deep learning networks have been applied in various fields.This thesis studies the deep learning model forecasting of precipitation nowcasting based on the perspective of computer vision,focusing on Doppler radar images to precipitation images.The task is abstracted as cross-modal spatiotemporal prediction task,integrating traditional algorithms,and fully utilizing deep learning methods to extract spatiotemporal meteorological characteristics within the precipitation area,and proposes solutions to over-smoothing issues in generated images.The main results of this work are as follows:1.Based on the existing research results of deep learning spatiotemporal prediction models,a feedback mechanism is introduced,and an adaptive loss function is proposed based on the long-tail distribution problem of meteorological precipitation label data,effectively improving the forecast indicators for high precipitation.Through experiments,the FED-Net proposed in this thesis is superior to traditional methods and deep learning benchmark methods.2.We analyze the characteristics of radar images.A data augmentation module for radar features is designed.Through wavelet transformation,noise in radar feature images is suppressed,and the different performances of temporal and spatial wavelets under meteorological data are analyzed and compared.Optical flow is used to extract inter-frame information of radar image sequences.Through the data augmentation module,the spatiotemporal features of radar images are highlighted,which helps the model to learn the spatiotemporal information in radar images more effectively and improve the overall forecast indicators.3.To address the over-smoothing problem of the output forecast results of traditional spatiotemporal models,a generative adversarial network architecture is used to fit the precipitation label distribution and visually enhance the output forecast results.