Single-station Doppler weather radar heavy rainfall estimation method based on double-branch double-stage depth model

Abstract

The invention discloses a single-station Doppler weather radar heavy rainfall estimation method based on a double-branch double-stage depth model. The method comprises the steps of building a double-branch double-stage depth model BBDM framework which is provided with a double-branch network structure; preprocessing the single-station Doppler radar feature data; adopting a double-branch training strategy in the BBDM training process; adopting a double-stage training strategy in the BBDM training process; storing a model obtained by training for subsequent testing; and in a test stage, preprocessing radar feature data, and then inputting the preprocessed radar feature data into a previously stored model, only taking a regression branch as final output, and obtaining a final precipitation estimation result. By means of the model, precipitation of various different intensities can be rapidly and accurately estimated by utilizing single-station Doppler weather radar observation data, and meanwhile, the method is a universal method used for processing the regression problem with obvious sample imbalance.

Description

technical field [0001] The invention relates to the technical fields of image recognition and weather intelligent recognition. Specifically, it is a single-station Doppler weather radar heavy precipitation estimation method based on a dual-branch dual-stage depth model. Background technique [0002] The traditional Quantitative Estimation of Precipitation (QPE) technology using Doppler weather radar is to use the empirical relationship between the radar echo intensity (Z) and the surface precipitation (R) (usually known as the Z-R relationship), and the Z-R relationship can be described as: [0003] Z=aR b [0004] Among them, a and b are parameters, and the size of the parameters is closely related to the precipitation weather type, season, and region, and is usually determined by empirical relationship. Uncertainty in parameter setting is the main reason for the low accuracy of surface precipitation estimation using traditional techniques. [0005] In recent years, the...

AI Technical Summary

Problems solved by technology

Using the traditional deep learning framework often leads to good learning effect on small precipitation with high occurrence probability, but poor learning effect on stronger precipitation with more practical value

Images