Page 1 of 7

European Journal of Applied Sciences – Vol. 12, No. 1

Publication Date: February 25, 2024

DOI:10.14738/aivp.121.16248

Hua, G., Jie, L., Dongling, B., Wang, Z., Qi, Y., & Zuofang, Z. (2024). Revised Test of Precipitation Forecasting Product of RMAPS-ST

System Based on U-Net. European Journal of Applied Sciences, Vol - 12(1). 561-567.

Services for Science and Education – United Kingdom

Revised Test of Precipitation Forecasting Product of RMAPS-ST

System Based on U-Net

Gao Hua

Institute of Urban Meteorology (IUM),

China Meteorological Administration (CMA), Beijing

Liang Jie

Corresponding author: jliang3@gzu.edu.cn

School of Mathematics and Statistics, Guizhou University

Bao Dongling

(bdl0124@sina.cn)

Institute of Urban Meteorology (IUM),

China Meteorological Administration (CMA), Beijing

Zaiwen Wang

Institute of Urban Meteorology (IUM),

China Meteorological Administration (CMA), Beijing

Yajie Qi

Institute of Urban Meteorology (IUM),

China Meteorological Administration (CMA), Beijing

Zheng Zuofang

Institute of Urban Meteorology (IUM),

China Meteorological Administration (CMA), Beijing

ABSTRACT

Using the U-Net algorithm idea in the field of artificial intelligence, a deep learning

framework is designed and the loss function is redefined. This learning framework

revised the precipitation forecast results of RMAPS-ST after training on two

summer samples. Approved the revised precipitation forecast for June-July 2022

and precipitation on August 18, 2022, the test results show that the revised forecast

level is significantly improved, especially for magnitude greater

than 0.1mm/3 hours, and the CSI score is increased by an average

of 30% compared with before revision.

Keywords: Gridded data, Deep learning, Precipitation correction, U-Net.

INTRODUCTION

The origins of deep learning can be traced back to the 40s of the 20th centuries, its prototype

appeared in cybernetics. In the past decade, deep learning has re-entered people's vision, and

Page 2 of 7

Services for Science and Education – United Kingdom 562

European Journal of Applied Sciences (EJAS) Vol. 12, Issue 1, February-2024

the research on deep learning problems and algorithms has also experienced a new wave. While

the design of convolutional networks is inspired by biology and neuroscience, the current

development of deep learning has long gone beyond the neuroscientific perspective in machine

learning models. It uses relatively simple functions to express complex representations, from

low-level features to more abstract high-level features, allowing computers to mine implicit

information and valence from experience. The idea of combining deep learning with numerical

models is increasingly recognized in the field of meteorology, and a large number of valuable

results have emerged. The main research and development directions of numerical models

include initial field improvement based on assimilation, improvement of model parametric

schemes, and output statistics of model prediction results (MOS)[1-6] et al. Artificial

intelligence starts from the perspective of numerical model post-processing, which has the

advantages of obvious effect and high efficiency, and is the mainstream direction of artificial

intelligence algorithm application in the field of meteorology[7-10]。The U-Net method is a

deep learning method based on convolutional neural network, which has the advantages of high

training efficiency and low sample demand, and is widely used in neighborhoods such as

medical imaging and face recognition. In the field of meteorology, research on the revision of

forecasts based on the U-Net method has focused on short-term forecasting [11-14].

Internationally, Hess (2022) adopts the two-layer U-Net method to use the ECMWF based on

TRMM data After training on 3-hour cumulative precipitation forecast and vertical velocity data

of 11 layers of barometric layer, the prediction effect test shows that U-Net is more likely to

precipitate than the traditional numerical model ECWMF-IFS. There are obvious

improvements, but only precipitation probability results, cannot meet the actual demand. In

China, Guo et al., 2019 et al. compared the linear extrapolation prediction of recurrent neural

network with strong convective system within 1 hour, and the test results show that the

prediction effect of recurrent neural network is significantly improved compared with linear

extrapolation. Zhang Debiao et al. revised the ground elements of ECMWF forecasts using the

improved U-Net method (Zhang et al., 2022). The test results show that the revised ground

temperature, humidity, wind direction and wind speed have been significantly improved, but

there is no precipitation correction result. Based on the above results, the U-Net method was

used to revise the 3-hour cumulative precipitation forecast of the RMAPS-ST system within the

24-hour time limit, in order to improve the Ritu system (RMAPS-ST) precipitation forecast

level.

MATERIALS AND METHODS

U-Net was designed by Olaf Ronneberger et al. and published in 2015, it is a variant of FCN. U- net's original intention was to solve the problem of biomedical images, and because it worked

really well, it was later widely used in all directions of semantic segmentation. In this study,

according to the size of the region and the number of grid points, the two-layer U-Net design

was adopted, and multiple convolutional layers (two convolution operations per layer), down

sampling layer (reduced resolution), up sampling layer (increased resolution) and fully linked

layers were combined. Its core technology is to perform convolution operations at different

scales in the research area to extract the feature information of the corresponding scale. The

convolutional layer is a special network layer. First, the input data is filled with bounds (0

supplemented) and then convolved, resulting in multiple 16*2^L (L is the number of sequential

layers, take 0, 1, 2....) size of 3X3 and a step size of 1. Each feature is then transformed using a

nonlinear activation function (ReLU) (blue arrow in Figure 1).

Page 3 of 7

563

Hua, G., Jie, L., Dongling, B., Wang, Z., Qi, Y., & Zuofang, Z. (2024). Revised Test of Precipitation Forecasting Product of RMAPS-ST System Based on

U-Net. European Journal of Applied Sciences, Vol - 12(1). 561-567.

URL: http://dx.doi.org/10.14738/aivp.121.16248

The down sampling layer (pooling layer, red arrow in Figure 2) is located after the

convolutional layer, which reduces the dimensionality of the data and extracts the multi-scale

information of the data, and the result is finally output to the next set of transformations.

The up-sampling layer (transpose convolution, green arrow in Figure 2) interpolates the data

processed in the previous stage in order to bring the data to the initial number of grid points.

The structure of the fully connected layer is the same as that of the multilayer perceptron. By

decomposing the feature information of the original data at scale, training can be performed

more efficiently and accurately in the fully linked link. The 3-hour cumulative precipitation

forecasts of RMAPS-ST and ECMWF were trained as ensemble samples, and then RMAPS-ST

was trained within 3 of the 24-hour time, Hourly cumulative precipitation forecasts are revised.

Figure 1: U-Net flowchart

Compared with the conventional MSELoss loss function, the above function has a stronger

constraint on small precipitation and weakens with the increase of precipitation level. The

function is to reduce the area covered by the revised precipitation of small magnitude, thereby

reducing the air reporting rate.

DATA

The high-resolution multi-source land surface fusion data (CLDAS) released by the China

Meteorological Administration was used as the grid live field, and the RMAPS-ST precipitation

forecast data developed by the Beijing Urban Meteorological Research Institute and the

European central fine grid precipitation data were used as the grid forecast field Algorithm

parameters for training. The CLDAS dataset uses multiple grid variational assimilation

(STMAS), optimal interpolation (OI), probability density function matching (CDF), and uses

multiple grid variational assimilation (STMAS), optimal interpolation (OI), Developed by

physical inversion, terrain correction and other technologies, the quality in China is better than

that of similar international products, and the spatiotemporal resolution is higher. The range

covers the Asian region (0-65°N, 60°°E-160°E° spatial resolution of 0.0625°×0.0625°, temporal

resolution of 1 hour. The forecast data were based on the RMAPS-ST system and ECMWF