Rainfall Prediction Model based on Radar Image Analysis Processing

Oudomseila Phok, Jiwan Lee and Bonghee Hong

Department of Electrical and Computer Engineering, Pusan National University, Busan, Republic of Korea

Keywords: Radar Image, Correlation Analysis, Multiple Linear Regression, Predictive Model, Big Data.

Abstract: The radar image represents the intensity of the rainfall measured at the observatory by the image pixel color

value. It is the goal of this paper to find that the radar image values at a given point calculate the rainfall at a

given time. Correlation analysis between radar images and rainfall data provided by rainfall gauges installed

at very rare intervals is performed first. Based on this correlation analysis, we find out how to calculate the

rainfall in the area where AWS is not installed by radar image. The biggest challenge of this paper is to find

a predictive model of rainfall that takes into accounts the movement patterns of radar images affected by wind

direction, wind speed, temperature and humidity.

1 INTRODUCTION

Rainfall forecasts in the weather forecasts are not al-

ways accurate, so there is great dissatisfaction with

the citizens. Existing rainfall is predicted using radar

reflectivity data and terrestrial meteorological data

collected by weather radar observation network.

However, there is a problem that the accuracy of rain-

fall forecasts judged by rainfall radar images and

forecasters' experience based on surface AWS data is

low as can be seen in Table I.

Table 1: Sample rainfalls of Radar and AWS data.

Date Radar Rain Surface Rain

2017-09-11 6:00 3.85mm 1mm

2017-09-11 6:10 6.8mm 3mm

2017-09-11 6:20 6.53mm 4.5mm

2017-09-11 6:30 20.3mm 4.5mm

This paper aims to develop rainfall forecasting

model through correlation analysis between weather

radar value and ground rainfall observation value.

Simple linear regression analysis shows that it is dif-

ficult to analyze the correlation between radar reflec-

tivity data and surface rainfall.

To conduct the analysis using real data, we used

the weather radar observations in Busan city (KMA,

2017). Actual rainfall radar image value is csv file,

image value is color pixel and classified into 32 kinds

according to rainfall type. The color of the radar im-

age is mm of rainfall per hour and is updated every

10 minutes (KMA, 2017).

Figure 1: Radar Image for representing rainfall (KMA,

2017).

This paper deals with the problem of using rainfall

radar image pixel color values to calculate rainfall at

a given location and to develop a model to forecast

Phok, O., Lee, J. and Hong, B.

Rainfall Prediction Model based on Radar Image Analysis Processing.

DOI: 10.5220/0006809903150319

In Proceedings of the 3rd International Conference on Internet of Things, Big Data and Security (IoTBDS 2018), pages 315-319

ISBN: 978-989-758-296-7

315

future rainfall along the direction of rainfall. To ana-

lyze the relationship between rainfall radar images

and rainfall at specific points, we develop rainfall cal-

culation model by correlation analysis with AWS (au-

tomatic weather observation equipment) (KMA,

2017) and future wind speed-based prediction model

considering various weather variables such as wind

direction, wind speed, temperature and humidity.

Section 2 shows the results of correlation analysis

between the rainfall radar image and the observed

rainfall on the ground. We also define the problem of

finding the best corresponding aerial position where

the rainfall difference between the location of the

ground observation point and the aerial position of the

rainfall radar image is minimized. In Section 3, we

show the difference between the rainfall calculated by

the rainfall radar image and the rainfall measured by

AWS (Automatic Weather System) on the ground.

Based on these experimental results, it is shown that

rainfall forecasting model based on radar image should

be made as non-deterministic function. Section 4 ex-

plains the related research cases and concludes with a

remaining breakthrough research topic in Section 5.

2 CORRELATION COEFFICIENT

ANALYSIS

Figure 2: Correlation Analysis (Correlation Coefficient =

0.8501).

The correlation index between rainfall radar images

and AWS measured rainfall is not constant. The rain-

fall data of the rainy season of July, August, Septem-

ber and October of 2017 were analyzed and the cor-

relation index with the radar image was analyzed and

it was 0.8501 as shown in Figure. 2. The more rain-

fall, the lower the correlation. In order to generate bet-

ter model, we need to consider attributes that could

influence the amount of rain, which are, wind speed,

wind direction., temperature, and humidity.

2.1 Correlation with Wind Speed

We also analyzed the results of the analysis of wind

speed as a weighting factor in the correlation analysis

between rainfall radar image and AWS rainfall data.

As shown in Figure. 3, the correlation index increases

as the wind speed is included as a weight. The radar

rainfall considering the wind speed can be calculated

by following the Eq. (1).

_



=



+(





) (1)

k is a constant to determine how much the wind-

speed should affect the calculation and we used k=12

in this current research. d is determined whether the

rain is decreasing or increasing in the current time

frame. d = -1 in the case of rainfall is decreasing

where d = 1 in the case of rainfall is increasing. The

correlation coefficient between rainfall and rainfall

radar images affected by wind speed slightly in-

creased to 0.8574.

Figure 3: Correlation analysis considering wind speed (Cor-

relation Coefficient = 0.8574).

2.2 Correlation with Wind Direction

The correlation index between the rainfall radar image

considering the wind direction and rainfall was found

to be rather low. Figure 4 shows that correlation coef-

ficient is 0.5321 as a result of correlation index analysis

considering east-southeast (ESE) wind direction.

Figure 4: Correlation analysis using ESE wind direction

(Correlation Coefficient = 0.5321).

IoTBDS 2018 - 3rd International Conference on Internet of Things, Big Data and Security

316

2.3 Correlation with Temperature

We also conducted the research using temperature as

a weighting factor in the correlation analysis between

rainfall radar image and AWS rainfall data. As shown

in Figure. 5, the correlation index increases as the

temperature is included into the calculation. The radar

rainfall considering the temperature can be calculated

by following the Eq. (2).

_



=



−





(2)

k is a constant to determine how much the tem-

perature should affect the calculation. In this experi-

ment we used k = 100. The correlation coefficient be-

tween AWS rainfall and radar image rainfall affected

by temperature has been increased to 0.8754.

Figure 5: Correlation analysis considering temperature

(Correlation Coefficient = 0.8754).

2.4 Correlation with Humidity

In addition to windspeed, wind direction and temper-

ature, we also analyzed the correlation index between

the rainfall radar and AWS rainfall using ASOS hu-

midity data as weight. As shown in Figure. 6, the cor-

relation index increases as the humidity is included

into the calculation. The radar rainfall considering the

humidity can be calculated by following the Eq. (3).

_



=



−

ℎ



(3)

k is a constant to determine how much the tem-

perature should affect the calculation. In this experi-

ment we used k = 100. The correlation coefficient be-

tween AWS rainfall and radar image rainfall affected

by temperature has been increased to 0.8799.

Figure 6: Correlation analysis considering temperature

(Correlation Coefficient = 0.8799).

2.5 MLR Predictive Model

Finally, we used Multiple Linear Regression analyze

to generate a more efficient prediction model, consid-

ering all major factors such as windspeed, tempera-

ture and humidity. This model achieved Correlation

Coefficient result of 0.88. The equation of this model

is detailed in Eq. (4) and Table II.

Y = -2.683 + 0.372X

- 0.002X

+ 0.03X

+ 0.025X

(4)

Table 2: Description of MLR Model variables.

Variable Description

Y Predicted Surface Rain (mm/10min)

Radar Rain (mm/10min)

Windspeed (m/s)

Temperature (°C)

Humidity (%)

A method for finding the correspondence between

the location of the rainfall radar image and the loca-

tion of the rainfall observation point has been re-

quired. Correlation analysis confirmed that the rain-

fall is not the same even if the latitude and longitude

coordinates of the radar image are the same for the

ground rainfall observation point. Through these ex-

periments, we propose a very breakthrough method to

devise a method to locate the radar image nearest to

the rainfall of the latitude and longitude coordinates.

Let P (x, y) be the position of the longitude and

latitude of the observation point of the rainfall on the

surface. Let the observed rainfall measured at the lo-

cation P (x, y) where the rainfall system is installed is

Rainfall Prediction Model based on Radar Image Analysis Processing

317

RF (x, y). Let AP (a, b) be the aerial position of the

radar image. Let the rainfall radar value observed at

AP (a, b) be PR (a, b). It is necessary to obtain the

latitude and longitude coordinates (a, b) of the radar

image in which the difference between RF (x, y) and

PR (a, b) becomes minimum for a given P (x, y). This

analysis is based on the given longitudinal coordi-

nates (x, y). The aerial position (a, b) of the radar im-

age at which the value of min (RF (x, y) - PR (a, b))

would be minimum is obtained and then a new learn-

ing model is newly created for different weather con-

ditions of wind direction, wind speed and temperature

and humidity.

It is necessary to find out the same zone of the

similar rainfall forecasts. In other words, it must be

resolved to find and determine the boundary point

where the minimum of the difference between RF (x,

y) and PR (a, b) is maintained spatially and tempo-

rally. It is a break-through research theme to further

discover that these boundary points are formed differ-

ently depending on the weather changes and the topo-

graphic form.

3 EXPERIMENTS

3.1 Experimental Environment

As we have mentioned in the previous section, in this

experiment, we used the input data used the weather

data in Busan region such as rainfall, windspeed, hu-

midity, wind direction and temperature daily during

July, August, September and October of 2017. To

compare and generate the result of the MLR model,

we pick all of the rainy days from each of those

month. All these data were calculated using method

shown in Section 2.

3.2 Experimental Results

As a result of both regression analysis and correlation

analysis, correlation index between rainfall radar im-

age and AWS image is 0.8 or more. However, it can

be seen that the correlation index changes as the rain-

fall layer image value moves away from the observa-

tion point. When the wind speed, humidity, and tem-

perature are included as a weight, the correlation in-

dex increases but the wind direction does not affect

the correlation coefficient. The most important prob-

lem is to make an optimal algorithm of best matching

a given ground AWS location with the corresponding

aerial rainfall radar image.

Figure 7: Multiple Linear Regression Model Result.

Figure. 7, compare the rainfall estimates using the

correlation-based MLR model with the actual rainfall

using the rainfall radar image value as the x value of

the function. Rainfall radar images are physically and

logically different from aerial rainfall locations as

aerial observations. It is important to develop a rain-

fall prediction model that can accurately calculate the

rainfall at a certain point because the rainfall radar

image is synthesized from radar images produced

from several radar observation devices. The rainfall

prediction model should be a non-deterministic func-

tion that finds the aerial location (a, b) of the radar

image with minimum difference between RF (x, y)

and PR (a, b). Rainfall forecasting model based on

rainfall radar image should be solved by development

of data analysis model considering surrounding ter-

rain in addition to wind velocity, wind direction, and

temperature and humidity. By having this experi-

mental result, we will be able to use it and focus on

increase the accuracy of our predictive model through

further research on analyzing the solution to fix the

radar and AWS location problem.

4 RELATED WORKS

In previous works, some researchers have tried to use

other method to estimate rainfalls. One of the method

is using data mining and deep learning method to pre-

dict the rainfall. The researcher of this method had

tried to combine radar data and AWS data with wind

speed and wind direction during the CNN training

process. However, a beneficial model was not found,

and the research end up using only AWS data. The

model was able to forecast rain in the next hour with

over 90% accuracy but does not provide the amount

of rainfalls data (Suryong, 2017). Another method is

through using road CCTV camera which achieved

over 87% of accuracy. However, the method requires

IoTBDS 2018 - 3rd International Conference on Internet of Things, Big Data and Security

318

CCTV camera in the observation location and the ra-

dar weather data was not utilized to improve the ac-

curacy rate in the estimation process (Lee, 2016).

5 CONCLUSIONS

The purpose of this paper is to solve the method of

forecasting the amount of rainfall on the ground using

radar images. However, it is difficult to find a effec-

tive rainfall forecast model only by analyzing radar

images and existing rainfall data. Finding a big data

analysis model as a method of forecasting the accu-

rate rainfall is a breaking-through theme because the

location of airborne radar image values does not co-

incide with the rainfall point on the ground and rain-

fall drop points depend on the wind direction and

wind speed.

ACKNOWLEDGEMENTS

This work was funded by the Korea Meteorological

Administration Research and Development Program

under Grant KMIPA2017-5030.

REFERENCES

Korea Meteorological Administration: July, August,

September, October 2017 Radar Image, Radar Footage

Open Portal, url: https://data.kma.go.kr.

Korea Meteorological Administration: July, August,

September, October 2017 AWS Data, AWS Observation,

url: http://www.weather.go.kr/weather/observation/

aws_table_popup.jsp.

Kwon Suryong, Jwae Oujong, Han Yelim. 2017. Rainfall

forecasting using data mining and deep learning, Grad-

uation Final Report of Pusan National University.

Jiwan Lee, Yongdeok Shin, Bonghee Hong., Jan. 2016.

Extraction of weather information on road using CCTV

video, IEEE 2016 International Conference on Big

Data and Smart Computing.

Rainfall Prediction Model based on Radar Image Analysis Processing

319