Deep-Learning-Based Precipitation Nowcasting with Ground Weather Station Data and Radar Data Jihoon Ko Kyuhan Lee Hyunjin Hwang and Kijung Shin

2025-05-06 0 0 1.61MB 8 页 10玖币

侵权投诉

Deep-Learning-Based Precipitation Nowcasting with

Ground Weather Station Data and Radar Data

Jihoon Ko∗, Kyuhan Lee∗, Hyunjin Hwang, and Kijung Shin

Kim Jaechul Graduate School of AI

Korea Advanced Institute of Science and Technology

Seoul, South Korea

{jihoonko, kyuhan.lee, hyunjinhwang, kijungs}@kaist.ac.kr

Abstract—Recently, many deep-learning techniques have been

applied to various weather-related prediction tasks, including

precipitation nowcasting (i.e., predicting precipitation levels and

locations in the near future). Most existing deep-learning-based

approaches for precipitation nowcasting, however, consider only

radar and/or satellite images as inputs, and meteorological

observations collected from ground weather stations, which are

sparsely located, are relatively unexplored. In this paper, we

propose ASOC, a novel attentive method for effectively exploiting

ground-based meteorological observations from multiple weather

stations. ASOC is designed to capture temporal dynamics of

the observations and also contextual relationships between them.

ASOC is easily combined with existing image-based precipitation

nowcasting models without changing their architectures. We

show that such a combination improves the average critical

success index (CSI) of predicting heavy (at least 10 mm/hr) and

light (at least 1 mm/hr) rainfall events at 1-6 hr lead times

by 5.7%, compared to the original image-based model, using

the radar images and ground-based observations around South

Korea collected from 2014 to 2020.

Index Terms—precipitation nowcasting, ground-based meteo-

rological observations, attention mechanism with sparse features

I. INTRODUCTION

Recently, deep learning techniques, especially computer

vision techniques, have been applied to forecasting various

weather-related events, and such approaches [1]–[6] often

outperform traditional methods in the ﬁeld. A representative

example is precipitation nowcasting, which is short-term (e.g.,

at 0-6hour lead times [7], [8]) location-speciﬁc forecasting of

precipitation. According to [9], current approaches equipped

with deep convolutional networks outperform HRRR [10],

which is one of the state-of-the-art numerical weather pre-

diction models, with lead times up to 12 hours.

For precipitation nowcasting, U-Net [11] and ConvLSTM

[1], which were originally designed for semantic segmentation

and spatio-temporal sequence forecast, respectively, have been

used mainly as backbone network architectures. For example,

Agrawal et al. [4] and Lebedev et al. [6] adapted U-Net and

used radar-reﬂectivity images and satellite images as inputs.

Ko et al. [3] also adapted U-Net to demonstrate the effective-

ness of their proposed training strategies for deep-learning-

based precipitation nowcasting. Shi et al. [1] demonstrated

that ConvLSTM outperforms optical-ﬂow-based methods and

∗Equal contribution.

fully-connected LSTM on precipitation nowcasting. In or-

der to improve the performance of precipitation nowcasting,

ConvLSTM was extended to learn additional location-variant

structures [2], and it was also extended with exponentially

dilated convolution blocks, which enhance expressive power

by capturing additional spatial information [9].

Most deep-learning-based approaches (e.g., [1]–[4], [6])

consider only radar images and satellite images as inputs,

and meteorological observations from ground weather stations

have been underutilized. Although radar and satellite images,

which are in a grid format, are naturally fed into deep

convolutional neural networks (e.g., U-Net and ConvLSTM),

ground-based meteorological observations are not naturally

represented in a grid format since ground weather stations

are sparsely located. Although interpolation techniques, such

as Inverse Distance Weighting [12] and Kriging [13], can be

used to resolve this issue, they are expensive both in time

and memory, especially to obtain high-resolution data. Thus,

in order to utilize ground-based meteorological observations

together with radar and satellite images, deep-learning models

should be able to utilize input data of different formats

efﬁciently and effectively.

In this paper, to address the aforementioned challenge, we

propose Attentive Sparse Observation Combiner (ASOC), a

novel deep-learning model for precipitation nowcasting based

on meteorological observations collected from multiple ground

weather stations. In a nutshell, ASOC combines LSTM [14]

and Transformer [15] to capture temporal dynamics of the

ground-based observations and also contextual relationships

between them. Speciﬁcally, ASOC uses LSTM, which pro-

cesses observations in chronological order, to capture temporal

dynamics, and it uses Transformer-style attention blocks be-

tween LSTM cells to capture contextual relationships between

observations. Another advantage of ASOC is that it is easily

combined with existing image-based models, without any

change in their design. In our experiments, we use ASOC+,

where ASOC is combined with DeepRaNE [3], which is one

of the state-of-the-art image-based models.

We evaluate our approaches using radar-reﬂectivity images

and ground-based observations (from 714 weather stations)

around South Korea collected for seven years (spec., from

2014 to 2020). We demonstrate that ASOC+ improves the

average critical success index (CSI) of predicting heavy (≥10

arXiv:2210.12853v1 [physics.ao-ph] 20 Oct 2022

mm/hr) and light (≥1mm/hr) rainfall events at 1-6 hr lead

times by 5.7%, compared to DeepRaNE. For reproducibility,

we made the source code used in the paper publicly available

at https://github.com/jihoonko/ASOC.

In Section II, we brieﬂy review related studies. In Sec-

tion III, we introduce the notations used in this paper and

deﬁne the precipitation nowcasting problem. In Section IV, we

present ASOC and ASOC+. In Section V, we provide experi-

mental results. Lastly, in Section VI, we provide conclusions.

II. RELATED WORK

In the machine-learning literature, precipitation nowcasting

is often formulated as pixel-wise classiﬁcation of precipitation

levels in the near future from input radar-reﬂectivity images,

and satellite images are often used additionally as inputs.

Among convolutional neural networks (CNNs), U-Net [11]

has been widely used for precipitation nowcasting [3], [4], [6],

[16]. U-Net was originally designed for an image segmentation

task, i.e., pixel-wise classiﬁcation. For example, Lebedev et al.

[6] used U-Net for precipitation detection, which is formulated

as a pixel-wise binary-classiﬁcation problem. Agrawal et al.

[4] divided precipitation levels into four classes and used U-

Net for pixel-wise multiclass classiﬁcation. Based on a similar

multiclass classiﬁcation formulation, Ko et al. [3] proposed

training strategies for precipitation nowcasting (spec., a pre-

training scheme and a loss function) and demonstrated their

effectiveness using a U-Net-based model.

Moreover, in order to aggregate both spatial and temporal

information, there have been several attempts to combine

recurrent neural networks (RNNs) (e.g., LSTM [14]) into

CNNs [1], [2], [5]. For example, Shi et al. [1] proposed

ConvLSTM, which has convolutional structures in the input-

to-state and state-to-state transitions in LSTM. Shi et al. [2]

extended ConvLSTM to TrajGRU, which can learn location-

variant connections between RNN layers. Sønderby et al. [5]

proposed MetNet, which uses ConvLSTM as its temporal

encoder and adapts axial attention structure for its spatial

encoder. Ravuri et al. [16] pointed out that deep-learning

based approaches tend to provide blurry predictions, especially

at long lead times, and they used a conditional generative

adversarial network [17] consists of ConvGRU-based [18]

generator and the spatial and temporal discriminators to ad-

dress this limitation. Espeholt et al. [9] extended ConvLSTM

with exponentially dilated convolution blocks, which enhance

expressive power by capturing additional spatial information.

Several studies utilized meteorological observations from

multiple weather stations as inputs to predict weather-related

events. For example, Seo et al. [19] considered temperature

forecasting. They generated a graph, where each node corre-

sponds to a weather station, and inferred the data quality of

each station, during training, by applying the graph convo-

lutional network (GCN) to the generated graph. Wang et al.

[20] focused on short-term intense precipitation (SIP) now-

casting. They generated a graph and its feature, by identifying

and clustering convective cells from radar-reﬂectivity images,

and used them, together with ground-based observations, as

TABLE I

FREQUENTLY USED NOTATION.

Notation Description

ttime (unit: minutes)

R(t)

x∈Rradar reﬂectivity at time tin each region x(unit: dbZ)

R(t)radar reﬂectivity image at time tin all regions

Iset of regions where ground weather stations are located

O(t)

x∈Rdground-based observations at time tin each region x

O(t)ground-based observations at time tin all regions

C(t)

xground-truth precipitation class at time tin each region x

C(t)

xpredicted probability distribution over

precipitation classes at time tin each region x

the inputs of a random forest classiﬁer. In contrast to our

deep-learning-based approach, they did not employ any deep-

learning techniques to process radar images and ground-based

observations together.

III. BASIC NOTATIONS & PROBLEM DEFINITION

In this section, we introduce basic notations and formulate

the precipitation nowcasting problem.

A. Basic Notations

The frequently-used symbols are listed in Table I. We use

R(t)

x∈Rto indicate the radar reﬂectivity in dBZ at time

tin each region x, and we use R(t)to indicate the whole

radar-reﬂectivity image at time t. We use Ito denote the set

of regions where ground weather stations are located. Then,

O(t)

x∈Rddenotes the ground-based observations in each

region x∈Iat time t, and O(t)denotes the ground-based

observations at time tfrom all regions in I. Lastly, C(t)

indicates the ground-truth precipitation class (see the following

subsection for precipitation classes) in each region x∈Iat

time t, and ˆ

C(t)

xindicates the predicted probability distribution

over all precipitation classes for each region xat time t.

B. Problem Deﬁnition

The goal of precipitation nowcasting is to predict precipita-

tion levels and locations at very short lead times. In this paper,

we formulate the problem as a location-wise classiﬁcation

problem, as in [3]. Speciﬁcally, we split precipitation levels

into three classes: (a) HEAVY for precipitation at least 10

mm/hr, (b) LIGHT for precipitation at least 1 mm/hr but

less than 10 mm/hr, and (c) OTHERS for precipitation less

than 1 mm/hr. We also frequently use a combined class

named RAIN (=HEAVY+LIGHT) for precipitation at least 1

mm/hr. As inputs, we use ground-based observations collected

from multiple weather stations and radar-reﬂectivity images

collected for an hour. We assume that both are collected

every 10 minutes. For example, if we perform prediction at

time tin minutes, the inputs are (a) seven radar reﬂectivity

images at times {t−60, t −50,· · · , t}, i.e., R(t−60),R(t−50),

· · · ,R(t), and (b) seven snapshots of ground observations at

times {t−60, t −50,· · · , t}, i.e., O(t−60), O(t−50),· · · , O(t).

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

Deep-Learning-BasedPrecipitationNowcastingwithGroundWeatherStationDataandRadarDataJihoonKo,KyuhanLee,HyunjinHwang,andKijungShinKimJaechulGraduateSchoolofAIKoreaAdvancedInstituteofScienceandTechnologySeoul,SouthKoreafjihoonko,kyuhan.lee,hyunjinhwang,kijungsg@kaist.ac.krAbstractRecently,manydeep-le...

展开>> 收起<<

Deep-Learning-Based Precipitation Nowcasting with Ground Weather Station Data and Radar Data Jihoon Ko Kyuhan Lee Hyunjin Hwang and Kijung Shin.pdf

共8页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Deep-Learning-Based Precipitation Nowcasting with Ground Weather Station Data and Radar Data Jihoon Ko Kyuhan Lee Hyunjin Hwang and Kijung Shin

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: