Citation: Hu, Y.; Chen, L.; Wang, Z.;
Pan, X.; Li, H. Towards a More
Realistic and Detailed Deep-
Learning-Based Radar Echo
Extrapolation Method. Remote Sens.
2022, 14, 24. https://doi.org/
10.3390/rs14010024
Academic Editors: Yangquan Chen,
Subhas Mukhopadhyay, Nunzio
Cennamo, M. Jamal Deen, Junseop
Lee and Simone Morais
Received: 29 November 2021
Accepted: 21 December 2021
Published: 22 December 2021
Publisher’s Note: MDPI stays neutral
with regard to jurisdictional claims in
published maps and institutional affil-
iations.
Copyright: © 2021 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
4.0/).
Article
Towards a More Realistic and Detailed Deep-Learning-Based
Radar Echo Extrapolation Method
Yuan Hu
1,
*
,†
, Lei Chen
1,†
, Zhibin Wang
1
, Xiang Pan
1,2
and Hao Li
1
1
DAMO Academy, Alibaba Group, Beijing 100102, China; fanjiang.cl@alibaba-inc.com (L.C.);
zhibin.waz@alibaba-inc.com (Z.W.); panxiang@smail.nju.edu.cn (X.P.); lihao.lh@alibaba-inc.com (H.L.)
2
Key Laboratory of Mesoscale Severe Weather/MOE, School of Atmospheric Sciences, Nanjing University,
Nanjing 210033, China
* Correspondence: lavender.hy@alibaba-inc.com; Tel.: +86-8116-9963
† These authors contributed equally to this work.
Abstract:
Deep-learning-based radar echo extrapolation methods have achieved remarkable progress
in the precipitation nowcasting field. However, they suffer from a common notorious problem—they
tend to produce blurry predictions. Although some efforts have been made in recent years, the
blurring problem is still under-addressed. In this work, we propose three effective strategies to
assist deep-learning-based radar echo extrapolation methods to achieve more realistic and detailed
prediction. Specifically, we propose a spatial generative adversarial network (GAN) and a spectrum
GAN to improve image fidelity. The spatial and spectrum GANs aim at penalizing the distribution
discrepancy between generated and real images from the spatial domain and spectral domain,
respectively. In addition, a masked style loss is devised to further enhance the details by transferring
the detailed texture of ground truth radar sequences to extrapolated ones. We apply a foreground
mask to prevent the background noise from transferring to the outputs. Moreover, we also design
a new metric termed the power spectral density score (PSDS) to quantify the perceptual quality
from a frequency perspective. The PSDS metric can be applied as a complement to other visual
evaluation metrics (e.g., LPIPS) to achieve a comprehensive measurement of image sharpness. We
test our approaches with both ConvLSTM baseline and U-Net baseline, and comprehensive ablation
experiments on the SEVIR dataset show that the proposed approaches are able to produce much
more realistic radar images than baselines. Most notably, our methods can be readily applied to any
deep-learning-based spatiotemporal forecasting models to acquire more detailed results.
Keywords:
realistic radar echo extrapolation; generative adversarial networks; style loss; power
spectral density
1. Introduction
Precipitation nowcasting, especially very-short-term (e.g., 0~3 h) forecasting, has
attracted much research interest in recent years, as it is beneficial to many practical ap-
plications such as thunderstorm alerting, flight arrangement, public decision making,
etc. Precipitation nowcasting is mostly performed based on extrapolation of observa-
tion data, such as radar echo maps [
1
,
2
]. Traditionally, the extrapolation of radar echo
images is conducted either by storm-tracking methods [
3
–
5
] or optical flow-based meth-
ods [
6
,
7
]. These methods often work well for capturing simple advection characteristics,
whereas they struggle to predict more complex evolution of the precipitation system
(e.g., convective development).
Recently, with strong power to extract features from ever-increasing streams of geospa-
tial data [
8
], deep learning (DL) has been successfully applied to solving remote sensing
problems, like vegetation detection [
9
] and building extraction [
10
]. For precipitation
nowcasting, DL-based methods have also achieved noticeably good performance, and
significantly surpass numerical weather prediction (NWP) and traditional extrapolation
Remote Sens. 2022, 14, 24. https://doi.org/10.3390/rs14010024 https://www.mdpi.com/journal/remotesensing
