DATA-DRIVEN SHORT -TERM DAILY OPERATIONAL SEAICEREGIONAL FORECASTING Timofey Grigoryev Ilya Troﬁmov Nikita Balabin Evgeny Burnaev Vladimir Vanovskiy

2025-05-06 0 0 4.49MB 18 页 10玖币

侵权投诉

DATA-DRIVEN SHORT-TERM DAILY OPERATIONAL

SEA ICE REGIONAL FORECASTING

Timofey Grigoryev∗

, Ilya Troﬁmov, Nikita Balabin, Evgeny Burnaev, Vladimir Vanovskiy

Applied AI Center

Skolkovo Institute of Science and Technology

Moscow, Russia

{t.grigorev, ilya.trofimov, nikita.balabin, e.burnaev, v.vanovskiy}@skoltech.ru

Polina Verezemskaya, Mikhail Krinitskiy, Alexander Gavrikov, Sergey Gulev

Shirshov Institute of Oceanology

Moscow, Russia

{verezem, krinitsky, gavr, gul}@sail.msk.ru

Nikita Anikin

Moscow Institute of Physics and Technology

Dolgoprudny, Russia

anikin.nn@phystech.edu

Aleksei Shpilman, Andrei Eremchenko

Gazprom Neft

St. Petersburg, Russia

{Shpilman.AA, Eremchenko.AYu}@gazprom-neft.ru

ABSTRACT

Global warming made the Arctic available for marine operations and created demand for reliable

operational sea ice forecasts to make them safe. While ocean-ice numerical models are highly

computationally intensive, relatively lightweight ML-based methods may be more efﬁcient in this

task. Many works have exploited different deep learning models alongside classical approaches

for predicting sea ice concentration in the Arctic. However, only a few focus on daily operational

forecasts and consider the real-time availability of data they need for operation. In this work, we

aim to close this gap and investigate the performance of the U-Net model trained in two regimes for

predicting sea ice for up to the next 10 days. We show that this deep learning model can outperform

simple baselines by a signiﬁcant margin and improve its quality by using additional weather data and

training on multiple regions, ensuring its generalization abilities. As a practical outcome, we build a

fast and ﬂexible tool that produces operational sea ice forecasts in the Barents Sea, the Labrador Sea,

and the Laptev Sea regions.

Keywords:

data-driven models

short-term sea ice forecasting

deep learning

computer vision

U-Net

remote

sensing ·satellite imagery analysis ·Arctic sea ice

1 Introduction

The rapid Arctic warming [

] is characterized by a twice as substantial temperature increase compared to the global

mean [

]. According to the ERA5 reanalysis, the annual Arctic warming trend from 1979 to 2020 is 0.72

°C

/decade

[

], which is 2-3 times stronger than the global mean. Arctic rapid warming is closely associated with an unprecedented

decline of sea ice extent by more than 30% over the last four decades [

], and a decrease of sea ice thickness [

These changes allow for faster and cheaper sea routes, such as the Northeast Passage [

]. Sea ice jams are one of the

most critical problems in marine navigation security. Accurate operative forecasts of sea ice properties and dynamics

can mitigate that problem, allowing ships to adjust their routes to avoid regions of ice accumulation. At the same time,

∗Correspondence: timofey.a.grigoryev@gmail.com

arXiv:2210.08877v1 [cs.LG] 17 Oct 2022

Data-Driven Sea Ice Forecasting Timofey Grigoryev et al.

new routes through the Arctic will cause an increase in the ocean and atmospheric pollution risks, primarily due to

ﬁshing, oil/gas extraction, and transportation. For the delivery of natural gas and oil to long-distance destinations,

transport by deep-sea vessels is more economical compared to offshore pipelines [

]. To decrease ocean pollution and

the carbon footprint [

] caused by transportation, gas/oil companies must optimize the routes [

] to make them

faster and to reduce associated ecological risks (for example, reduce the atomic icebreaker usage).

Coupled ocean-ice numerical modeling is the evident source of a reliable forecast of sea ice conditions. Newest sea

ice models, such as NextSIM [

] demonstrate fascinating results on sea ice concentration, thickness and drift

vectors representation comparing to the observational data (OSI SAF SSMI-S [

], AMSR2 [

], GloblICE dataset,

http://www.globice.info

). NextSIM is a fully-Lagrangian ﬁnite-element model, making it tough to couple with

Euler method-based ocean models. Eulerian sea ice models have been evolving for the last two decades and can

reproduce some aspects of sea ice and its recent changes. However, detailed comparisons between satellite remote

sensing data with Eulerian-model results reveal big differences in certain aspects of the sea ice cover, e.g., for fracture

zones and small-scale dynamic processes [

]. It remains unclear whether the current model physics (elastic-

viscous-plastic rheology) is suitable for reproducing these observed sea ice deformation features [

] and

provides a reliable forecast. Furthermore, coupled ocean-ice numerical modeling requires signiﬁcant computational

resources.

Statistical or data-driven machine learning approaches, on the other hand, are more ﬂexible and lightweight. They

do not need a complex physical model of processes in the ocean and atmosphere to work. Once trained, such a

model only needs appropriate recent observations and comparatively little computational resources to make a forecast.

However, the training part in this case is quite difﬁcult for several reasons. First, most of the input data used for training

(including sea ice concentration) is presented as 3d or even 4d spatiotemporal maps with a huge amount of highly

correlated input channels. It has been found, that usage of modern convolutional [

], recurrent [

]

or attention-based [

] architectures can overcome difﬁculties associated with exploding number of trainable

parameters and overﬁtting. Second, the model’s output is expected to be a consistent SIC forecast retaining the same

spatiotemporal nature, which is hard to guarantee when training on a limited amount of data. In order to overcome

these difﬁculties, one can train a model not to predict the data itself but to compensate for the errors of simple baselines,

such as climatology mean, persistence, or cell-wise linear trend. Finally, operative climate and sea ice characteristics

data have their peculiarities. It usually consists of several patches obtained at different times each day, thus should be

combined and averaged daily. SIC can only be measured in the sea, leaving the land cells blank. Measurements can be

based on different sources inheriting different biases, making the signal-to-noise ratio lower than expected. Furthermore,

the actual changes in the sea ice condition occur in limited periods in fall and spring, making more than half of the data

barely usable. Considering everything above, one must be very thoughtful when designing training and testing pipelines

and choose proper metrics to assess obtained solutions adequately.

Many works are dedicated to sea ice forecasting in the Arctic region. However, research in this ﬁeld mainly focuses on

climate studies rather than operative sea ice forecasts for practical use. Fully-connected MLP is often used either as

the primary method for predicting monthly-averaged sea ice concentration [

] or as one of the benchmarks [

NSIDC Nimbus-7 SMMR and DMSP SSMI/SSMIS data are used there as SIC maps. Other approaches exploit

CNN, applied on patches cropped out of ice maps [

], or RF with an additional set of weather input features from

ERA-Interim [

]. Deep learning methods are compared with simpler baselines in these works and reported to perform

signiﬁcantly better in standard metrics, such as RMSE. Works [

] are of particular interest, as they consider more

advanced deep learning models that seem more suitable for sea ice forecasting. In [

] authors consider ConvLSTM

[

] model, which can fully make use of spatial-temporal structure of the climatological data. However, they use

weather maps (predictors) from ERA-Interim and ORAS4 NEMO reanalysis data for training, thus limiting model

applicability for operational sea ice forecasts. Authors evaluate the performance of ConvLSTM on a weekly-averaged

and monthly-averaged scale and obtain results comparable in terms of RMSE to those of the ECMWF numerical climate

model only for short lead times. Authors of [

] deal with U-Net [

] model and train it to predict probabilities for the

next 6 months for monthly-averaged SIC values in each cell to belong to each of three classes: open water, marginal ice

and packed ice. They thoroughly investigate the model properties and compare it with SEAS5, a numerical ocean-ice

model with state-of-the-art sea ice prediction skills. However, the paper does not consider possibilities for operating at

the daily temporal resolution.

In our work, we focus on the operative daily sea ice forecasting and imply corresponding restrictions on the weather and

sea ice data we use. To our knowledge, only a few papers consider this type of setting. However, all these works either

use non-operative reanalysis data or perform experiments with one or two currently outdated machine learning methods.

For example, [

] demonstrates the potential of machine learning in sea ice forecasting by comparing a numerical

ocean-ice model with simple CNN and cell-wise k-NN method. Unlike previous works, it focuses on short-term

predictions with a length of 1-4 weeks. [

] assesses the ability of different cell-wise GRU networks equipped with

feed-forward encoder and decoder to forecast SIC for up to the next 15 days. To overcome limitations of locality in

Data-Driven Sea Ice Forecasting Timofey Grigoryev et al.

this setting, the authors incorporate global statistics in the network inputs and report signiﬁcant improvement in the

prediction accuracy. Authors of [

] demonstrate the superiority of ConvLSTM over CNN when forecasting SIC data

in a patch-wise manner with patches of size 41 by 47 pixels. They use only NSIDC Nimbus 7 and DMSP SMMR

SIC data, which is available operatively but has a low resolution (25

25 km) to be of actual use in navigation, and

forecast daily-averaged SIC for the next 10 days. In [

] authors investigate variations of relatively modern PredRNN++

[

] architecture for SIC forecasting for the next 9 days and compare it to the ConvLSTM network, demonstrating the

superiority of the former. However, their model depends on ECMWF ERA5 reanalysis data, which is not available in

real-time, and thus limits its practical value.

In this work, we provide a thorough research on the prospects of machine learning in sea ice forecasting in a few regions

in the Arctic: the Barents and Kara Seas (Barents), the Labrador Sea (Labrador), and the Laptev Sea (Laptev). These

three regions demonstrate varying SIC inter-annual dynamics and allow the investigation of the model’s performance in

different conditions. We deal with SIC and weather data that is available in real-time and can be used in practice to

obtain operational SIC forecasts for marine navigation. A single simple yet effective classical U-Net neural architecture

is chosen as such a model. It is lightweight, thus not prone to overﬁtting, and suited well for image-to-image tasks, such

as sea ice forecasting. We treat JAXA AMSR-2 Level-3 imagery as ground truth of sea ice concentration maps and

train, validate and test our models on this data. As a result, we not only obtain a trained U-Net model for the operational

sea ice forecasts but also provide datasets we used for benchmarks and future comparisons for the research community.

All similar works test their models with different satellite data in different regions during different periods over varying

baselines and usually report improvement in MAE in the range 25% – 50% over considered baselines. Though the

comparison with them hardly makes sense, we obtain similar daily improvement over persistence around 25% in all

three regions.

The main contributions of our work are the following:

We collect JAXA AMSR-2 Level-3 SIC data and GFS analysis and forecasts data, process it and construct

three regional datasets, which can be used as benchmark tasks for future research.

We conduct numerous experiments on forecasting SIC maps with the U-Net model in two regimes and provide

our ﬁndings on the prospect of this approach, including comparison with standard baselines, standard metric

values, and model generalization ability.

We build a fast and reliable tool — trained on all three regions U-Net network that can provide operational sea

ice forecasts in these Arctic regions.

We compare U-Net performance in forecasting in recurrent (R) and straightforward (S) regimes and highlight

the strength and weaknesses of both.

2 Data

2.1 Sea Ice Data (JAXA AMSR-2 Level-3)

Plenty of sea ice concentration products are available, covering a period from the very beginning of the satellite era

to nowadays [

]. Most of them are available daily on the regular grid with spatial resolution varying from

12.5 to 50 km [

]. We were looking for a higher resolution satellite product to provide a forecast comparable to

high-resolution ocean sea ice modeling. We present our analysis of the sea ice conditions based on JAXA (

http:

//ftp.eorc.jaxa.jp

) as a High-Resolution Sea Ice Concentration Level-3 from Advanced Microwave Scanning

Radiometer-2 (AMSR2 hereafter) from the GCOM-W satellite. Daily sea ice concentration is available since July

2, 2012, with a spatial resolution of 5 km on the regular grid, which is the highest resolution compared to other SIC

datasets, that we could ﬁnd in openly available products (CRYOSAT, AMSR2 L4, SSMI, and other). SIC data is given

in percentages (%) from 0 to 100. We used data from July 2, 2012, to January 20, 2022, i.e., 6970 days. AMSR2 L3

research product of SIC is distributed in two daily entities corresponding to the composites combined from the data

acquired during ascending and descending satellite passes. In our study, we use the mean between these two daily

snapshots as the ones statistically closer to ground truth compared to individual composites.

Monthly statistical distributions of SIC are presented in ﬁgure 1 and its climatological anomalies in ﬁgure 2 (see

subsection 3.2 for details of its computation). Changes in sea ice are present only in about half of the months during the

year. In the remaining time, the regions are fully melted (Barents, Labrador) or fully frozen (Laptev).

2.2 Weather Data (GFS)

While sea ice concentration describes its condition and dynamics, there is an opportunity for potential improvement of

a statistical model using additional variables correlated with sea ice dynamics. For example, surface winds inﬂuence

Data-Driven Sea Ice Forecasting Timofey Grigoryev et al.

Figure 1: Box and whisker plots of SIC data distribution in JAXA for different months of 2021, aggregated for all the

cells in each region. The box extends from the 25th percentile to the 75th percentile; whiskers extend the box by 1.5x of

its length. The orange line is the median (50th percentile); outliers are omitted in order not to clutter the plot.

Figure 2: Box and whisker plots of SIC climatological anomaly distribution in JAXA for different months of 2021,

aggregated for all the cells in each region. Climatological anomaly is the difference between the data and the climatology

of the respective channel (see subsection 3.2). The box extends from the 25th percentile to the 75th percentile; whiskers

extend the box by 1.5x of its length. The orange line is the median (50th percentile); outliers are omitted in order not to

clutter the plot.

sea ice drift, especially in shallow seas. Surface air temperature may also impact sea ice dynamics through ice melting

or growth. Our study explored the potential for improving data-driven SIC forecast by extending input features with

atmospheric properties such as 2-meter temperature, surface pressure, and u, v components of wind and its absolute

speed.

We used NCEP operational Global Forecast System (GFS) for atmospheric and ice condition data. The GFS core is

based on coupled atmospheric-ocean-ice models and provides an analysis and forecast globally at 0.25

◦

horizontal

resolution and 127 vertical levels (for atmosphere) [

]. Model forecast runs up to 16 days in advance at a 3 hourly time

steps interval at 00, 06, 12, and 18 UTC daily. The output is openly available in WMO GRIB2 format. It is available

with no delay and a minimal number of temporal gaps, making it the best choice between the weather data sources for a

reliable operative forecasting system.

2.3 Regions

We conduct experiments on three regions with varying SIC inter-annual dynamics (ﬁgure 3). It allows us to enlarge

the dataset and to adjust the model for different sea ice conditions. The Labrador Sea presents the Atlantic type of

ice regime, characterized by the shortest period of pack ice in the basin (1-3 months) with mean SIC below 50%

during the coldest month in a year. High interannual variability of the SIC in the Labrador Sea is caused by the sea

ice fragmentation observed in the marginal ice zone (15-80%, MIZ hereafter) due to the ocean-wave-ice-atmosphere

interaction. The Laptev Sea shows the typical inner-Arctic icing: SIC is above 90% spanning 7-9 months in the annual

sea ice duration with strong sea ice freeze-up and slight sea ice freeze-up. The highest inter-annual variability is

observed in summer, resulting in large open-water areas (SIC



80%). The Barents Sea and the Kara Sea regions are a

mixture of these two types. The Barents Sea is an ice-free region due to the inﬂuence of intense warming from the

North Atlantic Current. On the other hand, the Kara Sea is separated by the island of Novaya Zemlya and is similar to

the Laptev Sea.

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：

DATA-DRIVENSHORT-TERMDAILYOPERATIONALSEAICEREGIONALFORECASTINGTimofeyGrigoryev,IlyaTromov,NikitaBalabin,EvgenyBurnaev,VladimirVanovskiyAppliedAICenterSkolkovoInstituteofScienceandTechnologyMoscow,Russiaft.grigorev,ilya.trofimov,nikita.balabin,e.burnaev,v.vanovskiyg@skoltech.ruPolinaVerezemskaya,Mi...

展开>> 收起<<

DATA-DRIVEN SHORT -TERM DAILY OPERATIONAL SEAICEREGIONAL FORECASTING Timofey Grigoryev Ilya Troﬁmov Nikita Balabin Evgeny Burnaev Vladimir Vanovskiy.pdf

共18页,预览4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！

DATA-DRIVEN SHORT -TERM DAILY OPERATIONAL SEAICEREGIONAL FORECASTING Timofey Grigoryev Ilya Troﬁmov Nikita Balabin Evgeny Burnaev Vladimir Vanovskiy

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: