Improving Ocean Forecasting Using Deep Learning and Numerical Model Integration

Abstract

In this paper, we propose a novel method to enhance the accuracy of a real-time ocean forecasting system. The proposed system consists of a real-time restoration system of satellite ocean temperature based on a deep generative inpainting network (GIN) and assimilation of satellite data with the initial fields of the numerical ocean model. The deep learning real-time ocean forecasting system is as fast as conventional forecasting systems, while also showing enhanced performance. Our results showed that the difference in temperature between in situ observation and actual forecasting results was improved by about 0.5 °C in daily average values in the open sea, which suggests that cutting back the temporal gaps between data assimilation and forecasting enhances the accuracy of the forecasting system in the open ocean. The proposed approach can provide more accurate forecasts with an efficient operation time.

1. Introduction

In the past 20 years, ocean forecasting systems using data assimilation have been widely used to predict the physical phenomena of the ocean [1,2,3,4]. Most operational ocean forecasting systems assimilate satellite data [2,5,6,7,8]. Satellite data are the most widely used for data assimilation. Sea surface temperature (SST) data play a crucial role in atmospheric, ocean-ice, and coupled systems. Satellite SST data are normally assimilated along with the in situ SST data [9,10].

Synthesized satellite SST data are the accumulation of various satellite data and assimilate in situ observations. Since in situ observation data used for synthesized satellite SST often contain incorrect or missing information, the data are distributed for use after performing quality assurance (QA) and quality control (QC) following a set procedure. During this process, an expert manually performs a manual QC or visual inspection [11,12,13]. Recently, a more efficient method was suggested [14], but has not been widely used yet. Moreover, the synthetic process also takes some computational time [15].

The National Institute of Fisheries Science (NIFS) distributes multi-satellite sea surface temperature data of 1km around the Korean Peninsula by receiving Advanced Very High Resolution Radiometer (AVHRR) satellite data from the National Oceanic and Atmospheric Administration (NOAA). Satellite data from AVHRR equipped with an infrared sensor are distributed approximately two days later after going through the process of assimilating the data of ocean areas that were undetected due to cloud pollution with in situ data or climatology data [16]. Because of this QA/QC process, data from 1–2 days ago are generally used when assimilating satellite data into the operational ocean forecasting system. Therefore, the forecasting simulations do not carry out real-time data assimilation, but hindcast from several days ago [2,17,18,19].

Artificial neural networks, a statistical prediction technique, have been improved with deep learning [20], allowing them to perform highly complex calculations. Therefore, artificial neural networks are now being used to predict various ocean variables, such as water temperature, salinity, ocean currents, etc. [21,22,23,24,25,26,27]. Moreover, various satellite imageries are used for analysis and prediction using deep learning technology [28,29,30,31,32].

Kang et al. [33] proposed a deep-generative-inpainting-network (GIN)-based reconstruction method for sea surface temperature data, and they were able to obtain seamless satellite sea surface temperature data instantly using a deep-learning-based resilience approach. They successfully reconstructed the detailed missing patterns in the cloud regions in infrared satellite sea surface temperature data.

In this study, we reproduced a near-real-time satellite sea surface temperature with the GIN and assimilated with the initial fields of the numerical coastal ocean model in Gyeonggi Bay. Results from the new approach eliminate temporal gaps between the data assimilation and forecasting stage. The proposed system was compared with a control run of the conventional operation, which had a two-day temporal gap.

In this paper, the detailed method and specification of the data are described in Section 2. Section 3 contains the comparison of the results from the control run and the proposed method. A summary and discussion are given in Section 4.

2. Data and Methods

2.1. Data Assimilation Scheme

Most data assimilation schemes for ocean data can be formulated as:

$$X^a=X^b+K\left(Y-H{X}^b\right)$$

(1)

where $X^a$ is the assimilated field, $X^b$ is the background field, K is the weight, Y is the observations, and H is the conversion matrix, respectively.

In this study, we adopted optimal interpolation, and K is defined as:

$$K=B{H}^T{\left(HB{H}^T+R\right)}^{-1}$$

(2)

where B is the forecast error covariance matrix, which is proportional to the distance between grid positions, and R is the observational covariance matrix, respectively. In order to apply this assimilation scheme, this study used the extracted data in the model and restored satellite SST data. As the specific parameter, the extraction step was set to 5 intervals to make each spatial resolution equal to 25 km, which buffered the diameter of the horizontal direction set to 50 km. Then, the data assimilation scheme was applied according to the Gaussian distribution after setting the error ratio between the model and satellite data as 2-to-1.

2.2. Experiment Method

In this study, the SST produced by using a GIN-based reconstruction method for satellite sea surface temperature (GIN-SST) data was assimilated into the ocean model, then we examined the performance of the forecasting system. The forecasting system was built based on the region ocean modeling system (ROMS), a primitive equation, hydrostatic, finite-difference, free-surface model with the general kernel described by Schchepetkin and McWilliams [34]. The performance of the forecasting system was examined with (WGIN) and without (CTRL) the application of an initial field assimilated with the SST produced from the GIN-SST set as the initial date. Following the currently used settings of operational ocean forecasting systems [2,18], CTRL was designed to assimilate satellite SST into the previous state and had a temporal lag of 2 days from the present state (D-2), whereas WGIN assimilated the reconstructed data (GIN-SST) for D-1 and D (Figure 1).

The OSTIA system uses data from a combination of infrared and microwave satellites, as well as in situ data [10]. OSTIA data products are currently the most widely used operational sea surface temperature and sea ice analysis data produced by the Met Office. The OSTIA SST field is produced daily with a resolution of approximately 5 km for the entire region, and it is assimilated by optimal interpolation using 7 types of satellite data (AATSR, AMSR-E, AVHRR-LAC, AVHRR-GAC, SSM/I, SEVIRI, TMI) and sea surface temperature data obtained from buoys and in situ observations (Figure 2).

The water temperature data from the Korea Meteorological Administration’s ocean observation buoys in three locations (Incheon, Deokjeokdo, Oeyeongdo) were used for comparison to verify the performance of the experiment. February, May, August, and November, which are the four seasons of the sea, were set as the comparative experiment period.

2.3. Gyeonggi Bay Numerical Ocean Model

The Gyeonggi Bay model was constructed using the ROMS to generate a sigma vertical grid suitable for the strong tides and shallow depth of the West Sea. The dimension of the model was 534 × 545 in longitude and latitude, respectively. The horizontal grid with a 500 m resolution was generated with 15 vertical layers. The domain and bathymetry of the model is shown in Figure 3.

Lateral boundary input data, such as sea surface height (SSH), temperature, salinity, and velocities, were generated by interpolating the 1/12°-resolution global ocean forecasting data provided by Copernicus Marine Service. The 0.25°-resolution prediction model data produced by the National Centers for Environmental Prediction (NCEP) Global Forecast System (GFS) were interpolated according to the domain of the West Sea model to produce the meteorological boundary data. TPXO (OTPS) provided by Oregon State University was used for tides. TPXO8.0 is the most accurate global ocean tidal model, which provides TOPEX/Poseidon and Jason satellite data, as well as OTIS data [35]. In this study, we adopted 8 tidal components (M2, S2, N2, K2, K1, O1, P1, and Q1).

There is a difference between the two initial fields (WGIN and CTRL). For CTRL, an initial field produced by assimilating OSTIA SST data using the optimal interpolation (OI) at the time of forecast d-2 was used. For WGIN, an initial field produced by assimilating the GIN-SST data from d-day, the starting point of the forecast, and the background (model result) from CTRL using OI were used as the input data.

Table 1. Comparison of the configuration of the two systems, CTRL and WGIN.

Name	CTRL	WGIN
Model	ROMS
Horizontal grid	534(idx) × 535(idy)
Vertical layer	15 sigma layers
Ocean boundary	Myocean (1/12° physics analysis forecast data)
Surface boundary	GFS 0.25° (NCEP)
Tide	OTPS with TPXO8.0
Bathymetry	GBECO
Initial field	OI with OSTIA	OI with GIN-SST
Forecast period	+6 day with hindcast (2 day)	+6 day

shows a comparison of CTRL and WGIN.

2.4. GIN-SST (Instantly Reconstructed SST Field Using GIN)

The undetected areas in the satellite SST fields are instantly restored by the GIN method [33]. The GIN uses random masks to generate artificial missing regions, and the adversarial networks, which consist of a generator and a discriminator, were optimized by loss function values between the original and generated data. The size of the random masks was randomly generated rectangles whose size was randomly determined in the range from 32 px × 32 px to 256 px × 256 px.

The restored SST fields were near-real-time and assimilated into the initial fields of the ocean numerical model. The satellite images are provided at 1 km resolution by the NIFS, which are the synthesis of AVHRR, in situ observations, and climatology. The structure and learning process of the GIN model is shown in Figure 4, and the data including the missing data due to clouds (Level 2) were used as the input image to the GIN-SST model. The learning was carried out in two stages to enhance the accuracy, which consisted of rough learning and fine-tuning.

3. Result

In this study, we used ocean temperature data provided by Korea Meteorological Administration’s (KMA) ocean observation buoys located at Incheon, Deokjeokdo, Oeyeongdo, and the Korea Hydrographic and Oceanographic Agency’s (KHOA) ocean observation buoys located at Pyeongtaek Dangjin Port, Daecheon Beach, and Gunsan Port for verification. The locations of the buoys are shown in Figure 3. Numerical forecasting results from CTRL and WGIN were compared with observations.

3.1. GIN-SST

In order to conduct the experiments for February, May, August, and November of 2018, the missing data were reconstructed from the NOAA Level 2 SST data for each month. Figure 5 is the result (output data) produced by the GIN-SST model from the Level 2 SST images (input data) of the 13th of each month. The result was used as the observation data for data assimilation in the WGIN. Since the 2D image data generated through the GIN-SST model are represented in RGB color, we converted the RGB color into ocean temperatures (Figure 6). SST produced via the GIN showed RMSE differences ranging from 0.8 to 2.05 compared to OSTIA (Figure 5). To examine the performance for each season, the monthly average of the RMSE for February, May, August, and November are provided in Figure 5. The results showed the highest RMSE in August and the lowest in February, respectively.

Comparing the GIN production results and OSTIA SST for each buoy observation, the RMSE for the entire period for observation buoys located off the Gyeonggi coast (Deokjeokdo, Incheon, Oeyeongdo) was 0.34 in the GIN, which was slightly less than OSTIA (0.4), and the coastal observation buoys (Pyeongtaek, Daecheon, Gunsan Port) showed a large difference with 0.82 in the GIN and 0.44 in OSTIA (

Table 2. Comparison of the GIN and OSTIA with buoy observations.

Date	DEOKJEOKDO			INCHEON			OEYEONGDO			RMSE
	Obs	GIN	OST	Obs	GIN	OST	Obs	GIN	OST	GIN	OST
02.13	1.0	1.3	0.5	1.8	1.5	2.5	3.1	2.6	3.7	0.34	0.40
05.13	9.9	9.9	10.7	10.8	10.5	11.1	11.5	11.2	11.8
08.13	23.8	23.8	23.5	27.5	27.8	27.1	30.1	29.8	29.5
11.13	15.6	16.3	15.6	14.5	14.9	14.6	14.9	15.5	15.1
Date	PYEONGTAEK			DAECHEON			GUNSAN-PORT			RMSE
	Obs	GIN	OST	Obs	GIN	OST	Obs	GIN	OST	GIN	OST
02.13	0.4	2.7	1.6	2.0	1.6	2.0	1.1	0.8	2.1	0.82	0.44
05.13	11.9	12.1	12.2	13.7	13.1	13.6	14.1	14.1	14.5
08.13	25.3	23.9	26.5	26.9	26.8	27.4	28.3	25.3	28.4
11.13	15.5	16.1	15.4	15.2	16.1	15.0	14.6	14.7	14.4

). The considerable difference in the RMSE between the coastal and open sea region was due to the amount of cloud in the NOAA SST Level 2 input images. In the West Sea coastal area, sea fog or clouds are frequently formed, and consequently, satellite-mounted infrared sensors are not able to detect the SST. Overall, compared to OSTIA, the GIN method showed better performance in the open sea, but the error in the coast was 0.38 higher than that of OSTIA.

3.2. Result of CTRL and WGIN

3.2.1. Sea Surface Temperature

We compared the CTRL and WGIN in February and August, which showed the smallest monthly average errors between the GIN-SST and OSTIA (

Table 3. Comparison of surface temperature distribution of the WGIN and CTRL.

Date	CTRL	WGIN	DIFF CTRL−WGIN
13 February
13 August

). The results from February showed different characteristics between the near-coast region and the open sea. The WGIN showed lower surface temperatures than CTRL overall; however, due to the positive bias of the reconstructed area around the estuary of the Han River (February in Figure 5), the WGIN showed significantly high temperatures in the estuary of the Han River.

The results from August, as the results from February, showed CTRL with a higher surface temperature distribution in the entire region compared to the WGIN. However, CTRL simulated a lower surface temperature distribution in the boundary region between the coast and the open sea.

3.2.2. Comparison between Forecasts and Buoy Observations

The prediction accuracy of CTRL and the WGIN in the Gyeonggi bay forecasting were compared with buoy observations in times series of daily mean ocean temperature (

Table 4. Time series comparison of CTRL (red), the WGIN (blue), and observations (black) at the Deokjeokdo (left) and Incheon (right) buoy stations, respectively. The top, second from the top, second from the bottom, and bottom figures indicate February, May, August and November, respectively.

Date	DEOKJEOKDO	INCHEON
13–18 February
13–18 May
13–18 August
13–18 November

,

Table 5. Same as Table 4, except for Oeyeondo (left) and Pyeongtaek (right).

Date	OEYEONDO	PYEONGTAEK
13–18 February
13–18 May
13–18 August
13–18 November

and

Table 6. Same as Table 4, except for Daecheon (left) and Gunsan Port (right).

Date	DAECHEON	GUNSAN PORT
13–18 February
13–18 May
13–18 August
13–18 November

). The time series represents the six-day temperature prediction from the 13th to the 18th of each month. The red line represents CTRL, the blue line the WGIN, and the black line the observation data, respectively.

The results of February showed higher bias errors in CTRL, while the WGIN showed better results at five points, except for Deokjeokdo. However, in the results from May, August, and November, the WGIN showed better performances in the open ocean area (Deokjeokdo, Incheon, Oeyeondo), whereas CTRL showed better results in the coastal area (Daecheon, Gunsan, Pyeongtaek).

The higher errors of the WGIN in the coastal areas were correlated with the higher errors of the GIN-SST results in May, August, and November.

Table 7. RMSE with buoy observations of CTRL and the WGIN.

POINT	Open Ocean			Coastal Area
	DEOKJEOKDO	INCHEON	OEYEONGDO	PYEONGTAEK	DAECHEON	GUNSAN
WGIN	2.34	1.58	1.11	1.57	0.65	1.09
TOTAL	1.10			1.44
CTRL	1.3	1.28	0.86	2.17	1.42	0.72
TOTAL	1.68			1.15

represents the RMSE with buoy observations of CTRL and the WGIN. Open ocean and coastal area are distinguished by the distance (10 km) from the coast. At the observational sites in the open ocean, the WGIN showed a lower RMSE (1.10) than CTRL (1.68), whereas in the coastal areas, CTRL (1.15) was better than the WGIN (1.44).

4. Summary and Discussion

In this study, we developed a real-time operational ocean forecasting system using the assimilation of satellite data restored instantly using the GIN. The conventional operation of data assimilation has a temporal delay of several days, due to the complex reconstruction process of satellite data. The proposed system reduced the temporal gaps between the start of prediction and assimilation.

The data assimilation process attempted to minimize the initial condition errors by incorporating observations into the model data. However, due to the delayed provision of observations due to the QA/QC procedure, a forecast would start several days prior to the present state. Therefore, errors increased during the temporal gaps between the data assimilation and the start of forecasting.

The GIN method had structural errors. Firstly, the satellite imagery had noises around the coastal area due to land. Secondly, when the ratio of the missing region was high, the GIN depended on climatological monthly average values and generated more errors. Due to these problems, the WGIN showed enhanced surface temperatures in the open ocean area, whereas CTRL was better in the coastal area.

Recently, deep learning has been widely used for oceanic data analysis. We used the GIN for the reconstruction of satellite images and instantly restored undetected areas due to clouds. Time-series-based deep learning prediction [36] is widely used for marine forecasting. Moreover, spatio-temporal prediction is also available using a deep learning method [37]. Consequently, various prediction data can be assimilated into the numerical forecasting system to enhance the accuracy of the forecasts.

The constraint of assimilated observation was determined by the covariance matrix of observations, which were given by instrumental errors. However, the reconstructed data by the GIN contained additional errors during the generator network process. More investigations are required to clarify the error levels due to the GIN and the covariances of the reconstructed data.

The circulation of Gyeonggi Bay consists of complex processes, such as freshwater discharge from the Han River, strong tidal mixing, etc. More accurate data assimilation methods using realistic background error covariance are necessary for advanced research to reduce the initial condition problems [38].

Finally, since the deep learning method is a linear algebraic approach that carries out a massive number of simple instructions, general-purpose computing on graphics processing units (GPGPU) is widely used for calculation. The critical problem of the GPGPU is the amount of memory, which is 24 GB at most. Consequently, the size of the images is restricted to around 256 × 256. Due to such limitations of the GPGPU, the proposed system can only be applied to bay-scale areas only. However, with the recent development of a system-on-chip (SoC) that integrates a CPU and GPU, the size of the memory available for deep learning operations is expected to expand to hundreds of giga-bytes. Therefore, the utilization of deep learning for large scale array data such as numerical model data and satellite images will increase in the future, and the proposed method can be applied not only for bay-scale forecasting, but also used for wide-area simulation.