Numerical Weather Prediction of Sea Surface Temperature in South China Sea Using Attention-Based Context Fusion Network

Abstract

Numerical weather prediction of sea surface temperature (SST) is crucial for regional operational forecasts. Deep learning offers an alternative approach to traditional numerical general circulation models for numerical weather prediction. In our previous work, we developed a sophisticated deep learning model known as the Attention-based Context Fusion Network (ACFN). This model integrates an attention mechanism with a convolutional neural network framework. In this study, we applied the ACFN model to the South China Sea to evaluate its performance in predicting SST. The results indicate that for a 1-day lead time, the ACFN model achieves a Mean Absolute Error of 0.215 °C and a coefficient of determination ($R^2$) of 0.972. In addition, in situ buoy data were utilized to validate the forecast results. The Mean Absolute Error for forecasts using these data increased to 0.500 °C for a 1-day lead time, with a corresponding $R^2$ of 0.590. Comparative analyses show that the ACFN model surpasses traditional models such as ConvLSTM and PredRNN in terms of accuracy and reliability.

1. Introduction

Sea surface temperature (SST) is a critical parameter that influences both atmospheric and oceanic processes [1,2,3]. SST modifies air-sea turbulent heat fluxes and serves as an indicator of the thermal state of the upper ocean [4]. It also affects the development and intensity of tropical cyclones [5,6], and it can impact regional ocean biology due to extreme temperatures [7]. Therefore, accurate numerical weather prediction (NWP) of SST is essential for operational forecasting [8,9,10].

Geographically, the South China Sea (SCS) serves as a zonal intermediate area between the Pacific and Indian Oceans. The SCS acts as a bridge for interannual atmospheric patterns and facilitates marine circulation between these two major oceans [11,12]. Furthermore, the land–sea temperature gradient in the SCS influences seasonal wind patterns, including the monsoon [13]. Accurate short-term forecasting of SST is also essential for tropical cyclone predictions [14]. Due to complex multiscale air-sea interactions, short-term SST forecasting in the SCS presents significant challenges. Despite these challenges, the SCS remains an ideal region for conducting short-term SST forecasting experiments.

The NWP of SST can be categorized into two main approaches: data-driven methods and numerical models. Data-driven methods, particularly those utilizing deep learning, are becoming increasingly prominent [15,16,17]. Deep learning excels in capturing nonlinear variations and managing stochastic signals [17,18,19]. Numerical models, such as state-of-the-art oceanic general circulation models, incorporate multiscale oceanic dynamics. However, these sophisticated models still face challenges in computational efficiency, model resolution, and forcing reliability [20,21,22,23].

Both MultiLayer Perceptrons and Long Short-Term Memory (LSTM) networks demonstrate efficiency in time series prediction [24,25,26]. Subsequently, Convolutional LSTM (ConvLSTM) integrates both the convolutional neural network and the recurrent neural network, making it more suitable for spatiotemporal image-like datasets [27,28]. ConvLTSM may also serve as a foundation for more advanced spatiotemporal deep learning models. The Predictive Recurrent Neural Network (PredRNN) enhances ConvLTSM with a hierarchical structure that facilitates information combination across different LSTM layers [29]. The Convolutional Gated Recurrent Unit (GRU) replaces the LSTM with a GRU to explore temporal dependencies in data [30]. Recently, we introduced an advanced deep learning model, namely, the Attention-based Context Fusion Network (ACFN) [31,32]. ACFN features a more complex architecture compared with ConvLSTM and PredRNN, including a novel module called the Attention-based Detail Context Block. This module enhances ACFN’s ability to extract detailed spatiotemporal information from the input data. In SST forecasting experiments conducted in the Bohai Sea, ACFN demonstrated superior performance compared with both ConvLSTM and PredRNN [31].

In the SCS, deep learning models are increasingly demonstrating strong performance in short-term SST forecasting. For instance, Hao et al. utilized ConvLSTM-based models [28], while Xie et al. implemented a 3D U-Net architecture [33]. Miao et al. developed a multivariable convolutional neural network [34], and Zhang et al. adopted an enhanced LSTM model [35].

As a relatively new model, the reliability of ACFN is still in doubt. Questions remain regarding its utility in the SCS and its quantified performance. This paper is motivated by the need to address this gap. Furthermore, despite the availability of over 40 years of satellite SST observations [36] commonly used to train and validate deep learning models [31], the precision of satellite SST data, with errors on the order of O (0.1 °C) [36], still leaves room for improvement in short-term SST forecasting. Fortunately, several in situ buoy SST observations are available in the SCS [37], providing an opportunity to evaluate the model performance against direct measurements. Therefore, the objective of this study is to conduct a comprehensive evaluation of ACFN in the SCS using both satellite and buoy SST data. Specifically, this work focuses on the following three aspects:

• Quantifying the performance of the ACFN model in short-term SST forecasting in the SCS.
• Conducting intercomparisons among multiple deep learning models to assess the advantages of the ACFN model.
• Using third-party in situ buoy SST data to evaluate model performance and present the errors.

2. Data and Methods

2.1. Bathymetry Data

We use bathymetry data from ETOPO 2022, with an initial spatial resolution of 1/60° × 1/60°. The water depth data are used to set the land/sea mask. The data are then resized to a lower resolution of 1/4° × 1/4°, aligning with the computational grid. The research area encompasses the region defined by [107.125°, 122.875°]E × [8.125°, 23.875°]N (Figure 1).

2.2. Satellite SST Data

For training and validating the deep learning models, we use daily SST data from Optimum Interpolation SST (OISST) [36]. OISST integrates both satellite and in situ SST observations, with a spatial resolution of 1/4° × 1/4°.

2.3. In Situ SST Data

In situ buoy SST data were provided by Zhang et al. [37]. Four fixed buoys were deployed in the SCS, located approximately at (117°E, 18°N) with water depths ranging from 2000 to 4000 m (Figure 1). These buoys operated in 2014, and the SST measurements were averaged to produce daily mean time series (Figure A1). To the best of our knowledge, this dataset has not yet been utilized in the OISST product.

2.4. Forecast Models

We utilize three advanced deep learning models for our analysis: ACFN, PredRNN, and ConvLSTM [31,32]. ConvLSTM represents an essential framework that combines convolutional layers with LSTM units, enabling the effective processing of spatiotemporal data by preserving spatial information while modeling temporal dependencies. PredRNN enhances ConvLSTM with a hierarchical structure that facilitates interactions across different LSTM layers, improving the representation of complex temporal relationships. ACFN further integrates an attention-based context fusion block into the ConvLSTM framework, which generates attention maps sequentially across both the channel and spatial dimensions, allowing for the detailed exploration of intricate spatiotemporal correlations between the previous context state and the current input state in ConvLSTM. Details about the architecture and specifications of ACFN can be found in Shi et al. [31].

2.5. Evaluation Metrics

Four skill indices are used for model evaluation: Mean Absolute Error (MAE), coefficient of determination ($R^2$), root-mean-squared-error (RMSE), and correlation coefficient (CC) [32]. Their formulas are as follows:

$$\mathrm{MAE}={\displaystyle \frac{1}{N}}\sum _{i=1}^N|P_i-O_i|,$$

(1)

$$R^2=1-{\displaystyle \frac{{\sum }_{i=1}^N{\left(O_i-P_i\right)}^2}{{\sum }_{i=1}^N{\left(O_i-\overline{O_i}\right)}^2}},$$

(2)

$$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{1}{N}}\sum _{i=1}^N{\left(P_i-O_i\right)}^2},$$

(3)

$$\mathrm{CC}={\displaystyle \frac{{\sum }_{i=1}^N\left(P_i-\overline{P_i}\right)\ast \left(O_i-\overline{O_i}\right)}{\sqrt{{\sum }_{i=1}^N{\left(P_i-\overline{P_i}\right)}^2}\ast \sqrt{{\sum }_{i=1}^N{\left(O_i-\overline{O_i}\right)}^2}}},$$

(4)

where $O_i$ and $P_i$ are observational truth and predicted value, respectively.

2.6. Numerical Experiment

The ground truth for the numerical experiment is satellite SST data from OISST. The dataset spans from 1 January 1982, to 31 December 2022. Figure 2 illustrates our flowchart. In the first step, the deep learning models, which include ConvLSTM, PredRNN, and ACFN, are trained using the OISST data from 1982 to 2013. Next, these models are employed to conduct forecasting experiments for the period from 2014 to 2022. Finally, both OISST and in situ buoy SSTs are utilized to evaluate the skill metrics of the deep learning models. The model configurations for ACFN, PredRNN, and ConvLSTM follow the specifications outlined by He et al. [32]. The input length for the deep learning models is 10 days, and the lead time for numerical predictions ranges from 1 to 10 days.

3. Results

3.1. Characteristics of Regional SST

Climatological SST in the SCS exhibits a pronounced meridional variation (Figure 1a). SST is warmer in the near-tropical regions compared with the subtropical areas, with the isothermal lines (SST contours) generally oriented from the northeast to the southwest. For instance, SST reaches 28 °C in the southwest of the Philippine Islands, while dropping to as low as 22 °C in the Taiwan Strait. In the northern SCS, the variation (standard deviation) of SST is significantly higher than in the central and southern SCS (Figure 1b). Near the northern coast, the standard deviation of SST can reach up to 5 °C, whereas it remains below 1.5 °C around southwest of the Philippine Islands. It is important to note that the seasonal signal was not excluded in the standard deviation computation.

3.2. Example of NWP of SST

Figure 3 presents snapshots of SST forecasts for 2 January 2014. According to the ground truth (OISST), SST exhibits short-term variations. On the initial day ($D_0$), the OISST shows a primarily meridional distribution, with a warmer SST in the southern SCS compared with the northern SCS. The isothermal lines are oriented from the northeast to the southwest. In the southern SCS (latitude band 0°N to 15°N), SST is warmer in the eastern SCS compared with the western SCS. Near the Philippine Islands (120°E, 12°N), SST reaches its highest value in the SCS at 30 °C. In contrast, the SST near the northern coast is relatively cooler, at about 21 °C. On the following day ($D_0$ + 1), significant changes in SST are observed compared with the previous day. In the southwest of the Philippine Islands (120°E, 12°N), SST warms, and the 30 °C isothermal contour expands southwestward. By the third day after the initial day ($D_0$ + 3), SST in the southwest of the Philippine Islands cooled compared with the first day, and the 28 °C isothermal line shifted eastward to around (115°E, 10°N). However, from the sixth to the eighth day after the initial day ($D_0$ + 6 to $D_0$ + 8), the 28 °C isothermal line extended westward again near (115°E, 10°N). In the northern SCS, between days $D_0$ + 8 and $D_0$ + 10, a mesoscale eddy developed near the thermal front around (118°E, 20°N) as time progressed.

Using ACFN, the SST on the initial day ($D_0$; lead time of 0 day) is provided by the ground truth. For the 1-day lead time prediction ($D_0$ + 1), the predicted SST retains the overall spatial pattern. The relatively cold water along the northern near-coast is preserved, and the 25 °C contour line extends from the south of the Taiwan Island in a northeast to southwest direction. However, the area with SSTs exceeding 28 °C is smaller compared with the ground truth. This underestimation of warm water persists for 2-day and longer lead times, with predicted SSTs failing to match the extent observed in the ground truth. Furthermore, for 8-day and longer lead times, ACFN does not capture the mesoscale eddy near 118°E, 20°N. With PredRNN, the predicted SSTs are also colder than the ground truth for lead times of 1 to 3 days. Similar to ACFN, PredRNN does not reproduce the mesoscale eddy around 118°E, 20°N for lead times of 8 to 10 days. In contrast, ConvLSTM predicts SSTs that are notably higher than ground truth for lead times of 1 to 2 days, introducing some noise into the predictions. This noise likely arises from insufficient extraction of spatial modes in the ConvLSTM. For lead times of 3 days and longer, the predicted SSTs in the northern near-coast area align more closely with the ground truth. However, for the 3-day lead time, the 25 °C isothermal line in the western SCS is distributed further north compared with the ground truth.

We also evaluate the regional SST to provide an overall assessment, as shown in Figure 4. In 2014, the ACFN model generally aligned well with the ground truth; however, it exhibited some discrepancies in capturing the SST peak for lead times of 2 days and longer. PredRNN, on the other hand, tends to predict colder SSTs, particularly in spring and early summer (April to June). The ConvLSTM model displays a noticeable warm bias from January to April for 1-day lead times, followed by a cold bias from April to June. For lead times of 3 days or longer, ConvLSTM’s predictions are consistent with the ground truth. Nevertheless, ConvLSTM underestimates the SST peak in June across all lead times (1 day to 10 days).

3.3. Ground Truth Evaluation

The geographic distribution of the overall evaluation is illustrated in Figure 5. The ACFN model demonstrates superior forecasting skill compared with both PredRNN and ConvLSTM. Using the ACFN model, the MAE is higher in the northern near-coast compared with the central and southern SCS (latitude band of 8°N to 15°N). For a 1-day lead time, the MAE in the northern near-coast is approximately 0.5 °C, while in the central and southern SCS, it is around 0.2 °C. The MAE increases with lead time; for a 10-day lead time, MAE in the northern near-coast reaches 1.3 °C, compared with about 0.6 °C in the central and southern SCS. PredRNN shows a larger MAE in most areas of the SCS. For a 1-day lead time, MAE is approximately 0.6 °C. ConvLSTM exhibits the poorest performance in the northern near-coast, with a maximum MAE exceeds 3.0 °C for a 1-day lead time. For a 3-day lead time, the MAE decreases in the northern near-coast compared with the 1-day and 2-day lead times but increases in the central SCS (around 115°E, 13°N). For 4-day and 5-day lead times, the MAE in the central SCS further rises to 1.2 °C. However, for 7-day and longer lead times, ConvLSTM’s MAE improves relative to 4-day to 6-day lead times, with the MAE in the central SCS (around 115°E, 13°N) decreasing to about 0.5 °C for a 10-day lead time.

Figure 6 illustrates the coefficient of determination ($R^2$). For a 1-day lead time, the ACFN model achieves an $R^2$ close to the perfect value of 1. However, $R^2$ gradually decreases with longer lead times. At a 10-day lead time, $R^2$ remains above 0.8 in the northern SCS but drops to 0.6 in the southern SCS. The PredRNN model exhibits lower $R^2$ values compared with ACFN. For a 1-day lead time, $R^2$ in the southern SCS near (117°E, 9°N) is below 0.4, and in the southwest near the Philippine Islands, $R^2$ is generally below 0.6. Nevertheless, in the northern SCS (latitude north of 15°N), $R^2$ is predominantly above 0.8. For lead times of 2 days and longer, $R^2$ continues to decrease. By the 10-day lead time, the $R^2$ contour line at 0.4 expands around (115°E, 10°N), while $R^2$ in the northern SCS (north of 15°N) declines compared with the 1-day lead time. The ConvLSTM model has lower $R^2$ values than PredRNN. For a 1-day lead time, $R^2$ in the central SCS is approximately 0.7. For a 2-day lead time, $R^2$ in the northwest SCS (near 113°E, 20°N) increases to 0.8. By the 3-day lead time, $R^2$ recovers to around 0.8 in the northern SCS (north of 15°N). However, in the southern SCS near (115°E, 12°N), $R^2$ drops below 0.4. For lead times of 4 to 7 days, $R^2$ in the southern SCS near (115°E, 12°N) further declines. Nevertheless, $R^2$ generally recovers to above 0.4 for lead times of 8 days or longer. With a 10-day lead time, ConvLSTM’s $R^2$ is significantly higher than that of PredRNN.

Furthermore, we integrate the SCS data to obtain spatially averaged metrics, as shown in Figure 7 and

Table 1. Overall metrics of forecasting models including ACFN, PredRNN, and ConvLSTM. The metrics include MAE (units: °C), $R^2$, RMSE (units: °C), and CC. Lead time spans from 1 to 10 days. The computational area is the SCS only.

Lead Time	ACFN				PredRNN				ConvLSTM
	MAE	${\mathit{R}}^{\mathbf{2}}$	RMSE	CC	MAE	${\mathit{R}}^{\mathbf{2}}$	RMSE	CC	MAE	${\mathit{R}}^{\mathbf{2}}$	RMSE	CC
1	0.215	0.972	0.288	0.987	0.601	0.776	0.724	0.949	0.940	0.614	1.176	0.905
2	0.309	0.945	0.404	0.974	0.643	0.753	0.779	0.937	0.864	0.661	1.087	0.906
3	0.372	0.924	0.479	0.964	0.675	0.735	0.821	0.928	0.726	0.714	0.902	0.906
4	0.427	0.903	0.543	0.956	0.699	0.721	0.851	0.920	0.822	0.595	0.993	0.905
5	0.469	0.884	0.594	0.948	0.719	0.708	0.876	0.914	0.943	0.470	1.112	0.904
6	0.503	0.868	0.635	0.940	0.736	0.697	0.898	0.908	0.922	0.510	1.100	0.893
7	0.532	0.853	0.670	0.933	0.751	0.687	0.917	0.903	0.808	0.634	0.986	0.880
8	0.557	0.838	0.702	0.926	0.764	0.678	0.934	0.899	0.706	0.725	0.877	0.880
9	0.581	0.824	0.731	0.920	0.776	0.669	0.950	0.895	0.656	0.762	0.826	0.884
10	0.604	0.810	0.759	0.915	0.789	0.660	0.967	0.891	0.643	0.770	0.813	0.887

. The ACFN model outperforms both PredRNN and ConvLSTM in terms of MAE and $R^2$. For a 1-day lead time, ACFN’s MAE is nearly 0.22 °C, while PredRNN’s MAE is approximately 0.60 °C, and ConvLSTM’s MAE exceeds 0.94 °C. In terms of $R^2$, the spatially averaged value of ACFN is roughly 0.97, while PredRNN and ConvLSTM exhibit significantly weaker $R^2$ performance. For a 2-day lead time, ConvLSTM’s MAE improves to about 0.86 °C, which remains higher than those of ACFN and PredRNN. The $R^2$ value for ConvLSTM on the 2-day lead time increases to 0.66. However, by the 4-day lead time, ConvLSTM’s $R^2$ falls below 0.6. On the 7-day lead time, ConvLSTM’s $R^2$ slightly exceeds 0.6, and for lead times of 8 days or longer, $R^2$ further improves.

3.4. In Situ Evaluation

We also evaluated the performance of the model using in situ observations. At buoy 4, the time series is shown in Figure 8. The OISST (ground truth) exhibits considerable errors, particularly during the 30-day period from August 15 to September 15, when in situ SSTs are significantly higher than OISST. On 15 September, a rapid cooling event, attributed to a typhoon, was observed. The phase of SST cooling in OISST slightly lags behind that in the in situ observations. Using the ACFN model for a 1-day lead time during this 30-day period, the predicted SST is generally cooler than the in situ observations. On 15 September, when typhoon-induced cooling occurred, the predicted SST did not decrease as rapidly as the in situ observations; however, it did show a cooling tendency on the following days. This indicates an at least 1-day delay in the predicted SST cooling relative to the in situ observations. As the lead time increases, the deviation between the predicted and in situ SSTs grows, with a more pronounced delay in capturing the SST cooling event. For PredRNN, the 1-day lead time predictions exhibit significantly colder SSTs compared with in situ observations, with part of this discrepancy attributed to the limitations in the ground truth. The model does predict SST cooling during the typhoon, but with a time lag similar to that observed in ACFN. Longer lead times also show delayed predictions of the SST cooling event. For ConvLSTM, predictions for lead times ranging from 1 to 6 days consistently show lower SSTs than those observed in situ. However, beyond a 7-day lead time, the model predictions become more consistent with the in situ data. Despite this, ConvLSTM fails to accurately predict the SST cooling event even with a 1-day lead time.

A bar chart showing buoy-based evaluation metrics is shown in Figure 9 and summarized in

Table 2. Metrics of forecasting models including ACFN, PredRNN, and ConvLSTM based on in situ buoy.

Lead Time	ACFN				PredRNN				ConvLSTM
	MAE	${\mathit{R}}^{\mathbf{2}}$	RMSE	CC	MAE	${\mathit{R}}^{\mathbf{2}}$	RMSE	CC	MAE	${\mathit{R}}^{\mathbf{2}}$	RMSE	CC
0 *	0.431	0.693	0.506	0.921	0.431	0.693	0.506	0.921	0.431	0.693	0.506	0.921
1	0.500	0.590	0.584	0.907	0.700	0.208	0.812	0.843	0.706	0.178	0.827	0.575
2	0.497	0.581	0.591	0.890	0.708	0.164	0.834	0.809	0.749	0.134	0.849	0.390
3	0.527	0.524	0.629	0.859	0.723	0.120	0.856	0.777	0.863	−0.267	1.027	0.431
4	0.570	0.450	0.677	0.829	0.734	0.094	0.868	0.754	1.021	−0.703	1.191	0.566
5	0.595	0.402	0.706	0.807	0.741	0.078	0.876	0.735	1.149	−1.064	1.311	0.661
6	0.616	0.372	0.723	0.790	0.748	0.065	0.882	0.717	1.089	−0.897	1.257	0.692
7	0.631	0.341	0.741	0.770	0.753	0.056	0.887	0.700	0.891	−0.336	1.055	0.671
8	0.652	0.301	0.763	0.742	0.756	0.050	0.890	0.681	0.737	0.053	0.888	0.660
9	0.667	0.259	0.785	0.715	0.753	0.055	0.887	0.666	0.692	0.097	0.867	0.630
10	0.685	0.219	0.806	0.692	0.749	0.065	0.882	0.655	0.698	0.025	0.901	0.610

* Lead time equals 0 means the results are provided by the ground truth (OISST).

. The baseline of OISST has an MAE of 0.431 °C and an $R^2$ of 0.693. For a 1-day lead time, ACFN’s MAE is slightly higher than OISST, at approximately 0.500 °C, and its $R^2$ is marginally lower at 0.590. ACFN’s performance slightly degrades with longer lead times, with MAE reaching 0.685 °C and $R^2$ decreasing to 0.219 with the 10-day lead time. For PredRNN, the MAE for a 1-day lead time is significantly larger at 0.700 °C, and it increases slightly to 0.741 °C for a 5-day lead time. The $R^2$ for PredRNN is significantly lower than that of OISST, at 0.208 for the 1-day lead time, and decreasing further to 0.078 for the 5-day lead time. ConvLSTM has a MAE of 0.706 °C for a 1-day lead time, with a corresponding $R^2$ of 0.178. Notably, its $R^2$ falls below 0.0 for lead times ranging from 3 to 7 days, indicating relatively poor performance. In particular, ConvLSTM outperforms PredRNN in terms of MAE for lead times of 8 days and longer.

4. Discussion

The present results confirm that ACFN outperforms both PredRNN and ConvLSTM. However, it is important to note that the input length of 10 days used for all three models may have slightly affected the performance of ConvLSTM and PredRNN, as suggested by Hao et al. [28]. This aspect should be considered when interpreting the results, as optimizing the input length could potentially enhance the performance of these models.

In comparison with our previous forecasting experiment in the Bohai Sea, the present forecasting experiment demonstrates a lower MAE. It is noteworthy that seasonal sea ice influences SST in the Bohai Sea but is absent in the SCS. However, the two experiments utilize different satellite products and define varying spatial resolutions in the models. Therefore, the predictability of SST in the Bohai Sea and its differences from the SCS require further investigation.

The spatial distribution of MAE and $R^2$ indicates that both MAE and $R^2$ are higher in the north than in the south. In this context, MAE represents absolute error, while $R^2$ characterizes the similarity of changes. Furthermore, the variation in SST is greater in the north compared with the south. This increased SST variation in the north results in a stronger signal, making its tendency more easily predictable by the forecasting model, which contributes to a higher $R^2$.

From a forecasting perspective, the predictability of natural processes varies across different spatial points, indicating that SST itself exhibits a spatial pattern in predictability. Assuming a perfect model, we can analyze the forecast skill to assess this predictability. For instance, if we define $R^2\ge$ 0.95 as a criterion for good predictability, we can determine the effective lead time for short-term SST forecasts (Figure 10). The results show that the northern SCS is more predictable than the central and southern SCS. Furthermore, when adopting $R^2\ge$ 0.95 and MAE ≤ 0.4 °C as criteria for good predictability, the effective lead time is shortened in the northern SCS. The pattern of effective lead time can also inform future observational strategies. To extend the effective lead time, it is advantageous to enhance in situ observations in regions where the effective lead time is currently short.

This study is limited to univariate forecasting, which may introduce some deviations from actual forecasting performance. The univariate approach does not account for external forcing factors that influence SST. Future improvements could involve incorporating additional input variables and extending the input length, which can improve forecasting performance by capturing a broader range of influencing factors. On the other hand, this study uses daily satellite SST data. In future research, hourly satellite SST can be employed. Tu and Hao [38] assessed the error associated with hourly satellite SST from Himawari-8. They found that the seasonal variation in SST error was not significant, while the diurnal variation in SST was estimated to be approximately 0.1 °C. Utilizing hourly SST has the potential to reduce forecasting errors attributed to diurnal variation.

Our in situ evaluation suggests that current satellite SST observations in the SCS are insufficient. From a spatial perspective, the horizontal resolution of the current satellite SST is 1/4° × 1/4°. This level of resolution is inadequate for resolving oceanic submesoscale processes. Consequently, errors in satellite SST may arise from the omission of subgrid dynamics when compared with in situ buoy observations. Although the variation in sub-mesoscale processes remains poorly understood, adopting high spatial resolution satellite SST is likely to improve the performance of deep learning models, though this requires further investigation. Therefore, it is essential to conduct additional in situ observation to obtain more realistic SST. Furthermore, from an operational perspective, it is permissible to perform error correction on forecast results in the context of operational use.

Our forecasting experiments end in the year 2022, and these models can be easily operated in 2023 and beyond. The forecast model was trained using the satellite SST data from 1982 to 2013. The empirical relationship between predicted SST and historical SSTs is primarily learned from this earlier period. It is anticipated that better forecasting results can be achieved through dynamical model training. For instance, retraining the model with data from 1982 to 2014 when predicting for 2015 could enhance performance. The advantages of implementing dynamic training warrant further exploration.

5. Conclusions

The present study was motivated by the implementation of relatively advanced deep learning models for short-term SST prediction in the SCS, where these deep learning models are trained using satellite SST data but can be further evaluated with in situ SST observations. The models employed in this study include ACFN, PredRNN, and ConvLSTM, with ACFN being our recently developed deep learning model.

This study quantifies the forecasting skills of three state-of-the-art deep learning models for regional SST prediction. ACFN demonstrates strong suitability for numerical weather prediction of SST in the SCS. Using satellite data as the ground truth, ACFN achieves an MAE of 0.215 °C and an $R^2$ of 0.972 for a 1-day lead time. For a 5-day lead time, the MAE rises to 0.469 °C, with the $R^2$ value decreases to 0.884. At the 10-day lead time, the longest period evaluated in this study, the MAE further increases to 0.604 °C, and the $R^2$ drops to 0.810. Furthermore, we incorporated in situ SST observations to evaluate the forecast results, providing a robust assessment of model performance. The model predictions that do not utilize in situ data achieve an MAE of 0.500 °C for 1-day lead time. Our deep learning model, ACFN, outperforms both PredRNN and ConvLSTM, suggesting its superiority in short-term SST forecast.

The significant bias between satellite and in situ SST data calls for future efforts in SST observation. The forecasting metrics presented in this study may also serve as a valuable reference for future numerical forecasting experiments in the SCS.