Significant Improvement in Short-Term Green-Tide Transport Predictions Using the XGBoost Model

Abstract

Accurately predicting the drift trajectory of green tides is crucial for assessing potential risks and implementing effective countermeasures. This paper proposes a short-term green-tide drift prediction method that combines green-tide patch characteristics, 1 h interval drift distances from GOCI-II images, and driving-factor data using the XGBoost machine learning model to enhance prediction accuracy. The results demonstrate that the proposed method outperforms the traditional OpenDrift model in short-term predictions. Specifically, at time intervals of 3, 5, and 7 h, the root mean square errors (RMSEs) of the OpenDrift model in the zonal direction are 1.81 km, 2.89 km, and 3.55 km, respectively, whereas the RMSEs of the proposed method are 0.80 km, 0.98 km, and 1.20 km, respectively; in the meridional direction, the RMSEs of the OpenDrift model are 1.77 km, 2.67 km, and 3.10 km, while the RMSEs for the proposed method are 0.82 km, 1.10 km, and 1.25 km, respectively. Furthermore, the proposed XGBoost method more-accurately tracks the actual positions of green-tide patches compared to the OpenDrift model. Specifically, at the 25 h interval, the proposed method continues to accurately predict patch positions, while the OpenDrift model exhibits significant deviations. This study demonstrates that the proposed method, by learning drift patterns from historical data, effectively predicts the short-term drift process of green tides. It provides valuable support for early warning systems, thereby helping to mitigate the ecological and economic impacts of green-tide disasters.

1. Introduction

With global climate change and increasing seawater eutrophication driven by human activities, the frequency of algal bloom events, such as green tides, is rising, particularly in the Yellow Sea of China [1,2,3]. Since 2007, large-scale green-tide outbreaks have occurred annually in the Yellow Sea, and from 2016 onwards, golden tides caused by Sargassum horneri have emerged as another recurring cross-regional harmful algal bloom, affecting both the Yellow Sea and the East China Sea [4,5,6]. These algal blooms have had severe negative impacts on coastal aquaculture, tourism, marine ecosystems, and maritime traffic. The frequent occurrence of green tides, especially along the coastal regions of Jiangsu and Shandong, has prompted local governments to allocate substantial resources annually in early monitoring, control, and removal efforts to mitigate their harmful impacts [7,8]. Currently, ship-based removal remains the primary method for clearing green tides [9]. Therefore, accurately predicting the scale, affected areas, and drift trajectories of green-tide disasters is crucial for decision-making in green-tide salvage operations [10].

Remote sensing technology is currently the primary method for monitoring green tides, owing to its characteristics of long-term, continuous, and large-scale monitoring [11,12,13]. Methods such as vegetation indices, algal indices, and machine learning applied to cloud-free imagery during bloom periods effectively detect green-tide patches [14,15,16]. These approaches facilitate the rapid evaluation of the spatial distribution and extent of green-tide outbreaks. In addition, models that link algal indices to green-tide biomass provide estimates of the magnitude of the green tide [17,18,19]. Meanwhile, remote sensing monitoring distributions are also used as reference values for initializing numerical simulations, recalibrating parameters, and validating and refining models [20,21,22]. However, remote sensing technology has inherent limitations. Issues such as cloud cover and the temporal gaps between satellite observations can hinder continuous monitoring [23]. Despite efforts to integrate multi-source satellite imagery for more-comprehensive observations, ensuring sustained, uninterrupted monitoring over extended periods remains a significant challenge [11,16,24].

In recent years, numerical drift models have emerged as a crucial method for monitoring green tides to improve prediction accuracy [23]. By parameterizing the drift process of green-tide patches and simulating their movement with particle tracking models, this approach addresses the spatiotemporal limitations of remote sensing, and enables more comprehensive monitoring of dynamic green-tide changes. Currently, numerical simulations are primarily used to trace the origins of algal blooms, predict the affected areas, and evaluate the biological and physical factors influencing the scale of these blooms [25,26,27]. However, due to uncertainties in driving-force products and inappropriate parameter setting, numerical simulations inevitably contain errors that tend to accumulate over time, reducing the reliability of drift predictions and potentially leading to ineffective response actions. For instance, windage, a parameter that characterizes the direct effects of wind, and may also include the combined influences of wind forcing, Stokes drift, and additional surface shear, which is often set as a constant, despite its dependence on the driving-force data utilized [28]. To mitigate long-term accumulative simulation errors, Zhou et al. (2021) proposed dividing the overall simulation into multiple short-term segments [23]. By reducing the duration of each segment and utilizing remote sensing observations to reinitialize the starting location, error accumulation can be minimized. Therefore, improving the accuracy of short-term predictions for green-tide patches has the potential to precisely simulate the drift process over the entire duration of the green-tide outbreak.

Machine learning offers a promising alternative for short-term drift prediction by directly learning from historical green-tide data. For instance, Wang et al. (2023) successfully generated daily average green-tide distribution maps by combining long-term Moderate-resolution Imaging Spectroradiometer (MODIS) green-tide interpretation maps with a convolutional long short-term memory (ConvLSTM) model [29]. When applying machine learning to enhance green-tide drift prediction accuracy, two strategies can be considered: (1) The first strategy is to develop a relationship model between patch attributes, driving factors, and windage, with the goal of adjusting windage to more closely approximate its true value, thereby indirectly improving patch drift simulations. (2) The second strategy aims to directly model the correlations between driving factors, patch attributes, and drift distance to enhance the precision of drift predictions. A case study of the first strategy was conducted by Yu et al. (2023), who introduced a deep learning method to parameterize windage by minimizing the discrepancies between simulated and observed drifter trajectories [30]. While this method reduced simulation errors, its overall impact on accuracy was limited, providing only a slight improvement. This finding suggests that the first strategy may be less effective in significantly enhancing simulation accuracy. The limited performance may be attributed to the discrepancy between the windage derived by minimizing the simulated displacement across all samples and the true windage values for individual samples. Additionally, the deviation between the estimated and actual windage values, caused by the uncertainty in the zonal and meridional components of wind speed and current velocity, as well as the unreasonable drift-angle setting, also contribute to the limited improvement [20].

The Extreme Gradient Boosting (XGBoost) model effectively handles noise and outliers, along with mitigates overfitting, and it is widely applied in complex remote sensing classification and parameter estimation [31]. Based on the previous analysis, this study aims to enhance the short-term prediction accuracy by directly establishing the relationship between driving factors, patch attributes, and drift distance using the XGBoost model. Specifically, the prediction accuracy of the traditional numerical simulation method (choosing OpenDrift as an example) for green-tide drift at various time intervals is evaluated, with the green tides extracted from the Geostationary Ocean Color Imager-II (GOCI-II) satellite used for validation. Then, the proposed method is trained and tested using green-tide drift data at a 1 h interval and then applied to other time intervals. The feasibility of the proposed method is assessed by comparing its results with those from the OpenDrift model. This study explores the potential of machine learning methods to enhance the predictive accuracy of short-term green-tide drift.

2. Materials and Methods

2.1. Study Area

Green tides in the Yellow Sea primarily occur between May and August each year. The study area covers the main affected region, spanning from 32.5°N to 37°N latitude and 119°E to 124°E longitude (Figure 1). This area is a typical shallow sea with an average depth of approximately 44 m and a gently sloping seabed, while the northern part is characterized by distinct radial sand ridges [32]. The southern Yellow Sea is heavily influenced by the East Asian monsoon, leading to complex and variable ocean currents, tides, and longshore currents, which together create a unique hydrodynamic environment [33]. Along the coast of Jiangsu, extensive coastal aquaculture ponds and seaweed farming, particularly Porphyra cultivation, are important sources of local income [2]. Over 30 rivers flow into the study area, with the Sheyang River, Xinyanggang River, and Guan River being the largest contributors [1]. The substantial land-based nutrient input, combined with complex hydrodynamic conditions, has resulted in high levels of eutrophication in these waters [34]. In recent years, rising global sea temperatures have further exacerbated this trend, creating ideal conditions for algal bloom outbreaks [3].

2.2. Data and Processing

The data used in this study primarily consist of satellite remote sensing images and driving-force data, specifically ocean-current velocities and sea-surface wind speeds, collected during a green-tide outbreak.

2.2.1. Satellite Data

The satellite imagery is sourced from the GOCI-II geostationary satellite, that was launched in 2020. It provides a spatial resolution of 250 m and a temporal resolution of one hour. GOCI-II features 12 spectral bands, ranging from ultraviolet to near-infrared, and is capable of acquiring up to 10 images per day between 7:15 AM and 4:15 PM local time. The high temporal resolution of GOCI-II makes it an ideal platform for monitoring the dynamics and progression of green-tide events [35]. The available GOCI-II imaging data begin in September 2020, and the scale and duration of the 2022 green-tide outbreak were smaller than in previous years [36]. Therefore, this study utilizes GOCI-II images from the green-tide outbreaks in 2021, 2023, and 2024, with the 2021 and 2023 events representing the most prominent algal bloom occurrences [14]. The specific acquisition times of the satellite images are detailed in

Table 1. Details of the dates of GOCI-II images used in this study.

Year	Date
2021	4 June, 5 June, 6 June, 7 June, 19 June, 20 June, 22 June, 23 June, 1 July, 9 July, 10 July
2023	3 June, 6 June, 8 June, 9 June, 11 June, 12 June, 13 June, 14 June, 15 June, 22 June, 24 June, 27 June, 5 July, 9 July, 10 July
2024	3 June, 6 June, 7 June, 8 June, 9 June, 10 June, 13 June, 15 June, 18 June, 23 June, 25 June, 26 June, 27 June, 30 June, 3 July

. The Rayleigh-corrected reflectance data were filtered for low-cloud or cloud-free conditions and downloaded from the website: https://www.nosc.go.kr/eng/program/actionGociDownload.do?boardSelectPage=1 (accessed on 20 April 2025).

The processing of the downloaded images includes image preprocessing, green-tide interpretation, extraction of patch features, and the application of the Maximum Cross-Correlation (MCC) technique to determine the drift distance of green-tide patches. Initially, preprocessing involves geographic correlation based on a geographic lookup table and the application of land and cloud masks [37]. The alternative floating algae index (AFAI) and spectral angle index (SAI) are then calculated, and thresholds identified through visual interpretation are applied to produce binary images representing the spatial distribution of green tides [38]. Morphological processing is subsequently applied using OpenCV to extract patch feature information, such as the centroid, maximum enclosing rectangle, minimum enclosing rectangle, perimeter, and area of the green-tide patches. This information is then utilized in the MCC method to calculate the drift distance and for the construction of subsequent machine learning models. Using the SAI imagery and the MCC method, patch feature data are applied to set the MCC parameters, allowing for the calculation of the drift distance of green-tide patches. Finally, the newly constructed green-tide contours are compared to the actual distribution, with visual interpretation used to exclude patches that deviate from their true position. The calculation equations and detailed procedures for the spectral indices and MCC method are provided in the work of Ji et al. (2024) [39].

Four time intervals (1, 3, 5, and 7 h) were selected to quantitatively assess errors in short-term green-tide drift simulations under tidal influence. The 1-, 3-, and 7 h intervals correspond to the high-resolution image pair intervals, the time between MODIS morning and afternoon satellite passes, and the maximum interval between GOCI observations, respectively [39]. Increasing the 5 h interval extends the time interval by the same duration. Since green-tide patches undergo significant changes over periods longer than one day and are influenced by weather conditions, the number of matching patches becomes limited. Therefore, only a few patches with a 25 h interval were used to qualitatively evaluate the predicted net transport simulation errors. Evaluations over a 25 h period were not considered.

2.2.2. Driving-Factor Data

Ocean-current velocities and sea-surface wind speeds are the primary forces driving the drift of green tides at the ocean surface and provide the key input for drift models. For ocean-current velocities, this study utilizes the global analysis and forecast product (GLOBAL_ANALYSISFORECAST_PHY_001_024) from the Copernicus Marine Environment Monitoring Service (CMEMS), accessible at https://marine.copernicus.eu/ (accessed on 20 April 2025). This dataset has a spatial resolution of 1/12° and a temporal resolution of 1 h. The sea-surface current products include meridional and zonal Stokes drift velocities, geostrophic plus Ekman velocities, tidal currents, and a combination of these velocity components. The meridional refers to the north–south orientation, while the zonal refers to the east–west orientation.

Wind speeds are sourced from the European Centre for Medium-Range Weather Forecasts (ECMWF) ERA5, the fifth-generation global climate and weather reanalysis product, accessible at https://cds.climate.copernicus.eu/ (accessed on 20 April 2025). This dataset offers zonal and meridional wind speeds at 10 m above sea level, with a spatial resolution of 0.25° and a temporal resolution of 1 h.

2.3. Method

2.3.1. Numerical Simulation Model

In this study, the OpenDrift open-source framework (https://github.com/OpenDrift/opendrift (accessed on 20 April 2025)) was employed as a benchmark to evaluate the performance of the proposed method, as it is extensively utilized for simulating oceanic object drift [40]. In this framework, green-tide patches are represented as particles, and their drift is simulated using a Lagrangian particle-tracking approach. The horizontal movement of the patches is primarily influenced by ocean currents, wave-induced Stokes drift, and wind forces. The “wind_drift_factor” parameter in the model represents the influence of wind, typically set at 2% of the surface wind speed. When accounting for the Stokes drift, this factor is adjusted to 3.2%. According to Dagestad et al. (2018), the “wind_drift_factor” originates from the OpenOil model, where it is applied to the wind speed and added to the surface current field to better approximate actual drift [40]. This parameter does not represent windage but rather compensates for the significant shear effects in the upper centimeters of the ocean surface, which are not captured by ocean models (https://github.com/Opendrift/Opendrift/issues/362 (accessed on 20 April 2025)). For simplicity, this study does not distinguish between the “wind_drift_factor” and windage.

As parameter settings vary depending on driving-force data and the simulated region, recalibrating model parameters before conducting numerical simulations is essential. This process typically involves aligning the observed simulated green-tide distributions or minimizing the displacement between simulated and actual buoy trajectories. Windage is a key parameter that needs to be recalibrated before simulating green-tide drift with the OpenDrift model, and is recalibrated based on green-tide drift observed at 7 h intervals. However, the simulation results showed only a marginal improvement of approximately 0.2 km after recalibration. This may be because the windage derived by minimizing the simulated displacement of the overall patches does not fully represent the actual windage of each individual patch, suggesting that it is challenging to accurately reproduce the drift process of green-tide patches by calibrating windage solely based on minimizing the overall displacement. Furthermore, as the recalibration process is time-consuming and requires parameter adjustments before each simulation, this study does not involve the recalibration of the OpenDrift model to enhance efficiency. The observed green-tide patch distribution is used as the initial position for the model, with total ocean currents (including Stokes drift) and sea-surface wind speeds serving as the driving forces. The windage parameter is set to 0.02, and by defining the start and end times, the green-tide drift process is simulated at various time intervals.

2.3.2. Machine Learning Model

Using a variable value of windage for each individual green-tide patch would be an ideal optimization method for improving green-tide drift predictions, owing to its stronger physical relevance. However, since obtaining accurate windage is not feasible, a new approach for short-term green-tide drift prediction is proposed. This approach directly models the relationship between driving factors, patch attributes, and drift distance using the XGBoost model. The flowchart of the proposed method is illustrated in Figure 2. The attributes (location, area, and rotation) of green-tide patches extracted at 1 h intervals and the corresponding driving-force data are used as dependent variables, while the drift distance of green-tide patches served as the independent variable to construct the XGBoost model. A total of 12,345 samples are used, with 7000 samples randomly selected for model training, 3000 for testing, and 2345 for validation. After model development and validation, an iterative process is conducted to predict the drift distances of green-tide patches at other time intervals. These predictions are then compared with observed drift distances to assess the robustness and accuracy of the model.

2.4. Statistical Analysis

To evaluate the accuracy of the model performance, bias (Equation (1)), the root mean square error (RMSE) (Equation (2)), and the coefficient of determination (R²) (Equation (3)) are employed as key performance metrics. These metrics assess whether the model is overestimating or underestimating, how well the independent variables explain the variability in the dependent variable, and the overall precision of the predictions.

$$\mathrm{Bias}={\displaystyle \frac{1}{\mathrm{n}}}{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{n}}{(\mathrm{d}}_{\mathrm{pre}}^{\mathrm{i}}{-\mathrm{d}}_{\mathrm{t}}^{\mathrm{i}})}$$

(1)

$$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{1}{\mathrm{n}}}{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{n}}{{(\mathrm{d}}_{\mathrm{pre}}^{\mathrm{i}}{-\mathrm{d}}_{\mathrm{t}}^{\mathrm{i}})}^2}}$$

(2)

$${\mathrm{R}}^2=1-{\displaystyle \frac{{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{n}}{{(\mathrm{d}}_{\mathrm{pre}}^{\mathrm{i}}{-\mathrm{d}}_{\mathrm{t}}^{\mathrm{i}})}^2}}{{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{n}}{{(\mathrm{d}}_{\mathrm{t}}^{\mathrm{i}}-\overline{{\mathrm{d}}_{\mathrm{t}}^{\mathrm{i}}})}^2}}}$$

(3)

where ${\mathrm{d}}_{\mathrm{pre}}$, ${\mathrm{d}}_{\mathrm{t}}$, and $\overline{\mathrm{d}}$ represent the predicted distance, reference distance, and the average of reference distance, respectively, and $\mathrm{n}$ represents the sample size.

3. Results

3.1. Analysis of the Performance of the OpenDrift Method in Simulating Short-Term Green-Tide Drift

In this study, the drift process of green-tide patches is simulated using the OpenDrift model, and its performance is evaluated by comparing the simulated with the observed drift distances of the patches, derived from satellite imagery at multiple time intervals. Figure 3 illustrates the deviations of the simulated from the observed drift distances of green-tide patches at 1 h, 3 h, 5 h, and 7 h intervals. The results indicate that as the time interval increases, the deviation gradually increases. This trend is primarily attributed to the accumulation of errors in the model over time. Additionally, as the simulation duration extends, the simulated drift distances of the green-tide patches tend to exceed the observed drift distances, with the deviation being more pronounced in the meridional direction than in the zonal direction (Figure 3g,h). This is consistent with the observed phenomenon of green-tide patches predominantly drifting along the meridional direction [39]. However, the significant overestimation in the zonal direction may be due to the model neglecting the slight rightward deflection caused by wind acting at an angle on the green-tide patches. As a result, the zonal directional force applied in the model is too high.

The comparison between the simulated and the observed green-tide-patch drift distances from the satellite imagery is presented in Figure 4. The density of points is also shown, with yellow indicating areas of higher concentration and blue representing areas with fewer scattered points. The results show that the model performs better in the meridional direction compared to the zonal direction. In the zonal direction, the RMSE values for the 1 h, 3 h, 5 h, and 7 h time intervals are 0.64 km, 1.81 km, 2.89 km, and 3.55 km, respectively; in the meridional direction, the RMSE values for the same time intervals are 0.64 km, 1.77 km, 2.67 km, and 3.10 km, respectively. This indicates that as the model runs over time, the discrepancy between the simulated and observed positions of the green-tide patches increases, with greater deviation in the zonal direction than in the meridional direction. Additionally, the error growth between the 5 h and 7 h intervals is smaller compared to the increase between the 3 h and 5 h intervals. This may be due to the tidal cycle, which spans approximately 12.25 h. The 7 h interval coincides with a reversal phase in the green-tide movement, leading to a shorter actual drift distance compared to the 5 h interval, which partially mitigates the accumulation of errors.

Furthermore, by varying the windage values from 0% to 5% with a 0.1% increment and comparing the results with the observed drift distance at different time intervals, the optimal windage value for each patch is determined to assess its impact on the OpenDrift model predictions. Figure 5 shows the distribution of these optimal windage values. The results indicate that at various time intervals, the optimal windage values are predominantly 0% or 5%. This suggests that lower driving-force values result in simulated drift distances that are shorter than the observed distances, while higher values may cause overestimation. Windage is adjusted to minimize these errors. Additionally, the variation in windage values implies that different patches require distinct windage values owing to their inherent characteristics and errors in driving-force data. In the OpenDrift model, windage is typically set to a fixed value, leading to significant discrepancies between predicted and observed patch positions. Since the windage values are small, the improvement in prediction accuracy through optimization is limited over short time intervals, which are primarily influenced by ocean currents [39].

The comparison between the errors caused by ocean-current data (calculated as the difference between patch drift speed and average ocean-current velocities, multiplied by time) and the actual simulation errors of the OpenDrift model is shown in

Table 2. Regression model between simulated distance error and the approximate distance error due to ocean currents.

Time Interval (h)	Model	R2
1	$e_s^u=0.94e_o^u$+ 0.07 $e_s^v=0.73e_o^v$+ 0.12 $e_s=0.85e_o$+ 0.14	0.62 0.50 0.59
3	$e_s^u=0.78e_o^u$+ 0.32 $e_s^v=0.85e_o^v$+ 0.12 $e_s=0.85e_o$+ 0.27	0.57 0.72 0.68
5	$e_s^u=0.76e_o^u$+ 0.40 $e_s^v=0.81e_o^v$+ 0.21 $e_s=0.86e_o$+ 0.26	0.54 0.60 0.60
7	$e_s^u=0.78e_o^u$+ 0.43 $e_s^v=0.34e_o^v$+ 1.52 $e_s=0.65e_o$+ 0.40	0.64 0.11 0.43

Here, $e_s^u$, $e_s^v$, and $e_s$ represent the prediction errors of green-tide patches in the zonal, meridional, and total directions, respectively. Similarly, $e_o^u$, $e_o^v$, and $e_o$ denote the prediction errors attributed to ocean-current inaccuracies in the corresponding directions.

to assess the impact of inherent errors in the ocean-current product on simulation results. The results show that the R² values are all above 0.5 except for the 7 h interval, which indicates that the drift error of green-tide patches is primarily driven by inherent errors in ocean-current velocities over short time intervals. At the 7 h interval, the correlation in the zonal direction remains relatively high, while that in the meridional direction is low. This may be due to the significant influence of wind at the 7 h interval, which becomes more pronounced over longer time periods. This analysis suggests that errors in the driving-force data, along with uncertainties in windage parameter setting, lead to unavoidable discrepancies between the predicted and observed locations of green-tide patches in the OpenDrift model. As simulation time increases, these errors gradually accumulate. This implies that, before using the OpenDrift model for long-term predictions, further optimization of the driving data and recalibration of the windage parameters are essential to improve prediction accuracy.

3.2. Analysis of the Performance of the Proposed Method in Simulating Short-Term Green-Tide Drift

A comparison between the predicted drift distances of green-tide patches obtained by the proposed method and the observed drift distances is shown in Figure 6. In the training dataset (Figure 6a,b), the R² values for both the zonal and meridional directions are 0.98, with an overall bias of 0.00 km and RMSE values of 0.21 km and 0.23 km, respectively; for the testing dataset (Figure 6c,d), the R² values are 0.94, with biases of 0.00 km and −0.01 km, and RMSE values of 0.31 km and 0.35 km, respectively; the validation dataset (Figure 6e,f) also exhibits R² values of 0.94, with biases of −0.01 km and RMSE values of 0.30 km and 0.34 km, respectively. The results across the different datasets are relatively consistent. When compared to the results from OpenDrift model (Figure 4a,b), the RMSE values are lower, with the scatter points generally being closer to the 1:1 line, indicating that the proposed method demonstrates both accuracy and robustness.

To evaluate the performance of the proposed method at 3, 5, and 7 h intervals, the drift distances for these time periods are obtained through iterative simulations. Figure 7 illustrates the comparison between the model’s predicted and observed drift distances. These results indicate that the proposed method performs well across all three time intervals, with RMSE values of 0.80 km, 0.98 km, and 1.20 km in the zonal direction. In the meridional direction, the RMSE values are slightly higher at 0.82 km, 1.10 km, and 1.25 km, indicating a marginally lower predictive performance compared to the zonal direction. When compared to the predictions from the OpenDrift model (Figure 4c–h), the proposed method shows significant improvements in RMSE values, overall bias, and R² values, with biases remaining within 0.05 km and R² values exceeding 0.9. Notably, there is no evidence of significant accumulative errors. These findings suggest that the proposed method effectively simulates the drift process of green tides and is well-suited for short-term predictions of green-tide drift.

3.3. Visual Comparison of the Performance of the Proposed Method and OpenDrift Model

The comparison of the performances of the proposed method and OpenDrift model by visualizing the actual versus predicted drift of several patches at 1, 3, 5, and 7 h is presented in Figure 8. The analysis focuses on green-tide patches from 19 June, 22 June, and 23 June 2021, as these dates provide patches with similar distributions and characteristics, facilitating a more effective comparison of the models. As is shown in Figure 8, it can be observed that at the 1 h interval, the differences between the OpenDrift model and proposed method are minimal in most areas, with the predicted locations of both models closely aligning with the observed positions of the green-tide patches. However, as the time intervals increase, the results from the OpenDrift model gradually deviate from the observed patch positions, with deviation errors increasing over time. While the proposed method also exhibits some predictive offset as the time interval grows, its predicted positions remain largely consistent with the observed patch locations, showing smaller deviation errors compared to the OpenDrift model. This indicates that the proposed method is more effective at predicting the positions of green-tide patches over short time intervals. Additionally, Figure 8 reveals that, within the same time interval, the prediction offsets of the OpenDrift model vary across different regions. This suggests that discrepancies in the driving-force data and patch characteristics require different windage parameter settings for different areas.

The net transport predictions between the two models over a 25 h interval are also compared. Due to morphological changes in green-tide patches during drift and the influence of clouds causing gaps in satellite imagery, only a few patches can still be identified after this time period. Figure 9 shows the green-tide-drift tracking results that were searched over 3 years and can still be identified as the same patch after a 25 h interval. As is shown in Figure 9, it is clear that the OpenDrift model struggles to accurately predict the observed positions of the patches. In contrast, the proposed method consistently performs well in predicting the observed positions of the green-tide patches (illustrated in areas 9(a1–a4)), with deviations smaller than those of the OpenDrift model. Even in regions where the predictions from the proposed method exhibit a larger bias, the results remain comparable to those of the OpenDrift model (as seen in areas 9(c1,c3)). This indicates that the proposed method is effective in accurately predicting short-term green-tide-patch drift.

4. Discussion

This study explores the feasibility of using a machine learning method (XGBoost) for short-term green-tide drift prediction and compares this with a traditional numerical simulation model (OpenDrift). Due to inherent errors in the driving-force data and the challenges of numerical models in accurately simulating the effects of ocean currents, wind, and waves on drifting objects, along with the biological characteristics of green-tide patches during drift, there are discrepancies between the simulated and actual drift positions in the numerical models [41]. Additionally, these errors tend to accumulate over time. In contrast, the proposed method establishes the relationship between the attributes of green-tide patches, driving-force data, and observed drift distances, optimizing the simulated drift distance to more accurately replicate the drift process of green tides. The results show that across 1, 3, 5, 7, and 25 h intervals, the proposed method effectively tracks the observed green-tide drift locations, with the accuracy surpassing that of the OpenDrift model. This indicates that the proposed method is more adept at simulating short-term green-tide drift processes. However, the proposed method does have some limitations, particularly in regions with fewer patch location samples, where prediction errors tend to increase. For example, as is shown in Figure 10, some areas in Figure 9 (e.g., 9c1, 9c3, and 9c4) exhibit larger prediction deviations, primarily due to the lower number of training samples in these regions, which weakens the model’s generalization capability. Nevertheless, in areas with a larger sample size, the proposed method maintains a high prediction accuracy, even over longer time intervals, such as 25 h.

Additionally, this study utilizes observed drift distance data at 1 h intervals for model training, without exploring data from other time intervals. Given that the drift distance of green tides within a 1 h period is relatively short and may contain some errors in MCC extraction, future research should consider using data with 3 h intervals for training, to enhance the model’s predictive accuracy [39]. An analysis of the challenges in identifying green-tide patches at different time intervals on 23 June 2023 (Figure 11 and

Table 3. Statistics related to matched patches by the MCC method at various time intervals.

Time Interval (h)	Sample Size (N)	N (k > 0.50)	N (S > 10 km2)	Savg (S > 10 km2)	Smax (km2)	k (Smax)
1	152	152	55	48.88	352.31	0.53
2	134	133	53	43.99	189.48	0.57
3	133	129	54	41.24	157.98	0.62
4	116	109	50	41.18	157.98	0.57
5	98	90	39	39.16	142.43	0.47
6	60	50	32	38.15	142.43	0.43
7	57	52	27	37.28	136.08	0.55
8	57	52	28	36.71	136.08	0.54
9	52	51	23	32.45	136.08	0.51

N represents the sample size, S represents the patch area, k represents the similarity of the MCC method, Max represents the maximum, and avg represents the average.

) revealed that a 3 h interval not only ensures a sufficient number of green-tide patches but also maintains higher patch similarity and more substantial drift distances. Consequently, the relative error in MCC extraction is reduced. Furthermore, the 3 h interval coincides with the morning and afternoon observation windows of the MODIS satellite. This longer observation period enables MODIS to provide a greater number of green-tide patch samples, which would further support the model’s training and optimization.

Moreover, this study has developed a unified model for both zonal and meridional directions. However, due to angular deviations in the influence of wind on drifting objects and variations in the accuracy of driving forces in different directions, it may be necessary to consider the drift parameters for each direction separately [20]. Therefore, future research could explore the construction of distinct predictive models for zonal and meridional directions to further improve the accuracy of green-tide drift predictions. Finally, this study did not incorporate the growth and decay of green-tide patches during the drift process. In reality, green-tide patches undergo morphological changes as they drift, with shifts in both biomass and distribution over time [23]. Future studies could include the modeling of these growth and decay processes to better capture the biological dynamics of green tides and enhance the predictive capabilities of the model.

5. Conclusions

The recurring green tide in the Yellow Sea, a significant ecological threat, warrants accurate prediction of patch-movement trajectories and landing sites for effective disaster management and prevention. This study introduces an accurate tracking method for green-tide patches using the XGBoost model. By integrating the characteristic features of green-tide patches with driving-force data, the model establishes a direct relationship with patch drift distances at 1 h intervals. This approach minimizes prediction errors caused by inherent inaccuracies in driving-force data and numerical model parameter settings, enabling a more precise simulation of green-tide-patch drift over short time periods. This study assesses the prediction accuracy of the proposed method using satellite-observed drift distances and compares its performance to the traditional numerical model, OpenDrift, demonstrating the superior efficacy of the proposed method.

The results indicate that, at the 1 h interval, the OpenDrift model achieves RMSE values of 0.64 km in both the zonal and meridional directions. In contrast, the proposed method outperforms OpenDrift, achieving an average RMSE of 0.22 km in both directions on the training dataset, and an average RMSE value of 0.33 km on the testing and validation datasets. At the 3, 5, and 7 h intervals, the RMSE values for the OpenDrift model in the zonal direction are 1.81 km, 2.89 km, and 3.55 km, respectively. In comparison, the proposed method reports RMSE values of 0.80 km, 0.98 km, and 1.20 km. In the meridional direction, the RMSE values of the OpenDrift model are 1.77 km, 2.67 km, and 3.10 km, while the proposed method produces RMSE values of 0.82 km, 1.10 km, and 1.25 km. Moreover, the proposed method demonstrates higher R² values and a lower overall bias compared to the OpenDrift model across all time intervals. Visual tracking of the drift process of several real green-tide patches further validates the superior performance of the proposed method. Overall, these findings confirm that the proposed method is more effective in accurately simulating the drift process of green-tide patches over short time periods compared to the OpenDrift model. This study demonstrates that machine learning methods can effectively replicate the drift process of green-tide patches over 25 h by learning from drift data collected at 1 h intervals. This approach not only provides valuable insights for early warning systems for green-tide disasters but also offers convenience for further research on the prediction of drifting maritime objects.