A record-breaking agglomeration of Sargassum was packed along the northern Jeju coast in Korea in 2021, and laborers suffered from removing them from the beach. If remote sensing can be used to detect the locations at which Sargassum accumulated in a timely and accurate manner, we could remove them before their arrival and reduce the damage caused by Sargassum. This study aims to detect Sargassum distribution on the coast of Jeju Island using the Geostationary KOMPSAT 2B (GK2B) Geostationary Ocean Color Imager-II (GOCI-II) imagery that was launched in February 2020, with measurements available since October 2020. For this, we used GOCI-II imagery during the first 6 months and machine learning models including Fine Tree, a Fine Gaussian support vector machine (SVM), and Gentle adaptive boosting (GentleBoost). We trained the models with the GOCI-II Rayleigh-corrected reflectance (R_hoC) image and a ground truth map extracted from high-resolution images as input and output, respectively. Qualitative and quantitative assessments were carried out using the three machine learning models and traditional methods such as Sargassum indexes. We found that GentleBoost showed a lower false positive (6.2%) and a high F-measure level (0.82), and a more appropriate Sargassum distribution compared to other methods. The application of the machine learning model to GOCI-II images in various atmospheric conditions is therefore considered successful for mapping Sargassum extent quickly, enabling reduction of laborers’ efforts to remove them.

Red tides caused by Margalefidinium polykrikoides occur continuously along the southern coast of Korea, where there are many aquaculture cages, and therefore, prompt monitoring of bloom water is required to prevent considerable damage. Satellite-based ocean-color sensors are widely used for detecting red tide blooms, but their low spatial resolution restricts coastal observations. Contrarily, terrestrial sensors with a high spatial resolution are good candidate sensors, despite the lack of spectral resolution and bands for red tide detection. In this study, we developed a U-Net deep learning model for detecting M. polykrikoides blooms along the southern coast of Korea from PlanetScope imagery with a high spatial resolution of 3 m. The U-Net model was trained with four different datasets that were constructed with randomly or non-randomly chosen patches consisting of different ratios of red tide and non-red tide pixels. The qualitative and quantitative assessments of the conventional red tide index (RTI) and four U-Net models suggest that the U-Net model, which was trained with a dataset of non-randomly chosen patches including non-red tide patches, outperformed RTI in terms of sensitivity, precision, and F-measure level, accounting for an increase of 19.84%, 44.84%, and 28.52%, respectively. The M. polykrikoides map derived from U-Net provides the most reasonable red tide patterns in all water areas. Combining high spatial resolution images and deep learning approaches represents a good solution for the monitoring of red tides over coastal regions.

The spatial and temporal resolutions of contemporary microwave-based sea surface salinity (SSS) measurements are insufficient. Thus, we developed a gap-free gridded daily SSS product with higher spatial and temporal resolutions, which can provide information on short-term variability in the East China Sea (ECS), such as the front changes by Changjiang diluted water (CDW). Specifically, we conducted gap-filling for daily SSS products based on the Geostationary Ocean Color Imager (GOCI) with a spatial resolution of 1 km (0.01°), using a machine learning approach during the summer seasons from 2015 to 2019. The comparison of the Soil Moisture Active Passive (SMAP), Copernicus Marine Environment Monitoring Service (CMEMS), and Hybrid Coordinate Ocean Model (HYCOM) SSS products with the GOCI-derived SSS over the entire SSS range showed that the SMAP SSS was highly consistent, whereas the HYCOM SSS was the least consistent. In the < 31 psu range, the SMAP SSS was still the most consistent with the GOCI-derived SSS (R²=0.46; root mean squared error: RMSE = 2.41 psu); in the > 31 psu range, the CMEMS and HYCOM SSS products showed similar levels of agreement with that of the SMAP SSS. We trained and tested three machine learning models – the fine trees, boosted trees, and bagged trees models – using the daily GOCI-derived SSS as output, including the three SSS products, environmental variables, and geographical data. We combined the three SSS products to construct input datasets for machine learning. Using the test dataset, the bagged trees model showed the best results (mean R²=0.98 and RMSE = 1.31 psu), and the models that used the SMAP SSS as input had the highest level. For the dataset in the > 31 psu range, all the models exhibited similarly reasonable performances (RMSE = 1.25–1.35 psu). The comparison with in situ SSS data, time series analysis, and the spatial SSS distribution derived from models showed that all the models had proper CDW distributions with reasonable RMSE levels (0.91–1.56 psu). In addition, the CDW front derived from the model gap-free daily SSS product clearly demonstrated the daily oceanic mechanism during the summer season in the ECS at a detailed spatial scale. Notably, the CDW front in the zonal direction, as captured by the Ieodo Ocean Research Station (I-ORS), moved approximately 3.04 km d⁻¹ in 2016, which is very fast compared with the cases in other years. Our model yielded a gap-free gridded daily SSS product with reasonable accuracy and enabled the successful recognition of daily SSS fronts at the 1 km level, which was previously not possible with ocean color data. Such successful application of machine learning models can further provide useful information on the long-term variation of daily SSS in the ECS. The gridded gap-free SSS dataset at 0.01°×0.01° spatial resolution is freely available at https://doi.org/10.22808/DATA-2023-2 (Shin et al., 2023).

Since 1995, Margalefidinium polykrikoides blooms have occurred frequently in the waters around the Korean peninsula. In the South Sea of Korea (SSK), large-scale M. polykrikoides blooms form offshore and are often transported to the coast, where they gradually accumulate. The objective of this study was to investigate the synergistic effect of multi-sensor data for identifying M. polykrikoides blooms in the SSK from July 2018 to August 2018. We found that the Spectral Shape values calculated from in situ spectra and M. polykrikoides cell abundances in the SSK were highly correlated. Comparing red tide spectra from near-coincident multi-sensor data, remote-sensing reflectance (R_rs) spectra were similar to the spectra of in situ measurements from blue to green wavelengths. R_rs true-color composite images and Spectral Shape images of each sensor showed a clear pattern of M. polykrikoides patches, although there were some limitations for detecting red tide patches in coastal areas. We confirmed the complementarity of red tide data extracted from each sensor using an integrated red tide map. Statistical assessment showed that the sensitivity of red tide detection increased when multi-sensor data were used rather than single-sensor data. These results provide useful information for the application of multi-sensor for red tide detection.

Optical remote sensing using unmanned aerial vehicles (UAVs) is proposed to monitor changes in marine environments effectively. Optical measurements were performed using a UAV multispectral camera (RedEdge, five spectral wavelengths of 475, 560, 668, 717, and 842 nm) with high spatial (5 cm) and temporal (1 s) resolutions to monitor the rapidly changing suspended sediment concentration (SSC) in the Saemangeum coastal area on the western coast of Korea. To develop the SSC algorithm, optical field, and water sample measurements were obtained from outside (11 stations) and inside (three stations) regions separated by a seawall, accounting for 100 measurements from 2018 to 2020. Accordingly, the remote sensing reflectance (R_rs) was estimated at each sampling station and used to develop the SSC algorithm based on multiple linear regression. The algorithm reasonably estimated the SSC with an R² and root mean square error of 0.83 and 4.27 (mg L⁻¹), respectively. Continuous individual UAV measurements over the coastal area of Saemangeum were combined to generate a wider SSC map. For the UAV observational data, the atmospheric influence at each altitude was reduced to the surface altitude level using a relative atmospheric correction technique. The SSC map enabled front monitoring of SSC fluctuations caused by discharge water due to the sluice gate opening. These results demonstrated the usability of the UAV-based SSC algorithm and confirmed the possibility of monitoring rapid SSC fluctuations.

Increasing global plastic usage has raised critical concerns regarding marine pollution. This study addresses the pressing issue of floating marine macro-litter (FMML) by developing a novel monitoring system using a multi-spectral sensor and drones along the southern coast of South Korea. Subsequently, a convolutional neural network (CNN) model was utilized to classify four distinct marine litter materials: film, fiber, fragment, and foam. Automatic atmospheric correction with the drone data atmospheric correction (DROACOR) method, which is specifically designed for currently available drone-based sensors, ensured consistent reflectance across altitudes in the FMML dataset. The CNN models exhibited promising performance, with precision, recall, and F1 score values of 0.9, 0.88, and 0.89, respectively. Furthermore, gradient-weighted class activation mapping (Grad-CAM), an object recognition technique, allowed us to interpret the classification performance. Overall, this study will shed light on successful FMML identification using multi-spectral observations for broader applications in diverse marine environments.