Ki-Yeong Mun, Dahyeon Kang, Junhwa Chi. Korean J. Remote Sens. 2025;41:787-801. https://doi.org/10.7780/kjrs.2025.41.5.7

Ocean color remote sensing provides continuous observations of marine environments, with the Geostationary Ocean Color Imager-II (GOCI-II) offering high temporal resolution data for waters surrounding the Korean Peninsula. However, cloud cover and atmospheric interference frequently create data gaps that compromise the continuity and reliability of chlorophyll-a concentration measurements—a critical indicator of marine ecosystem health. This study proposes the Spatiotemporal Attention Partial Convolution model to reconstruct missing chlorophyll-a values in multi-day satellite imagery. The model extends two-dimensional partial convolution to three dimensions, enabling simultaneous exploration of spatial patterns and temporal dynamics. A 3D attention mechanism guides the network to focus on informative spatiotemporal features during reconstruction. To ensure robust training, we developed a comprehensive artificial cloud mask generation strategy using Gaussian Random Fields with parameterized Cloud Density Index and Missing Pixel Ratio, creating diverse missing data scenarios that exceed the variability found in observational data alone. Rigorous evaluation using actual GOCI-II cloud masks from 2024 was conducted to evaluate performance across varying missing data conditions, achieving mean absolute error of 0.036 mg/m³, root mean squared error of 0.068 mg/m³, and R² of 0.865. Long-term spatiotemporal consistency analysis across ten selected regions showed substantial improvements in spatial and temporal data coverage, with valid pixels increasing from 54% to 97%. Qualitative assessments confirmed that reconstructed images preserve natural spatial continuity without artifacts, significantly enhancing the practical utility of GOCI-II data for marine environmental monitoring and research applications requiring high temporal resolution.

Among the vessel operation data collected worldwide, static data represent a vessel’s inherent specifications. However, due to technical limitations and environmental factors, a significant portion of these data may contain missing values. In particular, dry docking operations are fundamental process for ensuring safe navigation through vessel maintenance, and the amount of paint discharge generated during ship repairs has traditionally been estimated using static data. However, dry docking data frequently contain a high proportion of missing values, which undermines the reliability of analytical results. To address this issue, this study proposes an AI-based methodology to impute missing static data for vessels engaged in dry docking operations using the Automatic Identification System (AIS). The proposed approach employs an AI-driven imputation method based on the MissForest algorithm, applied to data collected from shipyards in South Korea. This method was used to impute key vessel attributes, including length, breadth, and draft. The reliability of the imputed data was statistically validated through comparisons with baseline models using the Mann–Whitney U test and the Kolmogorov–Smirnov test. As a result, 473 missing values were successfully imputed, expanding the dataset to 1688 dry docking records.

Sea Ice Drift (SID) is an important parameter in understanding the Arctic climate dynamics and in maintaining navigation safety for the Arctic waters. SID is typically derived from keypoints extracted and matched from Synthetic Aperture Radar (SAR) imagery and is represented as a grid-based field. However, in feature tracking, the widely used Oriented FAST and Rotated BRIEF (ORB) is inherently vulnerable to low contrast, speckle noise, and complex sea ice deformation in feature tracking. To address these limitations, we propose an SID estimation framework that replaces the conventional ORB with deep learning-based methods such as SuperGlue and Local Feature TRansformer (LoFTR). In addition, multi-polarization is applied to exploit complementary information across both the feature tracking and pattern matching stages. Under polarization integration, SuperGlue reduced speed and directional RMSE by 50.8% and 37.3%, respectively, compared to ORB, while LoFTR achieved the best performance with reductions of 70.0% and 41.0%. These results demonstrate that deep learning-based methods can effectively replace the conventional ORB approach for SID estimation in the Arctic environments.