Tae-Hoon Kim
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Tae-Hoon Kim has a total of 14 co-authors. Here are the top co-authors by number of shared publications:
- Seung-Hyun Moon
School of Software, Kwangwoon University, 20 Kwangwoon-ro, Nowon-gu, Seoul 01897, Republic of Korea
4 shared publications - Yong-Hyuk Kim
Department of Computer Science, Kwangwoon University, Seoul 01897, Republic of Korea
4 shared publications - Chae-Rim Park
Kwangwoon University, Seoul, Republic of Korea
1 shared publication Do-Yeong Kim
1 shared publication- Do-Youn Kim
Ara Consulting & Technology, Incheon, Republic of Korea
1 shared publication - Hae-Sang Jeong
Korea Environment Institute, Sejong, Republic of Korea
1 shared publication
Latest Publications
- Open Access
- 4 May 2026
- Published by Elsevier in SSRN Electronic Journal
Knowledge-Guided Adaptive Learning Framework for Robust Drift Path Prediction in Maritime Environments
In maritime search and rescue (SAR), rapid and accurate prediction of a castaway’s drift trajectory is essential because survival probability decreases sharply over time. However, conventional physics-based drift prediction models often exhibit substantial uncertainty because they cannot fully capture complex environmental drivers and the diverse characteristics of drifting objects. To address this limitation, this study proposes a machine learning-based drift prediction framework that integrates observed surface drifter data, OpenDrift simulations, and environmental forcings (wind velocity and current velocity). Within this framework, physics-based drift models and historical observations are treated as structured knowledge that guides the learning process. Five prediction models (linear regression, support vector regression, multi-layer perceptron, light gradient boosting model, and long short-term memory) are employed for direct trajectory prediction and correction of OpenDrift forecasts. The framework further incorporates time-series feature expansion, feature selection using SHapley Additive exPlanations, and incremental learning to adapt to previously unseen drifting objects. In addition, an ensemble strategy is introduced that combines genetic algorithm-based weight optimization with dynamic updates via the weighted majority algorithm, enabling robust performance under real-world maritime uncertainty. Experiments using 2021–2023 surface drifter observations show that the proposed framework reduces mean absolute error by more than 70% and improves normalized cumulative Lagrangian separation by more than 80% compared with OpenDrift. These results demonstrate that the proposed framework not only enables reliable drift prediction but also serves as a practical, knowledge-driven decision support system for maritime SAR operations, supporting more efficient search and rescue planning.
- 4 May 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessKnowledge-Guided Adaptive Learning Framework for Robust Drift Path Prediction in Maritime Environments
In maritime search and rescue (SAR), rapid and accurate prediction of a castaway’s drift trajectory is essential because survival probability decreases sharply over time. However, conventional physics-based drift prediction models often exhibit substantial uncertainty because they cannot fully capture complex environmental drivers and the diverse characteristics of drifting objects. To address this limitation, this study proposes a machine learning-based drift prediction framework that integrates observed surface drifter data, OpenDrift simulations, and environmental forcings (wind velocity and current velocity). Within this framework, physics-based drift models and historical observations are treated as structured knowledge that guides the learning process. Five prediction models (linear regression, support vector regression, multi-layer perceptron, light gradient boosting model, and long short-term memory) are employed for direct trajectory prediction and correction of OpenDrift forecasts. The framework further incorporates time-series feature expansion, feature selection using SHapley Additive exPlanations, and incremental learning to adapt to previously unseen drifting objects. In addition, an ensemble strategy is introduced that combines genetic algorithm-based weight optimization with dynamic updates via the weighted majority algorithm, enabling robust performance under real-world maritime uncertainty. Experiments using 2021–2023 surface drifter observations show that the proposed framework reduces mean absolute error by more than 70% and improves normalized cumulative Lagrangian separation by more than 80% compared with OpenDrift. These results demonstrate that the proposed framework not only enables reliable drift prediction but also serves as a practical, knowledge-driven decision support system for maritime SAR operations, supporting more efficient search and rescue planning.
- Open Access
- 14 July 2025
- Published by Association for Computing Machinery (ACM)
- p. 843-846
Evolutionary Ensemble for Prediction of Drifter Trajectories Using Weighted Majority
- 14 July 2025
- Published by Association for Computing Machinery (ACM)
- p. 843-846
Open AccessEvolutionary Ensemble for Prediction of Drifter Trajectories Using Weighted Majority
- Open Access
- 14 July 2025
- Published by Association for Computing Machinery (ACM)
- p. 887-890
A Genetic Ensemble Method for Correcting Maritime Wind Forecasts
- 14 July 2025
- Published by Association for Computing Machinery (ACM)
- p. 887-890
Open AccessA Genetic Ensemble Method for Correcting Maritime Wind Forecasts
- Open Access
- 14 July 2024
- Published by Association for Computing Machinery (ACM)
- p. 675-678
Evolutionary Ensemble for Predicting Drifter Trajectories Based on Genetic Feature Selection
In the event of a maritime accident, accurately predicting the trajectories of objects is crucial for devising appropriate measures to rescue individuals in distress and prevent the spread of oil spills. To achieve this, we created four machine learning models using data obtained from the drifter and particle trajectory modeling framework. However, the data had only six input factors, which were insufficient for accurate trajectory prediction. Therefore, we conducted feature expansion using time series characteristics of the data and created 24 additional features to use as input factors, resulting in an average performance improvement of 74% compared to using only the basic six features. In addition, we performed feature selection using a wrapper method based on genetic search to remove unnecessary features, improving computing time by an average of 28% and performance by an average of 0.1%. Subsequently, we conducted an ensemble of the four machine learning models. We used weighted voting as the ensemble technique and optimized the weights using a genetic algorithm. The results showed better performance than the ensemble model with uniform weights by an average of 1.9% and the four base models by an average of 5.4%.
- 14 July 2024
- Published by Association for Computing Machinery (ACM)
- p. 675-678
Open AccessEvolutionary Ensemble for Predicting Drifter Trajectories Based on Genetic Feature Selection
In the event of a maritime accident, accurately predicting the trajectories of objects is crucial for devising appropriate measures to rescue individuals in distress and prevent the spread of oil spills. To achieve this, we created four machine learning models using data obtained from the drifter and particle trajectory modeling framework. However, the data had only six input factors, which were insufficient for accurate trajectory prediction. Therefore, we conducted feature expansion using time series characteristics of the data and created 24 additional features to use as input factors, resulting in an average performance improvement of 74% compared to using only the basic six features. In addition, we performed feature selection using a wrapper method based on genetic search to remove unnecessary features, improving computing time by an average of 28% and performance by an average of 0.1%. Subsequently, we conducted an ensemble of the four machine learning models. We used weighted voting as the ensemble technique and optimized the weights using a genetic algorithm. The results showed better performance than the ensemble model with uniform weights by an average of 1.9% and the four base models by an average of 5.4%.
- Open Access
Privacy Safe Representation Learning via Frequency Filtering Encoder
Deep learning models are increasingly deployed in real-world applications. These models are often deployed on the server-side and receive user data in an information-rich representation to solve a specific task, such as image classification. Since images can contain sensitive information, which users might not be willing to share, privacy protection becomes increasingly important. Adversarial Representation Learning (ARL) is a common approach to train an encoder that runs on the client-side and obfuscates an image. It is assumed, that the obfuscated image can safely be transmitted and used for the task on the server without privacy concerns. However, in this work, we find that training a reconstruction attacker can successfully recover the original image of existing ARL methods. To this end, we introduce a novel ARL method enhanced through low-pass filtering, limiting the available information amount to be encoded in the frequency domain. Our experimental results reveal that our approach withstands reconstruction attacks while outperforming previous state-of-the-art methods regarding the privacy-utility trade-off. We further conduct a user study to qualitatively assess our defense of the reconstruction attack.
Open AccessPrivacy Safe Representation Learning via Frequency Filtering Encoder
Deep learning models are increasingly deployed in real-world applications. These models are often deployed on the server-side and receive user data in an information-rich representation to solve a specific task, such as image classification. Since images can contain sensitive information, which users might not be willing to share, privacy protection becomes increasingly important. Adversarial Representation Learning (ARL) is a common approach to train an encoder that runs on the client-side and obfuscates an image. It is assumed, that the obfuscated image can safely be transmitted and used for the task on the server without privacy concerns. However, in this work, we find that training a reconstruction attacker can successfully recover the original image of existing ARL methods. To this end, we introduce a novel ARL method enhanced through low-pass filtering, limiting the available information amount to be encoded in the frequency domain. Our experimental results reveal that our approach withstands reconstruction attacks while outperforming previous state-of-the-art methods regarding the privacy-utility trade-off. We further conduct a user study to qualitatively assess our defense of the reconstruction attack.