5:15 PM - 6:45 PM
[AAS04-P04] Development of Bimodal deep learning model for very short-term coastal sea fog Prediction
Keywords:Coastal Sea Fog Prediction, Multi-modal Learning, Closed-circuit Television Image, Multivariate Time Series Observation Data, Visibility Class
Coastal sea fog is a natural phenomenon that can pose a threat to the safety of coastal ports. Therefore, its prediction is crucial to ensure port safety and reduce economic losses. Existing sea fog prediction methods often use a single modality to understand sea fog phenomena in space and time, or rely on satellite data to reflect differences in sea fog characteristics across regions. To overcome these limitations, an AI-based sea fog prediction model was developed using both closed-circuit television (CCTV) images and text-based time-series measured (TTSM) data. CCTV image-based labeling was used to generate data consisting of three classes: Normal Visibility, Low Visibility, and Sea Fog. In addition, this study applied a learning method to overcome the imbalance of each visibility class and a method to replace gappy TTSM data. Forecasts were made in one-hour increments from 0 to 6 hours using both LSTM and Swin transformers in the forecast model structure. The model was trained with three years of data and evaluated with the remaining one year of data. The data were collected at Incheon Harbor near Seoul, which is prone to sea fog. The performance of the model in recognizing the current visibility state, based on the macro F1 score performance metric, was 86.2% (0-hour). The prediction performance at one-hour intervals was 79.1% (1 hour), 73.4% (2 hours), 70.7% (3 hours), 64.7% (4 hours), 59.6% (5 hours), and 49.3% (6 hours) with an average prediction performance of 69.0%. The proposed sea fog prediction model is expected to enhance safe navigation and efficient port operations.