해양 재해 예측을 위한 멀티모달 학습·추론 기술 개발

Abstract

○ 멀티모달 해양 데이터 AI 표현학습 개발 - 지역-국지연안의 재분석·위성·부이관측을 통한 기온, 수온, 해상풍, 파랑 데이터 수집 및 학습데이터 구축 - 멀티모달 해양 데이터의 표현학습 기술 개발 - 확장된 수용영역 합성곱 신경망, 주의집중 메커니즘, 장단기 메모리 네트워크 구조 개발

○ AI 기반 해양 재해 예측 모델 개발 - 재해성 파랑, 이상 수온 예측 - 한반도 주변해역 적용, SOTA 대비 예측 정확도 개선 - 이상 수온 및 고파랑에 대한 성능을 구분하여 평가, 정량적 성능 향상치 제공

The accurate prediction of extreme water temperature is very essential in understanding the variability of the marine environment and in reducing marine disasters maximized by global warming. In this work, we propose a self-attention-based two-pathway approach consisting of separate spatial and temporal encoder networks for the precise prediction of coastal water temperature and coastal waves, particularly the extremely high temperature and hazardous coastal waves, through effective spatiotemporal representation learning. We assess the performance of the proposed vision Transformer,(ViT)-Transformer encoder,(TE) and multi-path ViT-TE networks with the best experimental conditions for a consecutive 7-day ahead prediction of the extreme coastal water temperature by applying the proposed framework to the waters of the Korean Peninsula and by performing various comparative ablation experiments to determine the combination of the self-attention-based state-of-the-art model, modality,and resolution of spatiotemporal data. Compared with conventional convolution and recurrent networks, the proposed framework based on the self-attention mechanism, which captures the long-range dependencies of the input data, obtained well a better predictability by allowing the learning of the spatiotemporal teleconnection of regional-scale water temperature features affecting the coastal water temperature with a proper time-lag. The explainabilty of the deep neural network is presented by visualizing the spatial and temporal attention maps of the trained model for prediction and confirming that the results are consistent with the major oceanic currents in the Korean Peninsula and autocorrelation in the time series data.