FogNet-V2: Multi-view Tensorized Transformer for Coastal Fog Forecasting

Reliable fog forecasting in coastal areas is a complex but vital task for safety, navigation, and environmental management. It involves multiple challenges, such as the spatiotemporal and variable inter-dependency, non-stationery, and heterogeneous data. Traditional approaches have leaned on recurrent or convolutional neural networks (2D and 3D CNNs), some even integrating attention mechanisms. However, these models often become entangled in complexity, making them difficult to interpret or to provide clear insights into the learned features. Recognizing the need for a more interpretable yet high-performing model, this research introduces a pioneering approach based on a self-attention mechanism. The proposed multi-view attention model aims to unravel various potential inter-correlations for a better understanding of the intricate interactions between inputs and targets, particularly in the spatial, temporal, and variable dimensions, and provide similar or improved performance. The novelty of this model lies in its ability to learn the spatio-variable correlation with temporal correlation for over 300 meteorological variables in parallel. By doing this, it promises a superior grasp of the complexities and uncertainties that characterize coastal fog forecasting. It aims to predict fog visibility categories below three critical thresholds (1600m, 3200m, and 6400m) by refining numerical weather prediction model outputs. In contrast to the previously proposed FogNet model, this new model (FogNet-V2) aims to provide a more robust, explainable and higher performance approach. It has been carefully crafted to eliminate some of the limitations of the older model, thereby enhancing its efficiency in learning variable and temporal-wise correlations. Data from 2009 to 2017 were used to calibrate the model, while data from 2018 to 2020 were used for testing. The comprehensive assessment of FogNet-V2's performance included comparisons with both the original FogNet and the High-Resolution Ensemble Forecast (HREF) system for 6, 12, and 24-hour lead times. The evaluation was based on eight standard metrics that cover various aspects of prediction performance. What sets this research apart is not only the innovative modeling approach but also the application of Explainable AI methods. These methods were employed to interpret the model's performance in a transparent and insightful manner. This involved detailed analyses of the importance of each variable and the sensitivity of the model to different input parameters. Such an approach contributes significantly to the transparency of AI in meteorology, a field often criticized for its 'black box' nature. In summary, this research presents a groundbreaking step towards more efficient and transparent fog forecasting in coastal regions. By leveraging a self-attention mechanism and introducing the multi-view attention model, it addresses some of the most pressing challenges in the field. The comparative analysis with existing models, coupled with the use of Explainable AI methods, ensures that this work stands as a significant contribution to both the academic community and practical applications. It offers valuable insights into not only the forecasting of fog but also the broader application of machine learning in meteorology, opening new avenues for exploration and innovation.