[M-GI27] Data-driven approaches for weather and hydrological predictions

[MGI27-05] Automatic Front Detection Using Deep Learning: Leveraging Temporal Data and Local Explanations with Attention Mechanisms

*Takumi Matsuda¹, Shoichi Shige¹, Kazumasa Aonashi¹ (1.Kyoto University)

Keywords:Fronts, Artificial intelligence, Deep learning, Neural networks

This presentation discusses the development of an automatic front detection method from atmospheric field data using deep learning. It highlights improvements in the accuracy of distinguishing adjacent fronts, which was challenging in conventional analyses based on single-time-step information, as well as enhancements in the interpretability of the model. As input, we used physical variables such as temperature, specific humidity, and wind speed at various pressure levels. To capture the continuity and temporal evolution of fronts, we integrated the front estimation results from the past 6 and 12 hours. We trained a U-Net++ based semantic segmentation model using frontal analysis data from the Japan Meteorological Agency. Furthermore, we integrated a Channel Attention mechanism at the input layer, enabling dynamic adjustment of the contribution of each channel, which allowed for the visualization of how the model interprets each input sample.
In the evaluation experiments, we compared our method with conventional approaches using metrics such as Probability of Detection (POD), Success Rate (SR), and Critical Success Index (CSI). While the overall performance was comparable to conventional single-time-step methods, a significant improvement was observed in reducing the confusion between adjacent fronts. Furthermore, the visualization of attention weights for individual input samples revealed that the importance of each variable varied depending on time and season. Additionally, we will present the results of a comparison with models trained on U.S. National Weather Service’s frontal analysis data.