近年來造成台灣地區重大的天氣災害事件，主要以劇烈降雨事件為主。而這些事件中，短延時極端降雨事件在天氣預警與防救災作業上，因為系統的時空尺度較小導致預警能力有限，進而導致在天氣預報與防救災作業上很大的壓力。在過去的天氣預報技術中，數值模式在這類時間尺度很短之天氣事件上，受限於觀測資料採樣頻率、人工預報流程與數值模式計算能力等因素。導致在短延時極端降雨事件的預報上有很大的不確定性，也無法提供即時且具充分反應時間的預警。本計劃嘗試利用高時間與空間覆蓋度的衛星遙測資料，配合目前人工智慧（ＡＩ）技術中較為成熟的機器學習模組，發展可用來診斷短延時極端降雨事件的分析工具。透過對於衛星資料之長期時空特徵分析與大量短延時極端降雨事件的研究，協助發展利用即時衛星資料診斷短延時極端降雨事件之預警模型。我們將利用所開發之極端降雨診斷技術，回溯高空間解析度之短延時極端降雨事件，期待可應用在天氣預報與防救災的即時預警作業上。本計畫預計分三年執行，第一年將依據地面觀測紀錄標定短延時極端降雨事件，並收集與整理相對應之衛星觀測資料。第二年規劃針對衛星資料之特性，測試合適之降尺度分析技術，並利用在發展機器學習技術為基礎之測站短延時極端降雨事件診斷工具上。第三年將應用機器學習診斷工具，回溯短延時極端降雨事件之空間分佈與變異特徵。期望能基於機器學習技術與即時衛星資料，建置短延時極端降雨事件之預警模型。
Extreme rainfalls become one major weather disaster events in Taiwan during the past several years. However, the warning system for this type of weather events are limited due to the small temporal-spatial scale. It caused serious issues for weather forecasting and disaster preventions and reductions. In the traditional weather forecasting method, there are some uncertainties of the extreme rainfall due to the limitation of numerical simulations. It is also hard to provide the reliable real-time predictions. In this study, we will use the advanced remote-sensing dataset which provide more comprehensive temporal and spatial coverage. We are expecting to develop a machine-learning based tool to diagnose short duration extreme rainfall events. We will also apply this technology to retrieve the short duration extreme rainfall events with high spatial and temporal resolutions in order to improve the warning system in both weather forecasting and disaster prevention and reduction. We plan to spend three years to finish this project. First year, we will be collecting, diagnosing, and organizing the data of short duration extreme rainfall events from satellite image based on the surface observations. In other hands, we will analyze the characters of satellite observations and will use that information to test different methods of dimension reduction in the second year. Those results can be used to developing the machine-learning based model to diagnose short duration extreme rainfall events. In the third year, the application of diagnosing tool will be tested by retrieving the extreme rainfall events with high spatial and temporal resolutions satellite observation. We are expecting to establish a warning system for short duration extreme rainfall by using machine learning technology in this project.
