探索三維雷達數據於對流雨細胞特徵演變預測之應用

Abstract

短延時雷達降雨預測模型是被廣泛地應用於預測對流性暴雨之技術，然而由於資料以及模型本身演算法的限制，大部分的短延時雷達降雨模型僅能預測雷雨胞的移動(advection)；然而，在真實世界中，雷雨胞除了隨著時間產生空間上的移動之外，其強度和面積等特性也會隨時間發生變化(evolution)。

此外從文獻中，我們可以發現三維雷達過去經常用來觀測對流雨胞之演化，尤其在演化的過程中，其對流核心之高度變化與降雨強度之變化息息相關。這些觀察表明了，三維雷達有潛力可以協助我們預測雷雨胞生命之演化過程。然而，過去大多數對於三維雷達與對流降雨之間關係的研究，都侷限於個案分析，對於如何將三維雷達的資訊融入短延時降雨預報模型仍是一大挑戰。

因此，本研究旨在利用深度學習方法來預測雷雨胞隨時間的演化。執行上，第一步是使用 Muñoz 等人所提出的雨胞分離及追蹤模型，用來擷取二維雷達中的雷雨胞以及其移動軌跡。然後，再利用二維雷雨胞的位置資訊，重新從三維雷達中提取完整之三維雨胞資訊，包含：雷達反射率、雨胞面積、擬和橢圓長短軸等。而第二步則是利用擷取出來的雨胞資訊，建構及訓練一 LSTM-Encoder-Decoder 深度學習模型，建立出利用過去15分鐘的雨胞資訊來預測未來15分鐘雨胞演化之預測模型。

本實驗使用了4708組雷雨胞生命週期資訊來進行LSTM-Encoder-Decoder的模型訓練。並使用1177組雷雨胞生命週期資訊來進行模型之驗證。實驗結果顯示，本研究提出之 LSTM-Encoder-Decoder 深度學習模型可以成功地預測雷雨胞的演化，並且通過使用三維雷達所提供的資訊，在預測未來15分鐘的雷雨胞強度上，可將現有不考慮雨胞強度演化預測誤差降低約20到25\%。

Object-based radar rainfall nowcasting is a widely-used technique for convective storm prediction. Due to the data and algorithmic limitations, most existing object-based nowcasting methods focus on predicting the movements of each rain object (or cell). The evolution of rain cells’ properties (e.g. cell size, shape and intensity) themselves is often neglected. It is however critical to account for the temporal changes in cells’ properties in order to improve the predictability for convective storms.

In the literature, three-dimensional (3D) radar images have been used for observing the vertical feature changes through the formation process of convective rain cells. This shows the potential of extracting useful information from 3D images to facilitate characterising the life cycle of rain cells. Most of these works however focused on analysing or reconstructing the life cycles of individual convective rain cells or storm events. It remains an open challenge to incorporate 3D radar rainfall information into object-based radar rainfall nowcasting.

In this research, we would like to explore the use of deep learning techniques to predict the evolution of convective rain cells. The proposed work comprises two main parts. The first part is rain cell data preparation. An enhanced TITAN storm tracking algorithm proposed by Muñoz et al. (2018) is employed to identify 2D rain cells and their temporal associations (or tracks) across successive time steps. The information of 2D cells are then used to extract cell properties from 3D radar images. These include mean reflectivity, area, major and minor axis lengths and the convective core altitude of each rain cell. In the second part of the work, a LSTM-Encoder-Decoder model is developed, which uses cells’ properties from the past 15 min to predict the evolution of these properties in the next 15 min.

A total of 4708 lifespans of rain cells extracted from high-resolution (5-min, 1 km, 24 levels) 3D radar images are used to train the model, and a total of 1177 extracted lifespans are used to validate the prediction result. The result suggests that the proposed LSTM-Encoder-Decoder model can well predict the evolution of cells’ properties, and, with the employed 3D information (core altitude), the prediction errors of mean reflectivity can be further reduced by 20-25 percent at 15-min forecast lead time.