利用數據驅動方法改進氣候類比法

Abstract

氣候類比法可透過計算相關大氣參數間的距離，來量化不同氣候狀態間的相 似性;此方法目前已被廣泛使用於大氣科學中的天氣預測和氣候風險研究。近年 來，深度學習以「特徵」的概念、數據驅動的方式，於各領域展現其卓越的能 力。因此，本研究旨在探索深度學習技術應用於氣候類比法中的潛力，並提出了 一種數據驅動的氣候類比方法 ClimaDist (CD)，其藉由提取氣候動態變化的特徵、 結合非傳統用於大氣領域之距離度量 (如:影像歐幾里得距離 (image Euclidean distance, IMED) 和結構相似性指數 (structural similarity index metric, SSIM)，取代 傳統方法中的原始大氣參數與歐幾里德距離。本研究以 ERA5 (ECMWF Reanalysis v5) 重分析大氣資料來測試不同模型和距離度量組合，其結果顯示，基於 CD 技術 之模型，無論是在重現觀測分佈或是預測參數變化之準確度，皆優於其他模型組 合。此外，本研究更近一步導入可解釋人工智慧技術 (explainable AI, XAI)，嘗試 以資料之角度，剖析 CD 在捕捉、學習氣候動態之觀點，包含關鍵特徵之選取以 及神經網路層間資訊傳遞之決策路徑，分析結果顯示氣候特徵差異及其時空變化 對於氣候類比之選取至關重要，然而傳統方法中經常忽略其重要性，而本研究提 出之數據驅動模型可以有效地捕捉氣候之時空特性變化。

Analogue methods are widely used in atmospheric science for weather forecasting and climate risk studies. The concept of weather/climate analogues is straightforward: it quantifies the similarity between weather conditions at target and candidate time periods by computing a distance using relevant atmospheric variables (e.g., temperatures and geopotential heights). Traditionally, Euclidean distance (ED) is used for this task, computed over the original value space of the variables.

In recent years, with the rapid advance in deep learning (DL) techniques, the concept of 'features' (or representations) has emerged, providing an alternative for analogue-related applications. This study aims to explore the potential of DL in climate-scale applications. Specifically, we propose a data-driven climate analogue approach, ClimaDist (CD), which involves constructing a data-driven climate dynamics model. ClimaDist identifies features relevant to climate dynamics, and we use these features, instead of the original input atmospheric variables, to search for analogues.

In addition to the traditional ED, we explore two other distance measures that have not been widely used in the atmospheric community: the image Euclidean distance (IMED) and the structural similarity index metric (SSIM). We tested a range of different combinations of models and distance measures using ERA5 reanalysis data. Our results suggest that the proposed CD-based models outperform all other model combinations under comparison in terms of their ability to reproduce both the observed distributions and variations.

We further employ explainable AI (XAI) techniques to investigate the perspectives of CD models in analogue identification, with focuses on feature importance and the decision flows between neural network layers. The analysis results indicate that it is critical to account for the spatial and temporal variations in the importance of different climate features, which are often neglected in traditional methods.