可解釋性深度學習方法於辨識平流層爆發性增溫事件之應用

Abstract

平流層爆發性增溫（Sudden stratospheric warming, SSW）事件是極地平流層中極端的天氣現象，在此一事件發生時，平流層極地渦旋的環流被破壞甚至逆轉，進一步在次季節的時間尺度上對地表產生後續的影響，了解平流層中的變異性與其和對流層之間的耦合關係，對於改進次季節至季節時間尺度上近地表場的預測十分重要。借助人工智慧在識別圖形及網格化資料上空間細節的潛力，我們可以藉由辨識SSW事件的任務來分析事件中平流層極地渦旋的空間結構。在本研究中，我使用全球氣候模式的系集模擬結果訓練機器學習模型辨識SSW事件，並採用了一種可解釋的深度學習方法測試機器學習的成效。我首先運用平流層位於60°N的一維緯向風場來訓練三種不同複雜度的類神經網路，從簡單到複雜依次為：邏輯回歸網路、淺層神經網路及深度神經網路。所有類神經網路均能以相當高的準確性識別SSW事件。為了瞭解這些類神經網路是如何學習區分SSW事件和非SSW事件，在測試階段，我沿著經度以不同長度的遮罩屏蔽掉一部分的緯向風場，以測試屏蔽一部分的資訊會如何影響類神經網路的表現，即藉由類神經網路對不同空間上的風場資訊的依賴性，來解析SSW的關鍵結構。當遮罩窗口較短時，淺層和深度神經網路並未表現出明顯的空間依賴性，而邏輯回歸網路則在約160°W附近表現出較強的空間依賴性，這一地區恰對應到緯向風的平均值為負、標準差較小之所在，隨著遮罩長度增加，淺層和深度神經網路逐漸顯現出空間依賴性。

為了進一步探討二維空間上的空間依賴性，我運用北半球以北極為中心的平流層二維緯向風場來訓練卷積神經網路，並透過不同大小的矩形遮罩來進行類似的測試。卷積神經網路對於二維風場的空間依賴性與前述一維的結果具一致性：空間依賴性較強的地方分布在負緯向風所在的區域，惟空間範圍擴展到60°N的北部和南部。除了緯向風場之外，我另使用了二維位勢高度場訓練卷積神經網路並進行相同的依賴性測試。卷積神經網路對於二維位勢高度場的空間依賴性可以分為兩個區域：1）阿拉斯加上空，對應到平均緯向風場負值所在區域，2）以北極為中心的圓形區域，前者與僅使用二維緯向風場訓練得到的結果相符，亦反映了風場和位勢高度場之間的地轉風平衡，而後者恰對應到SSW和非SSW事件之間的地形高度差異最大的地區。因此，卷積神經網路不僅能夠利用地轉關係，還學習到運用SSW和非SSW之間的對比，來區分SSW事件和非SSW事件。此研究的發現不僅表現了負緯向風場是SSW極地渦旋的關鍵特徵，也揭示SSW事件和Northern Annular Mode (NAM)負相位之間的關聯──NAM負相位的特徵之一為極區壓力高於正常水平──。此外，額外的測試當中還顯示了55°N處的緯向風場也是辨識SSW事件的一個重要特徵，此發現補充了常見的SSW定義中所依據的60°N的重要性。

本研究的結果凸顯了可解釋性深度學習工具於學習SSW空間訊息和攫取關鍵性特徵的能力，並對於SSW肇始的追溯和隨後地表影響的預測可能具有重要意義。

An advanced understanding of stratospheric variability and its coupling to the troposphere is critical to improving the prediction of near-surface fields at subseasonal-to-seasonal timescale. In the most extreme situation, a sudden stratospheric warming (SSW) event occurs and substantially perturbs the stratospheric circulation and could, subsequently, exert profound surface impacts. Artificial intelligence could be a powerful tool in recognizing SSW spatial details resulting in a better categorization of the types of disrupted vortices. Here I apply an interpretable deep learning approach to identify SSW events from non-SSW ones using a large-ensemble suite of outcomes from a global climate model. I start with a one-dimensional case by using the stratospheric zonal wind profile of SSW events circling the 60°N latitude to train neural networks with different complexity: logistic regression network, shallow neural network, and deep neural network. All neural networks can identify SSW events with fairly high accuracy. To address the interpretability of how these neural networks learn to distinguish SSW from non-SSW events, I mask out the zonal wind fields with longitudinal windows with varying lengths to test if the spatial structure of disrupted winds is decisive for the network learning. Neither shallow nor deep neural networks show apparent spatial dependence when the masking window is short, while the logistic regression network gives stronger spatial dependence centering around 160°W, where small variation and negative mean value of zonal wind resides. The dependence of shallow and deep networks emerges as the window length increases. To further explore the two-dimensional spatial dependence, I train a convolutional neural network (CNN) exploiting the two-dimensional zonal wind fields in the Northern Hemisphere. Similar tests are performed by strategically masking out the zonal wind fields by a rectangular region with varying sizes. The spatial dependence of two-dimensional neural network is largely consistent with one-dimensional networks, highlighting the region with negative zonal winds, but the spatial extents expand wider to the north and south of 60°N. In addition to zonal wind profile, I use two-dimensional geopotential height fields at 10 hPa to train CNN and perform the same mask-out analysis. Two regions of high spatial dependence emerge immediately: 1) the region where large negative zonal wind fields locate, and 2) a circular patch centering at the North Pole. The former corresponds well to the results obtained from the training using zonal wind profile solely, revealing the geostrophic wind balance between wind and geopotential height fields, whereas the latter is mainly associated with the largest geopotential height difference between SSW and non-SSW events. The CNN model, thus, utilizes not only the geostrophic relationship but also the large contrast between SSW and non-SSW, to learn the categorization task. These findings not only suggest the significance of negative zonal wind speed in characterizing a SSW polar vortex, as embodied in the common definition of SSW, but also unveil the association between SSW events and the negative phase of the Northern Annular Mode, which is characterized by the higher-than-normal geopotential height over the polar regions. Furthermore, additional tests revealed that the zonal wind field at 55°N is also an important feature in identifying SSW events, better capturing the core of negative zonal wind region and complementing the significance of 60°N on which common SSW definitions rely. The above results highlight the capability of interpretable deep learning tools in learning the SSW spatial information and revealing the spatial dependence, which may carry out important implications for the prediction of SSW genesis and subsequent surface impacts.