以深度學習方法分析衛星雲圖並估計熱帶氣旋風場結構

Abstract

熱帶氣旋結構的估計與分析為重要的研究與防災議題。然而因觀測的限制，熱帶氣旋最佳路徑資料於2004年後才提供較完整之結構參數估計。本研究使用卷積神經網路（CNN）建立一深度學習模式，利用紅外線衛星雲圖估計熱帶氣旋軸對稱風速剖面，為一套客觀、具全球一致性之熱帶氣旋結構分析方法，以期協助解決目前熱帶氣旋結構觀測資料時空解析度不足的問題。 CNN為一種監督式學習演算法，需足夠且可信的標籤資料方可計算損失函數並優化權重，故本研究結合ERA5再分析資料、最佳路徑資料之各項結構參數、及一個參數化風速模型，建立熱帶氣旋軸對稱風速結構標籤資料（2004– 2018）。利用這組標籤資料，我們以2004–2016年的資料訓練CNN模式，並以2017–2018年的資料進行測試；結果顯示模式所估計之熱帶氣旋強度之均方根誤差與暴風半徑之平均絕對誤差，分別為9.9 kt及43.6 km，具作業實用性。利用ASCAT及SAR海表風觀測對2017–2018熱帶氣旋進行獨立校驗亦顯示，深度學習模式能合理估計熱帶氣旋結構，且可能較標籤資料更接近觀測的風速值。訓練結果也顯示，模式對強度較弱的熱帶氣旋的估計誤差較大，可能是由於衛星圖像中，系統結構較鬆散所致。此外，對於登陸的熱帶氣旋，模式估計之誤差更大；而風速結構之估計誤差，與資料的空間分佈亦有關。 利用本研究的客觀方法，通過衛星雲圖重建1981年至2020年之熱帶氣旋結構再分析資料，並利用此資料分析熱帶氣旋強度和壯度的長時間趨勢，結果顯示，大西洋在2005年後之PDI（Power Dissipation Index）、角動量累積值（AM_accum）與海表面溫度之相關性並不明顯；即隨海溫之逐年提升，PDI和AM_accum卻無顯著增加趨勢。最後，本研究嘗試將模式輸出的一維軸對稱風速剖面資訊，反演成二維風場資料，並與實際觀測比對；結果顯示，此方法已可大致掌握熱帶氣旋之不對稱風速結構。

Estimating and analyzing tropical cyclone (TC) structure are very important for TC research and disaster prevention. However, due to observational limitations, the best track data provide more complete TC structure parameter estimations only after 2004. This study uses deep learning to establish a model based on the convolutional neural network (CNN), which can estimate TC structure using satellite observations. This model is an objective and globally consistent TC structure analysis method. It is expected to help solve the insufficient temporal and spatial resolution of tropical cyclone structure observations. Since CNN is a supervised learning algorithm, labeling data are required to compute the loss function through the training process. Therefore, this study first uses the structural parameters of the best track data and a physically-based parametric wind model to estimate the axisymmetric wind speed structure of the TCs. Since the wind profiles are not close enough to the real situation, we use the ERA5 reanalysis data to correct the maximum wind and wind field at outer radii. Using this set of labeling data, we train the CNN model with the data from 2004 to 2016. With the testing data, the performance of our model is comparable with other studies. The intensity RMSE and storm wind radius MAE are 9.9 kt and 43.6 km, respectively. Independent verification of the 2017-2018 TCs using ASCAT and SAR sea surface wind observations shows that the deep learning model can estimate the TC structure reasonably. Results also show that TCs with weaker intensity have larger estimation errors of the wind speed, which may be due to the looser structure in the satellite images. For landfall TCs, the intensity errors are even larger. In addition, the estimation error is related to the spatial distribution of the dataset. With the objective method of this study, the reanalysis data of TC structure from 1981 to 2020 was reconstructed through satellite images, and the long-term trend of TC intensity and strength was analyzed using this data. The results show that the PDI (Power Dissipation Index) and angular momentum accumulation (AM_accum) are not significantly correlated with sea surface temperature (SST) in the Atlantic Ocean after 2005; that is, with the increase of SST, PDI and AM_accum have no significant increasing trend. Finally, we attempt to convert the one-dimensional axisymmetric wind speed profile into the two-dimensional wind field, and compare it with actual observations. The results show that the method could estimate the asymmetric wind speed structure of TC.