每年的5到10月，正值臺灣的梅雨季以及颱風季，由於救災以及提前防範的需求，我們需要一套可靠的短期天氣預報系統。過去氣象局使用的氣象模型是透過大量的物理模型去計算，龐大的計算量導致經常需要花上數小時來產生一個預報，而本篇論文的目的，就是希望能透過深度學習的方式縮短預報時間，進行短期的雷達回波外延，去預測未來5小時的雷達回波走勢，進而為防災提前做好準備。
　　其它同樣使用深度學習進行雷達回波外延的方法，通常只有雷達回波單一種資料類型，本篇論文嘗試加入風場資料，目的是希望模型能從多元的資料中提取更多的特徵，使預測更為準確。我們提出了一個使用卷積自動編碼器(Convolutional Autoencoder)架構的資料特徵融合模型，將原本複雜的高維度資料，透過卷積及降維提取出包含兩種資料特徵的低維度特徵向量，然後結合這個特徵向量與經縮小成同樣維度的雷達回波與風場資料，組合成用來訓練雷達回波外延模型的異質資料。而雷達回波外延模型，我們使用了PredRNN來實作，不同於較常見的ConvLSTM，多了一個時空間記憶單元，能夠將該層的記憶傳遞到同一時間輸入的下一層，解決了以往只能在同階層傳遞的問題。
　　實驗結果方面，為了方便理解，我們幫雷達回波依強度畫成了假色圖，也透過計算得出了風速及風向，並幫它們畫成了假色圖及風向箭頭圖。我們舉了兩好一壞，共三個案例，在好的結果中，我們的模型於強回波區域的表現都較其它兩種方法好，更在第二個案例中，清楚地保留了颱風眼的特徵；而在壞的結果中，我們的方法受到了風場資料的影響，在風速強的區域預測出了強回波，此與真實情形不符，因而造成誤判。在實驗數據的表現上，我們方法的外延結果與實際觀測之間的均方誤差達到 71.37，相較使用PredRNN 的73.33有了2.7%的提升，與ConvLSTM 的85.75更提升了20.1%，而在臨界成功指數中，於選取30 dBZ、40 dBZ和50 dBZ的閾值之下，我們分別得到0.4312、0.3267和0.1149，也是三種方法中的最佳。

From May to October each year, it is the rainy season and typhoon season in Taiwan. To reduce the damage caused by the disaster, we need a reliable short-term weather forecast system. In the past, the weather bureau used complex physical models to produce weather forecasts; however, the huge calculation often takes several hours. In this thesis, we propose a deep learning approach to shortening the forecasting time, which can extrapolate radar reflectivity for the next 5 hours, so as to prepare for disaster prevention in advance.
Different from other studies that only use radar reflectivity data for training the extrapolation model, we add the wind field data to make predictions more accurate. Moreover, we use the convolutional autoencoder architecture to train a data feature fusion model. Through the convolution operation and dimensionality reduction, the important features of the two types of data are extracted and result in feature fusion data. Then we combine this feature fusion data with the dimension reduced radar reflectivity data and wind field data as the heterogeneous one that is used to train the radar reflectivity extrapolation model. And we adopt PredRNN as the deep learning method to build the model. Unlike ConvLSTM, a newly spatiotemporal memory is added to PredRNN, which surmounts the problem that the information of memory cells can only be transmitted in the same level.
To make it easier to understand, we visualize the experimental results. For the radar reflectivity data, we make a pseudo-color map according to its intensity. As for the wind field data, we first calculate both the wind speed and wind direction, then we also make a pseudo-color map and draw a wind direction arrows graph. Three examples are demonstrated: two for good cases, and one for bad. In good cases, the results of our method are better than those of PredRNN and ConvLSTM without wind field data in strong radar reflectivity region. Especially, the typhoon eye can be clearly found in the second good case. Nevertheless, in the bad case, our method is affected by the wind field data, and a strong reflectivity is predicted in the area with strong wind speed, which is inconsistent with the real situation, thus causing misjudgment. In terms of the performance of the experimental data, the sum of squares error (SSE) reaches 71.37 between the extrapolate result of our method and the actual observation, which is a 2.7% improvement compared to 73.33 using the PredRNN, and an increase of 20.1% compared to 85.75 of the ConvLSTM without wind field data. And in the critical success index (CSI), under the thresholds of selecting 30dBZ, 40dBZ, and 50dBZ, we obtain 0.4312, 0.3267, and 0.1149 respectively, which are also the best of the three methods compared.

[1] X. Shi et al., “Convolutional LSTM network: A machine learning approach for precipitation nowcasting,” in Proceedings of Conference on Neural Information Processing Systems., pp. 802-810, Montreal, Canada, 2015.
[2] Y. LeCun et al., “Gradient-based learning applied to document recognition,” in Proceedings of the IEEE., vol. 86, no. 11, pp. 2278-2324, 1998.
[3] S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural Computation, vol. 9, no. 8, pp. 1735-1780, 1997.
[4] P. Cheung and H. Yeung, “Application of optical-flow technique to significant convection nowcast for terminal areas in hong kong,” in Proceedings of WMO International Symposium on Nowcasting and Very ShortRange Forecasting., 2012.
[5] W. Woo and W. Wong, “Operational Application of Optical Flow Techniques to Radar-Based Rainfall Nowcasting,” Atmosphere, vol. 8, no. 3, pp. 48, 2017.
[6] I. Sutskever, O. Vinyals and Q. V. Le, “Sequence to sequence learning with neural networks,” in Proceedings of Conference on Neural Information Processing Systems., Montreal, Canada, 2014.
[7] K. Gregor et al., “DRAW: A recurrent neural network for image generation,” [Online]. Available: https://arxiv.org/abs/1502.04623.
[8] I. Sutskever, J. Martens and G. E. Hinton, “Generating text with recurrent neural networks,” in Proceedings of International Conference on Machine Learning., pp. 1017-1024, Bellevue, Washington, 2011.
[9] A. Graves, A.-R. Mohamed and G. Hinton, “Speech recognition with deep recurrent neural networks,” in Proceedings of IEEE international conference on acoustics speech and signal processing., pp. 6645-6649, Vancouver, Canada, 2013.
[10] A. Graves, “Generating sequences with recurrent neural networks,” Aug 2013. [Online]. Available: http://arxiv.org/abs/1308.0850.
[11] H. Sak, A. Senior and F. Beaufays, “Long short-term memory recurrent neural network architectures for large scale acoustic modeling,” in Proceedings of the Annual Conference of International Speech Communication Association., Singapore, 2014.
[12] G. Ayzel et al., “All convolutional neural networks for radar-based precipitation nowcasting,” in Proceedings of Procedia Computer Science., vol. 150, pp. 186-192, 2019.
[13] E. Shi, Q. Li, D. Gu and Z. Zhao, “A method of weather radar echo extrapolation based on convolutional neural networks,” in Proceedings of International Conference on Multimedia Modeling., Bangkok, Thailand, 2018.
[14] M. Jaderberg et al., “Spatial transformer networks,” in Proceedings of Conference on Neural Information Processing Systems., Montreal, Canada, 2015.
[15] F. Gers, N. Schraudolph and J. Schmidhuber, “Learning Precise Timing with LSTM Recurrent Networks,” Journal of Machine Learning Researc., vol. 3, pp. 115-143, 2002.
[16] J. Geng et al., “High-resolution SAR image classification via deep convolutional autoencoders,” in Proceedings of IEEE Geoscience and Remote Sensing Letters, vol. 12, no. 11, pp. 2351-2355, 2015.
[17] W. Sun et al., “A sparse auto-encoder-based deep neural network approach for induction motor faults classification,” Measurement, vol. 89, pp. 171-178, 2016.
[18] Y. Wang et al., “PredRNN: Recurrent neural networks for predictive learning using spatiotemporal LSTMs,” in Proceedings of Conference on Neural Information Processing Systems., pp. 879-888, Long Beach, California, 2017.
[19] Central Weather Bureau, Four weather radar stations and the radar coverage area in Taiwan. https://opendata.cwb.gov.tw/opendatadoc/MSC/A0058.pdf.
[20] D. Kingma and J. Ba, “Adam: A method for stochastic optimization,” 2014.
[21] R. Pascanu, T. Mikolov and Y. Bengio, “On the difficulty of training recurrent neural networks,” in Proceedings of the 30th International Conference on Machine Learning., pp. 1310-1318, Atlanta, 2013.