20200713_Convolutional LSTM Network: A Machine Learning Approach for Precipitation Nowcasting


 1. Title, Journal and Authors

Title : Convolutional LSTM Network: A Machine Learning Approach for Precipitation Nowcasting

Authors : a, Dit-Yan Yeunga, Hao Wanga, Zhourong ChenaXingjian Shi

a Hon Kong University of Science and Technology


2. Summary

In this paper, we use the Radar Echo Dataset, which is a weather radar photo dataset in Hong Kong for three years from 2011 to 2013. Since it deals with real-time prediction of rainfall (Precipitation Nowcasting), it aims to forecast weather radar from 1 hour to 6 hours by looking at weather radar data up to the present time. In this study, the FC-LSTM (Fully-Connected Long Short-Term Memory) model is used as a baseline. FC-LSTM is a multivariate version of the LSTM where the cell's input, output, and state are all 1-dimensional vectors. LSTM Future Predictor Model consists of an encoding model and a decoding model. The encoding model continuously receives input, finds a hidden representation of time series data, and the prediction model receives the final state of the encoding model and predicts the values ​​after the data given as input from the encoding model. Subsequently, the Future Predictor model is changed to a Convolutional LSTM based Encoding-Forecasting model.

This study compares two-stacked ConvLSTM and FC-LSTM and three different ROVER models. The ROVER model is one of the traditional Radar Echo Extrapolation based models. We can see that all evaluation indexes are ahead of the existing ROVER or FC-LSTM models. FC-LSTM lags behind all ROVER models, but it is impressive that the Convolutional LSTM leads all ROVER models.


3. Originality & Creativity 

When the input variable is 2D image data, it has a feature that the size increases significantly even if the data size is slightly increased. In general, when using image data, each image feature vector is extracted with CNN and used as the input of LSTM. However, the convolutional LSTM takes a completely different approach. It immediately puts convolution into the LSTM internal operation itself.


4. Application to research 

I will do research to predict future water quality using two-dimensional water quality data. It was helpful for data pre-processing and model selection. 



5. Contact

Jeongwoo Moon

Ph.D. Program Student

Environmental Systems Engineering Lab.

School of Earth Sciences and Environmental Engineering

Gwangju Institute of Science and Technology

1 Oryong-dong Buk-gu Gwangju, 500-712, Korea


Phone : +82-10-9384-8271

Office : +82-62-715-2461

E-mail : jeongwoomoon@gist.ac.kr


첨부 (0)