Sequential Images Prediction Using Convolutional LSTM with Application in Precipitation Nowcasting

Abstract

The precipitation nowcasting task is to predict the rainfall intensity in a local region in the short future. In this thesis, we formulate this task as a sequential images prediction problem in which both the input and output are temporal sequences of spatial images. The problem will be viewed from the machine learning perspective. Inspired by fully connected long-short term memory (FC-LSTM) which is good at capturing the temporal relationship, we build a convolutional long-short term memory (ConvLSTM) by adding convolutional operation in the input-to-state and state-to-state transitions. By incorporating an encoder-forecaster model structure, the experiments we have run show that stacked ConvLSTM is better to capture spatio-temporal relationships and outperforms FC-LSTM in the precipitation nowcasting task.