APPLICATION OF CONVOLUTIONAL NEURAL NETWORK TO OCCURRENCE PREDICTION OF HEAVY RAINFALL

Abstract

 A rainfall prediction model was proposed by applying Convolutional Neural Network to spatiotemporal two-dimensional data created from the time series information of meteorological observation at several sites. The rainfall threshold and the prediction lead time were set as the prediction setting. We used several meteorological variables as input data and studied a prediction model that does not use precipitation as input data. As the difference in the prediction accuracy in the prediction setting, it is confirmed that prediction accuracy decreases as the prediction lead time becomes longer, and rainfall prediction becomes more difficult as the threshold becomes higher. From the prediction model without precipitation, it turned out that the information of precipitation had an influence on the model accuracy, and it was suggested that other meteorological variables also contain information related to rainfall prediction.