RAINFALL OCCURRENCE PREDICTION WITH CONVOLUTIONAL NEURAL NETWORK

Abstract

 A rainfall occurrence prediction model was developed using a convolutional neural network (CNN), a representative machine learning algorithm in image recognition. A spatiotemporal data array was created from the time series of related atmospheric variables from multiple ground gauge observation sites and used as the image data set. By feeding the atmospheric data array into the CNN algorithm as an input, the algorithm was trained to classify whether there will be rain in the next 30 min. The trained models demonstrate promising results for three different cities in Japan, with a 64 – 76 % detection ratio for a 30 min prediction lead time. The high false alarm ratio is an issue that should be addressed in further research, with additional input data. This paper presents the basic concept of the developed model and the results from modeling tests, with various model structures and input data combinations.