Capturing Atmospheric Signatures with Convolutional Neural Networks to Predict Occurrences of Rainfall Events

Abstract

Rainfall occurrence prediction models were developed based on a convolutional neural network algorithm, which is one of the most representative machine learning algorithms for image recognition. Spatiotemporal map of atmospheric movement was prepared as image data based on ground gauged data and satellite image data. The spatiotemporal map contained information regarding the last 30 min of atmospheric movement to predict rainfall occurrence at the target area with a lead time of 30 min. The proposed model was developed for on-off (rain or no rain) forecasting at the target area to maximize the image classification functions of the CNN algorithm. Various forms of input combinations and hyperparameters were tested to evaluate the performance and applicability of the model. The evaluation index from the best model showed promising results with a detection probability of 0.836 and a critical success index of 0.456. This paper illustrates the concept of the developed model and summarizes the results from various model structures and input data combinations.