Journal of Japan Society of Civil Engineers, Ser. B1 (Hydraulic Engineering)
Online ISSN : 2185-467X
ISSN-L : 2185-467X
Annual Journal of Hydraulic Engineering, JSCE, Vol.64
SENSITIVITY ANALYSIS ON DATA ARRAY AND MODEL STRUCTURE OF CONVOLUTIONAL NEURAL NETWORK FOR RAINFALL OCCURRENCE PREDICTION
Moonsun PARKSunmin KIMTsuguaki SUZUKIYasuto TACHIKAWA
Author information
JOURNAL FREE ACCESS

2019 Volume 75 Issue 2 Pages I_1183-I_1188

Details
Abstract

 We have developed a rainfall prediction model based on Convolutional Neural Network (CNN), which is one of the standardized deep-learning algorithms for image processing. In our previous study, we used input data with 3-dimensional array and fed into the processes called convolution and pooling. In this paper, the data array has been modified into a 2-dimensional array to see how the models’ performance vary. Moreover, the models were also tested with variant filter sizes and model structures. From the experimental results, we found that using the modified data array as an input was good enough to lead to similar results and it gives more effective results in some cases than the previous results with the 3-dimensional data array. For an additional experiment, data with 10 minutes resolution was tested with the same model structure but the result was not improved in predicting rainfall occurrence 1 hour ahead compared to the simulation done with the hourly data.

Content from these authors
© 2019 Japan Society of Civil Engineers
Previous article Next article
feedback
Top