SHORT-TERM RAIN PREDICTION BASED ON SATELLITE IMAGE USING CONVOLUTIONAL NEURAL NETWORK (CNN) (CASE STUDY: MAJALAYA SUBDISTRICT AND SURROUNDINGS)
Yanuar Rizky Ramadhan1, Frank Leonard Tansaulu1, Edi Riawan2,3, Muhammad Rais Abdillah2, Wendi Harjupa4, Riki Waskito5, Mimid Suamid5, Yadi Mulyadi5, Indra Mustofa5

1Department of Meteorology, Faculty of Earth Sciences and Technology, Institut Teknologi Bandung, Indonesia
2Atmospheric Science Research Group, Faculty of Earth Sciences and Technology, Institut Teknologi Bandung, Indonesia
3Center for Water Resource Development, Institut Teknologi Bandung, Indonesia
4Research Center for Climate and Atmosphere, National Research and Innovation Agency (BRIN), Indonesia
5Garda Caah, Community Volunteers, Majalaya, Indonesia

Abstract

Majalaya Subistrict is an area that often experiences flooding due to rain. The existence of reliable rain prediction information is needed to support community preparedness in the area in the face of flooding. In this study, a short-term rain prediction model for Majalaya Subistrict and its surroundings was made by training the Convolutional Neural Network (CNN) algorithm to be able to distinguish between cloud features in satellite images that indicate rain and those that do not indicate rain at 30, 60, 90, and 120 minutes to come in Majalaya Subistrict and its surroundings.

The CNN algorithm is trained using a set of historical satellite image data sourced from the Himawari-8 satellite which has gone through several preprocessing stages (cutting, composite, and classification/labeling) to produce the expected model. The model generated through the training process is then validated using a validation data set consisting of four categories of images, each of which describes the characteristics of the dominant cloud depicted in the image. The validation results are written in the form of a contingency table and then used to calculate the probability of detection, bias score, false alarm ratio, and the overall accuracy of the prediction model to analyze the level of reliability.

The results of this study indicate that the rain prediction model with a time lag of 30 minutes is able to predict all categories of images in the validation data set well. Meanwhile, the rain prediction model with a time lag of 60, 90, and 120 minutes has difficulty in predicting the image category that represents the dominance of clear skies and the image category that represents the dominance of thick clouds with high-level top (Cb). Rain prediction models with time lags of 60, 90, and 120 minutes have a high tendency to under forecast when receiving input image categories that represent the dominance of clear skies and have a high tendency to over forecast when receiving input image categories that represent the dominance of thick clouds with high-level top (Cb).

Keywords: Convolutional Neural Network (CNN), Machine learning model, Rainfall, Satellite image

Topic: Atmospheric Sciences