DEEP LEARNING MODEL FOR HEAVY RAINFALL PREDICTION HIMAWARI SATTELITE AND RAPIDLY DEVELOPMENT CUMUULUS (RDCA) DATA

• Main Author: Fadhlan Putranto, Muhammad
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/78020
• Institution: Institut Teknologi Bandung

High-Intensity Rainfall in a short period can increase the likelihood of landslides and floods. With such a short period, the disaster mitigation process will be too late. This is also worsened by the absence of an early warning system. Therefore, an accurate prediction system is needed to predict heavy rainfall. By utilizing real- time observation data from Himawari-8, Rapidly Developing Cumulus Area (RDCA) can predict the possibility of high rainfall occurrence. However, RDCA has not been able to make predictions for a long period of time. Therefore, in this research, the author proposes an extreme rainfall prediction model from a deep learning perspective by utilizing cloud observation data from Himawari-8 and RDCA index data. In this study, a Convolutional Long-Short Term Memory Model (ConvLSTM) with an encoder-forecaster architecture is designed to perform heavy rainfall prediction. In the model training process, the results show that the performance increases as the training iterations grow. In addition, in experiments with test data, the model also produces probability values that are close to the same as the RDCA index value. This can be seen from the SSIM value which is close to one. Although like that, this model still has weaknesses. Among them, for threshold values greater than 0.5 model performance is still not good in predicting the event. This is due to the unbalanced data distribution between threshold values.