Model Klasifikasi Awan Konvektif dalam Peringatan Dini Banjir Sumatera Utara dengan Pendekatan CC-Unet

Abstract

Handling cloud classification in the context of natural disaster mitigation needs is a complex problem. The development of computer science is very much needed to assist flood disaster mitigation in providing tools and predictive models in solving the complexities that occur in handling natural disasters. Classification of convective clouds in flood early warning in North Sumatra is carried out to identify hydrometeorological disasters that currently often happen. The convective system that is formed can cause mesoscale rainfall, which is over a large area for a long duration of time. Currently, Weather Research and Forecasting (WRF) is a numerical weather model that is used to make weather forecasts at the North Sumatra BMKG. To improve the accuracy of weather forecasts, especially for predicting heavy rain events, it is necessary to improve the output of the WRF model with assimilation techniques to correct the initial data. To improve the accuracy of weather forecasts, especially for predicting heavy rain events, it is necessary to improve the existing models. In this study, a deep learning model was developed to improve the accuracy of the classification results of convective clouds or convective clouds. The model developed in this study is the CC-Unet model, which is a U-Net-based deep learning model. The CC-Unet algorithm is built on a Fully Convolutional Network (FCN) architecture which is trained using a convective cloud image dataset. This study uses two types of data, namely primary data, and benchmark data. The primary data used is the Himawari-8 satellite image, and the benchmark data used is the cloud image dataset 38-cloud dataset. In the pre- processing stage, Himawari satellite image data were verified using observation data. The training and testing of the model are based on a random train-test split to separate the training and test data. With a series of model tests, the results obtained that the CC-Unet model has a better performance in terms of accuracy of 97.29% compared to the U-Net model of 94,17%, SegNet model of 96.21% and Ref Unet of 95.97%. The significance test of the CC-Unet model with the Friedmann method at a significance level of 95% or an error level of 5% (0.05) gave significantly different performance results to the Segnet model with an exact p-value of 0.04296 <0.05 and Refined Unet of 0, 04882<0.05. The results of the ground truth image are also compared with the predicted image based on the root mean square error (rmse) value of 0.0218 which is very small. These results indicate that the level of similarity between the ground truth image and the predicted image is very close.