CORRECTION OF EXTREME RAINFALL DATA FROM GLOBAL PRECIPITATION MEASUREMENTS SATELLITE IN INDONESIA USING MACHINE LEARNING

• Main Author: Salsabila, Muhana
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/80399
• Institution: Institut Teknologi Bandung

Extreme rainfall information is very important for human life, especially in the field of infrastructure engineering design. However, information on extreme rainfall in Indonesia is limited due to the distribution of observation stations, which are only located in certain areas of Indonesia and not in every point of the region. On the other hand, along with technological developments, there is the development of satellite data that has rainfall data at every point in the region over a long period of time, one of which is the Global Precipitation Measurement (GPM) satellite. However, extreme rainfall information from satellites is limited, especially in terms of data accuracy. However, accurate rainfall data is needed for infrastructure engineering needs. In this research, extreme rainfall data from the Global Precipitation Measurement (GPM) satellite will be corrected using machine learning. This research uses Global Precipitation Measurement (GPM) rainfall data and observation station rainfall data from BMKG to correct extreme rainfall data using 3 different machine learning models, namely: linear regression, random forest, and deep learning. Extreme rainfall calculations are carried out using the Gumbel distribution and return periods fitted to annual maxima data. Extreme rainfall that has been corrected through the machine learning process is searched for the Normalized Root Mean Square Eror (NRMSE) value so that the return period that has been corrected using the machine learning method will be identified. The study's finding shows that extreme rainfall with a return period of 5 years in Indonesia which has been corrected through a machine learning process from 3 machine learning models, namely linear regression, random forest, and deep learning has a greater error value than before correction using machine learning. The error value of extreme rainfall with a return period of 5 years which has been corrected using machine learning has increased in the value range from 0 to 0.25. Linear regression, random forest, and deep learning models show that the difference in error values increases compared to extreme rainfall before correction. But out of the three models, the one with the lowest Normalized Root Mean Square Eror (NRMSE) value and the highest model performance is the linear regression machine learning model.