An unsupervised machine learning approach for estimating missing daily rainfall data in peninsular malaysia

Abstract

Rainfall data plays a vital role in various fields including agriculture, hydrology, climatology, and water resource management. Stakeholders had raised concerns over the issue of missing rainfall data as it presents a huge obstacle in achieving reliable climate forecasts. Therefore, it is necessary to perform accurate estimation for the missing daily rainfall data. Each year, the peninsular Malaysia experiences a significant rainfall event during the monsoon period due to the North-East monsoon (NEM) wind. The intricate spatial rainfall dynamics requires a computational model, capable of generating accurate estimates and deciphering hidden patterns from the missing data. An unsupervised machine learning model known as the Self-Organising Feature Map (SOFM) is developed to estimate the missing daily rainfall across 10 rainfall stations during the NEM period between 2010 and 2020. The SOFM exhibited reliable performance across the percentage of missingness between 10% to 50%. Below 50% missingness, the evaluated statistical metrics, coefficient of determination (R2) is attained above 0.5, ranging between 0.504 and 0.915; root mean square error (RMSE) between 15.9 to 22.7. The feature maps enabled the visualisation of the relationship between the rainfall intensity and studied rainfall stations. The feature maps suggested that the studied rainfall stations are inhomogeneous.