Assessing Flood Factors through Comparative Analysis of Machine Learning Techniques

Abstract

Floods are among the most frequent natural disasters, leading to loss of life, spreading of diseases, and destruction of property. Therefore, accurate and timely flood prediction is essential for effective disaster risk management. This study explores the application of Machine Learning (ML) techniques to improve flood forecasting systems in Malaysia using geological and hydrological features. The purposes of the study are to identify the factors influencing flood occurrence and their significance using Random Forest (RF), analyse hidden trends or patterns related to flood occurrence using Logistic Regression (LR) and K Nearest Neighbors, and evaluates the performance of the machine learning models proposed using evaluation metrics during the modelling process. A dataset consisting of 4217 samples was compiled for 42 study areas at 42 weather stations across Malaysia, incorporating 16 explanatory variables, including seven-day continuous rainfall characteristics, seasonal and geographical factors, and other relevant environmental attributes. Three classification models, RF, LR and KNN were trained to predict flood occurrence. The results indicate that recent rainfall (0.1444), cumulative (0.1088) and maximum daily rainfall (0.1058) are the most influential features. Among the models, LR performs better at classifying flood events with higher recall value (0.7218), while RF shows best performance in all other metrics. Overall, the best model identified is RF based on Receiver Operating Characteristic (ROC) curve (AUC=0.877).