Comparison of Machine Learning Algorithms for Flood Prediction in Puri, Odisha

Abstract

Floods pose a serious risk to communities, infrastructure, and overall regional development. Puri, a district in Odisha, India, is particularly prone to flooding due to its low-lying landscape and the heavy rains that occur during the monsoon season. In such areas, quick and accurate flood forecasting becomes crucial for protecting lives, planning evacuations, and reducing damage. This study aims to compare numerous machine learning methods, including Decision Trees, Logistic Regression, Random Forests, Support Vector Machines (SVM), and Lasso Regression, for predicting possible flood events using past flood data and environmental factors like rainfall, soil moisture, temperature, and other hydrological indicators. Soil moisture is an important variable, but its dataset was incomplete. To fill these gaps, three machine learning models were tested for soil moisture prediction. Lasso Regression performed the best, giving the lowest Mean Absolute Percentage Error (MAPE) of 0.17, and was chosen to generate the missing values. With this completed dataset, multiple algorithms were evaluated for flood prediction. Logistic Regression stood out, achieving a Recall Score of 1, a Matthews Correlation Coefficient (MCC) of 0.68, and an accuracy of 0.91. These results show that Logistic Regression is a strong and reliable choice for predicting floods in the Puri region.