Forecasting Floods Using Classification Based Machine Learning Models

Abstract

Floods are the most recurrent natural hazards that ravage the state of Bihar, India. Resultantly, crops, roads, buildings and other infrastructure are severely damaged. Moreover, thousands of people are forced to migrate from one place to another in addition to many who lose their lives due to such floods. Therefore, flood mitigation strategies with a focus on accurate flood forecasting in real time are required. However, traditional forecasting models, like physical and conceptual models, are very complex and require domain knowledge and thus machine learning (ML) based data-driven models are used for forecasting floods. Many ML based flood forecasting models take into account the historical rainfall-runoff data for floods. However, these models forecast floods having short lead time. For efficient flood mitigation, forecasting of floods providing a longer lead time is required. Accordingly, in this paper, ML techniques viz. Artificial Neural Network, k-Nearest Neighbor, Logistic Regression, Naive Bayes, Random Forest and Support Vector Machine have been used for forecasting floods with longer lead times. For this, the monthly mean of precipitation and temperature of twelve flood affected districts of Northern Bihar have been used. Performance of these ML models have been compared on parameters like accuracy, precision, recall, F-measure and area under the ROC curve. It can be inferred from the experimental results that the support vector machine based flood forecasting model performed comparatively better in terms of accuracy and precision whereas, Naive Bayes flood forecasting model performed comparatively better in terms of recall, F-Measure and area under the ROC curve.