Flood Event Classification Using Hybrid Random Forest and Deep Neural Network with Time-Based Feature Expansion

Abstract

Floods are one of the most frequent natural disasters that occur in Indonesia, causing extensive damage to roads, bridges, and homes while severely disrupting daily life and economic stability. This paper proposes a flood prediction-classification model that learns using a hybrid architecture of two machine learning algorithms: Random Forest and Deep Neural Network. The benefit of using this approach lies in its expansion of temporal and spatial data by systematically incorporating past values of each variable for the same locations, as well as values from nearby neighboring locations. The proposed model specifically includes the five-step lagged values (t-5 and t-6) for each spatial point to enhance its predictive capability. The training data used in this study covers the West Java region from 2018 to 2024 and consists of six key environmental and demographic indicators: rainfall, rain events, sunshine duration, green open space, population density, and infiltration wells. Experimental results from the evaluation show that the hybrid model with extended temporal features performs best, successfully achieving a Macro F1-Score of 0.84 and an accuracy of 0.83. The prediction-classification results of the models that are used to forecast flood risk for the years 2025 to 2030 indicate a clear increase in flooding across metropolitan and industrial areas. Furthermore, the "High" risk category demonstrates a noticeable upward trend expected to emerge prominently by the year 2028.