Benchmarking Lightweight Machine Learning for Disaster Prediction: Accuracy, Latency, and Readiness

Abstract

Natural disaster early warning systems are often constrained by high computational latency, which impedes the timely dissemination of critical information. This study proposes a disaster impact prediction framework designed to optimize inference speed through a lightweight machine learning approach. Utilizing a comprehensive historical dataset from the Indonesian National Board for Disaster Management (BNPB) comprising 28,773 disaster records across 34 provinces from January 2018 to May 2024, this research evaluates the performance of XGBoost and LightGBM against a Gated Recurrent Unit (GRU) model for predicting infrastructure damage. SMOTE was applied exclusively during the training phase to address class imbalance without affecting inference speed. The results demonstrate that tree-based models significantly outperform GRU in both accuracy and speed. XGBoost achieved the lowest Mean Absolute Error (MAE of 1.1649) and the fastest inference latency (0.0026 ms per sample), followed by LightGBM (MAE of 1.4006, latency of 0.0064 ms), while GRU yielded a substantially higher error (MAE of 1.4189) and latency (0.2493 ms). Statistical validation via the Wilcoxon signed-rank test confirmed the significance of these performance differences (p < 0.001). Beyond technical benchmarking, these findings support the development of Intelligent Decision Support Systems for disaster mitigation, where organizational readiness and technology acceptance are critical for successful adoption. This study discusses practical implications for deploying lightweight models on resource-constrained edge computing devices such as Raspberry Pi 4, demonstrating sub-millisecond response speed, horizontal scalability, and affordability under $100 per node.