A recall-optimized LSTM-XGBoost ensemble model for river flood forecasting using multivariate time-series data-case study in Sri Lanka

Riverine (fluvial) flooding remains a significant climate-related disaster in many countries. Extreme rainfall events caused by climate change cause significant damage and disruption to communities living near riverbanks or neighboring land areas, particularly when risk-sensitive development is ignored. This necessitates an effective mechanism to predict river flooding events and provide timely early warnings to the most vulnerable populations in high-risk flood-prone areas. The dynamics of flood risks require a localized, automated, and scalable solution that can be centrally monitored. This study presents a river flood forecasting system integrating artificial intelligence (AI) and Internet of Things (IoT). This system was deployed in one of the flood-prone river basins in Sri Lanka to assess the operational feasibility of real-time hydrological monitoring and flood-risk classification. The proposed system employed two machine learning models, Long Short-Term Memory (LSTM) and Extreme Gradient Boosting (XGBoost), for flood-risk prediction and IoT sensors (rainfall and ultrasonic sensors) for real-time hydrological data collection. The machine learning models were developed using rainfall, discharge rate, and river water level data to predict flood-risk states. A rule-based ensemble strategy was implemented by combining the outputs of LSTM and XGBoost to prioritize positive-class detection and reduce missed positive-labeled observations. Model performance was evaluated primarily using recall for the positive flood cases, while the overall accuracy was reported as the complementary metric alongside positive-class recall, providing an overall assessment of classification performance. Among the evaluated input window configurations using the same chronological holdout, the rule-based ensemble model with a 1-day input window showed the most favorable observed balance of performance metrics, achieving 100% positive-class recall, with 98.73% accuracy, 50.00% positive-class precision, and 66.67% positive-class F1-score. The 100.00% recall was based on correctly classifying all four positive-labeled observations in the held-out test set (4/4; 95% exact binomial CI: 39.8%-100.0%) and these observations occurred consecutively from November 25 to 28, 2024, and represent one continuous positive-labeled flood episode. The proposed safety-oriented AI-IoT framework demonstrates the operational feasibility of integrating real-time sensing with AI-based flood forecasting; however, longer-term field deployment, including actual flood events, is required to evaluate flood-event prediction performance under real-world flood conditions.