<![CDATA[Indonesia frequently faces hydrometeorological disasters, with flooding being one of the most common and damaging. This study examines recurrent inundation in the Ciledug Indah Housing complex, Tangerang, an area highly vulnerable to overflow from the Kali Angke river. To address this persistent issue, the research proposes and evaluates an autonomous, field-ready Flood Early Warning System (FEWS) integrating Internet of Things (IoT) sensing, machine learning, and real-time alert delivery. The system deploys JSN-SR04T ultrasonic and tipping bucket sensors, supported by solar power and dual connectivity (Wi-Fi and GSM), enabling continuous operation despite outages or unstable networks conditions frequently experienced during flood events. Its primary scientific contribution is a practical two-stage hybrid machine learning framework: a Long Short-Term Memory (LSTM) model forecasts short-term river water levels, while a Random Forest (RF) classifier translates those predictions into actionable risk categories—Safe, Alert, or Warning. Separating numerical forecasting from categorical decision-making enhances accuracy, interpretability, and usability compared with single-model approaches. Automated community notification is enabled through Firebase Cloud Messaging (FCM), ensuring rapid dissemination of warnings. Experimental evaluation using 49 days of continuous river level and rainfall data (September 27–November 15) demonstrates strong predictive performance (LSTM RMSE 0.4276 m). The RF classifier achieved 99.0% accuracy; however, this figure must be interpreted cautiously due to dataset imbalance dominated by non-critical conditions and the absence of actual flood events. Overall, the proposed FEWS offers a resilient, scalable, and field-validated solution for flood detection, prediction, and public warning, contributing to more proactive urban disaster mitigation.]]>
