Attention-Based CNN-LSTM Fusion of IoT Sensor Streams for Short-Term Urban Flood Early Warning

Abstract

Urban flooding is a recurrent hazard in many Indonesian cities, and reliable short-term early warning can substantially reduce loss of life and property. Effective warning depends on turning noisy, heterogeneous, real-time observations into accurate short-horizon forecasts, but conventional models typically capture either the spatial relationships among sensors or the temporal dynamics of each series, rarely both, and they degrade when sensor readings are missing. This study designs and evaluates FloodNet-IoT, an attention-based convolutional-recurrent framework that fuses heterogeneous Internet-of-Things streams, including water-level sensors, rain gauges, and upstream indicators, together with weather information, to predict flooding at a microcatchment scale. A convolutional component captures cross-sensor spatial structure, a long short-term memory component captures temporal dynamics, and an attention mechanism weights the most informative time steps, while an edge gateway performs preprocessing, temporal alignment, and missing-data imputation to improve robustness. Following a design science methodology, the framework was compared against a persistence baseline, classical and recurrent sequence baselines, and a convolutional-recurrent model without attention, using regression metrics, event-detection metrics, a lead-time analysis, and a missing-sensor robustness analysis. The results are intended to show whether attention-based multimodal fusion improves short-term flood warning relative to simpler models and to provide a deployable and robust blueprint for Internet-of-Things-driven flood early-warning systems in data-rich but noisy urban environments.