Implementing a Modular Low-Cost Sensor Network for Water Level Observation and Machine Learning Prediction

Flood forecasting in Mediterranean rivers remains challenging due to the highly intermittent nature of flow and the limited availability of real-time hydrological observations. This study presents a low-cost system for water level monitoring and short-term flood prediction using high-frequency sensor measurements and meteorological data collected in Kefalonia, Greece. A multi-stage preprocessing pipeline was applied to improve data quality, including session-based filtering, outlier removal, and Savitzky–Golay smoothing. An XGBoost regression model was then trained to predict the maximum water level within a 24-hour horizon using hydrological lag features and lagged meteorological variables. The proposed model achieved strong predictive performance (R² = 0.924, RMSE = 0.293 cm), demonstrating its ability to capture the temporal dynamics of flood events. The results also confirm a strong relationship between precipitation accumulation and river response.