Machine Learning and IoT Monitoring for Stream-Level Nowcasting: Event-Based Performance Assessment

Machine learning (ML) models utilizing high-frequency Internet of Things (IoT) monitoring data offer significant potential for stream-level nowcasting. The assessment of model performance is often a challenging task because conventional approaches are typically influenced by prevailing low-variability baseline conditions and are unable to assess a model’s ability, or inability, to reproduce rare but critical stream-level surges. To address this limitation, this study proposes a framework combining standard and event-based performance metrics. Using two years of multi-source observations (i.e., stream level, volumetric water content, suction, and rainfall) collected in a small catchment in southern Italy, a set of Random Forest (RF) models was developed to systematically investigate the effect of adopting alternative feature sets, calibration strategies, and target formulations. The assessment of the calibrated models was performed considering 29 stream-level surge events. The results of the analyses revealed that better statistical accuracy over the entire validation period does not necessarily translate into better capability in reproducing critical event dynamics. Three main findings emerged: (i) models predicting short-term stream-level variations consistently outperformed those directly predicting stream level; (ii) calibration using the complete continuous dataset produced more robust models than calibration based exclusively on surge events; and (iii) a reduced set of features provided the most favorable performance balance between predictive performance and model robustness among the tested configurations. Overall, these results suggest that combining conventional and event-based evaluation metrics can provide a complementary assessment of ML forecasting models, highlighting the importance of considering both overall predictive accuracy and event-scale dynamics.