Sensors prioritisation for hydrological forecasting based on interpretable machine learning

Abstract

The digitalisation of the hydrological sector introduces new challenges related to IoT network implementation, extensive data management, and real-time analysis while offering significant opportunities to improve hydrological forecasts. Reliable information is crucial for managing hydrogeological risks and optimising water usage, particularly in the current era of climate change, marked by frequent and severe extreme events such as intense precipitation and prolonged droughts. This study aims to enhance short-term hydrological predictions by prioritising sensors based on interpretable machine learning. We propose an evaluation framework that involves tuning machine learning-based hydrological models for different horizons, applying leave-one-out cross-validation to simulate sensor failures and evaluate their significance, and defining sensor priority levels. Conducted in the South Tyrol watershed (northern Italy), this study uses data from streamflow gauges and weather stations. The results show that specific sensors significantly impact forecasting accuracy, and prioritisation improves the reliability of hydrological predictions. These findings highlight the importance of maintaining critical sensors and provide a data-driven methodology for optimising resource allocation in monitoring system maintenance, ultimately enhancing the robustness of hydrological forecasting and risk mitigation strategies.