A common framework for the development of spring water contamination early warning system in western Mediterranean karst areas: Spanish and French sites.

Karst water resources, traditionally used worldwide for drinking purposes, are highly vulnerable to contamination. Scientific-technical efforts must therefore be done to ensure sufficient water quality for human consumption. Early-Warning-Systems emerge as an effective spring scale protection strategy for real-time identification of contamination episodes at drinking water capture points. With this ambition, the proposed stepwise procedure for the implementation of site-specific Early-Warning-Systems (EWS), focuses on three critical features: identification of groundwater contamination proxy parameters, warning dissemination and system validation. It was tested at three karst springs affected by temporary faecal contamination and intended for supplying drinking water to populations in Spain (Ubrique) and France (Montpellier). The applied statistical techniques, coupled to the analysis of individual flood events at karst springs, allowed to identify the optimal combination of groundwater and contamination proxy parameters at each study site according to the main contaminants and recharge mechanisms. The decision tree-like workflow, used to evaluate groundwater quality at hourly time step and set up the EWS, was then constructed considering a specific combination of hydroclimatic, hydrodynamic and physical variables together with national regulations for drinking water. Hence, warning thresholds were adapted to the behaviour of each karst system, and used to trigger the alarm when specific parameters and proxies exceeded the defined limits. The performance of the EWS was assessed by implementing 4 specific Key Performance Indicators (KPIs) dedicated to appreciate the effectiveness of the workflows in identifying contamination events and verifying estimated warning thresholds. The performed analysis demonstrated an overall successful functioning of the EWS for the three case studies, with mean anticipation times ranging between 12 and 45 h and <3 % of failure rate.