Beach oriented models for estimating bathing water quality

Traditional monitoring of bathing water quality is based on the detection and enumeration of faecal indicator bacteria (FIB). The process is both time-consuming and costly and also provides information with a delay, limiting the ability to take timely action to protect public health. To address this challenge, we designed and conducted an intensive study and developed a predictive model that estimates bathing water quality using data on environmental and meteorological parameters, providing a faster and more cost-effective alternative and support to traditional monitoring. The study focused on three official bathing sites, with data showing an imbalance between acceptable and unacceptable water quality at two of the three sites. Despite these imbalances, the predictive model achieved acceptable overall performances. The first site, Gojača (VK1), yielded the best performance, with a random forest algorithm trained solely on the site’s dataset achieving an accuracy of 0.73, f1 score 0.76 and informdness of 0.45. In contrast, the second site, Kamp (VK2), achieved the worst performance. For Kamp (VK2), the best results were obtained using a Logistic Regression model trained on aggregated data from all sites which achieved an accuracy of 0.65, f1 score of 0.22 and informdness of 0.02. Further research is needed to refine the model and improve prediction accuracy. The results showed meteorological data combined with measurements of salinity and turbidity can serve as a useful proxy for predicting bathing water quality in real time. These results demonstrate that site-specific modeling, combined with frequent environmental monitoring, can enhance real-time prediction of bathing water quality, potentially reducing reliance on traditional methods and enabling faster responses to protect public health. • Site-specific models significantly improve the accuracy of real time water quality predictions. • Machine learning based on intensive monitoring leads to the accuracy up to 75% • Meteorological data, salinity and turbidity improve performances of predictive model up to 25%