Author(s): Mario Suaza-Medina; Javier Lacasta; Sergio Martin-Segura; F.javier Zarazaga-Soria

Linked Author(s):

Keywords: Artificial intelligence; Eutrophication; Remote sensing

Abstract: This work presents a satellite-based early warning system for eutrophication in large reservoirs used for drinking water, recreation, and biodiversity protection. Current monitoring still depends mainly on in-situ sampling of nutrient and cyanobacteria concentrations, accurate yet resource-intensive and slow, often catching problems too late. The study first determines which Sentinel-2 spectral bands are most sensitive to eutrophication. Based on historical reports from the Ebro River Basin Authority in Spain, a reservoir with alternating eutrophic and non-eutrophic years was chosen to detect signal differences. Bands B09 and B11 proved the most discriminative. To train a predictive model, a far larger labelled dataset was needed. An open U.S. dataset containing over 13,700 in situ cyanobacteria measurements since 2017 was therefore used. Several Machine Learning models were trained with Sentinel-2 data as predictors. GradientBoost, AdaBoost, and XGBoost delivered the best performance, with a weighted average of F1 score of ~0.71 despite class imbalance. The results confirm the feasibility of issuing early alerts so authorities can deploy field teams proactively, laying the basis for scalable, cost-efficient eutrophication management tools.

DOI: https://doi.org/10.64697/iahr.proc.hic2026.8

Year: 2026