Freshwater ecosystems are increasingly affected by urbanization, industrial activity, agricultural runoff, and climate change, particularly in transboundary river systems where pollution propagates across national borders. The Kura, Araz, Ganikh, and Okchuchay rivers are key water resources in the South Caucasus region, supporting biodiversity, agriculture, and water supply. However, these systems are under significant anthropogenic pressure, leading to deterioration of physicochemical and microbiological water quality. Traditional monitoring methods are limited in capturing nonlinear dynamics and predicting future pollution trends, creating a need for data-driven approaches.
This study investigates the physicochemical and microbiological status of these rivers and develops machine learning models for water quality prediction. Seasonal sampling was conducted to measure temperature, pH, dissolved oxygen (DO), nitrate, nitrite, ammonium, phosphorus, heavy metals, and microbial communities, including saprotrophic, pathogenic, oil- and phenol-degrading, and nitrogen-cycle bacteria. Data preprocessing included normalization and feature selection based on physicochemical and nutrient variables. Random Forest, XGBoost, and Artificial Neural Network models were implemented and evaluated using R², RMSE, and MAE.
The results showed that the Kura and Ganikh rivers are highly polluted, with saprotrophic bacteria exceeding ecological thresholds and several parameters surpassing TLVs. In the Okchuchay River, untreated mining effluents caused severe contamination, with nickel concentrations 5–7 times, iron 3–4 times, and copper and molybdenum about two times above permissible limits. These conditions have significantly degraded aquatic ecosystems and reduced fish populations, particularly trout.
Among the tested models, XGBoost achieved the highest predictive performance, outperforming Random Forest and ANN by better capturing nonlinear relationships and spatiotemporal variability in water quality data. The findings demonstrate that integrating environmental monitoring with machine learning provides an effective framework for predicting river pollution and supports early warning systems and sustainable management of transboundary water resources.