Machine Learning Approaches for Predicting Trophic Status in Kotpally Reservoir, Telangana, India Using Limnological and Phytoplankton Indicators

Abstract

Freshwater reservoirs in semi-arid Telangana are critical for irrigation, domestic use, fisheries, and ecosystem services, yet are increasingly threatened by nutrient enrichment from agricultural runoff and other watershed inputs that can accelerate eutrophication and degrade water quality. This study assessed the trophic status of Kotpally Reservoir (Vikarabad District; ~70 km west of Hyderabad) by integrating limnological measurements with phytoplankton indicators and by developing supervised machine learning models to predict trophic categories. Water was sampled monthly for one year at five stations and key physicochemical parameters and phytoplankton diversity were analyzed using standard methods, representing key ecological zones, with field measurements (temperature, pH, dissolved oxygen) and laboratory analyses following standard methods; phytoplankton were collected using a 25 µm net, preserved in 4% formalin, identified microscopically (400×) and enumerated using a Sedgwick–Rafter chamber, and community structure quantified using Shannon diversity, evenness, and richness indices. Trophic status was determined using Carlson’s chlorophyll-a–based Trophic State Index (TSI), and predictive models (Random Forest, Support Vector Machine, Artificial Neural Network) were trained on normalized, quality-controlled data with feature selection emphasizing temperature, dissolved oxygen, nitrate, phosphate, chlorophyll a, and diversity indices, and evaluated using accuracy and related classification metrics. Seasonal dynamics were pronounced: summer temperatures peaked at ~34°C with concomitantly reduced dissolved oxygen, while monsoon conditions produced the highest nitrate and phosphate concentrations, consistent with runoff-driven loading. Forty-five phytoplankton species across five classes were recorded, dominated by Chlorophyceae and Bacillariophyceae, with increased Cyanophyceae abundance during the nutrient-rich monsoon, indicating incipient eutrophic tendencies. TSI results placed the reservoir overall in the mesotrophic-to-eutrophic range, with higher productivity in summer and monsoon. Among models, Random Forest achieved the best performance (92% accuracy) and identified chlorophyll-a as the most influential predictor, supporting integrated, data-driven early-warning and management strategies focused on nutrient-input control and sustained ecological monitoring.