Enhancing Water Quality Early Warning System Accuracy in Pangasius Aquaculture Using Machine Learning

Abstract

Intensive catfish (Pangasius sp.) aquaculture faces significant economic risks driven by mass mortality events linked to unstable water quality, particularly toxic ammonia spikes and pH fluctuations. Although Internet of Things (IoT) technology enables real-time monitoring, the resulting time-series data presents complex challenges, including high sensor noise, asynchronous transmission, and severe class imbalance, which compromise standard reactive monitoring methods. This study aims to enhance diagnostic accuracy by comparing Support Vector Machine (SVM), Random Forest (RF), and XGBoost algorithms to construct a robust Early Warning System (EWS). A quantitative experimental methodology was applied to real-world sensor data, with temporal aggregation preprocessing to reduce noise. To ensure rigorous validation simulating real-world deployment, the dataset utilized a strict chronological split (80% training, 20% testing) and was further tested using 5-Fold Time-Series Cross-Validation. The results demonstrated the definitive superiority of ensemble-based models; Random Forest and XGBoost achieved 100.00% accuracy on the test set, successfully eliminating the critical false negatives exhibited by the SVM model (99.80%). Stability analysis further confirmed the robustness of Random Forest (98.35%) and XGBoost (98.32%) compared to SVM (97.02%). Additionally, feature importance analysis unequivocally identified ammonia as the dominant predictor of critical conditions. Crucially, the study detected a “concept drift” phenomenon in which “Safe” conditions disappeared during the final cultivation phase. These findings conclude that ensemble models provide the optimal architecture for EWS. However, the presence of concept drift necessitates adaptive retraining strategies to ensure long-term reliability in dynamic pond environments.