Authors - Nu Yin Khaing, Win Lelt Lelt Phyu Abstract - Water quality monitoring is essential for sustainable aquaculture management and fish health assessment. However, monitoring a large number of physicochemical parameters in creases sensor deployment costs and system complexity. While traditional machine learning ap proaches struggle with complex, nonlinear relationships among water quality variables, feature reduction can optimize system efficiency. This study proposes a hybrid machine learning and deep learning framework to achieve accurate, cost-effective water quality classification using the Water Quality Dataset (WQD). The framework integrates Random Forest (RF) and XGBoost for feature selection with Support Vector Machine (SVM) and Long Short-Term Memory (LSTM) models as classifiers, evaluating four hybrid combinations (XGBoost+SVM, XGBoost+LSTM, RF+SVM, and RF+LSTM) across subsets of 10, 7, and 5 features. Experimental results demon strate that hybrid deep learning architectures consistently outperform traditional machine learn ing methods. Specifically, XGBoost+LSTM and RF+LSTM achieved the highest classification accuracy of 96.05% using 10 selected features, while maintaining reliable performance at lower dimensions. Furthermore, Shapley Additive explanations (SHAP) analysis was applied to en hance model interpretability, identifying Dissolved Oxygen (DO), Turbidity, BOD, H2S, and Nitrite as the most important attributes. Ultimately, the proposed framework minimizes sensor requirements and provides an accurate, interpretable, and economically viable solution for aqua culture water quality monitoring systems.
Open Zoom