Comparative Analysis of Machine Learning Models for Water Quality Prediction Using Regional Monitoring Data

Abstract

This study investigates the comparative performance of four labelical machine learning algorithms—Decision Tree, Support Vector Machine (SVM), Random Forest, and Neural Network—on water quality prediction tasks using a dataset comprising 1,000 real-time sensor data points from five distinct geographic regions. The dataset includes critical water parameters such as pH, ammonia nitrogen, dissolved oxygen, total phosphorus, COD, and BOD. Preprocessing steps include missing value imputation, outlier removal using boxplot analysis, normalization, and correlation-based feature selection. Each model is tuned through grid search for optimal performance. Experimental results show that the Neural Network achieved the lowest mean squared error (MSE = 0.047) and highest coefficient of determination (R² = 0.976), outperforming the other models. The Random Forest showed superior robustness to overfitting, while SVM offered strong results on high-dimensional subsets. Decision Trees, although less accurate (MSE = 0.130), provided high interpretability. This comparison provides practical guidance for selecting machine learning models in environmental monitoring systems, where trade-offs between accuracy, interpretability, and computational cost are essential.