A Hybrid Deep Learning–SVM Framework for Water Quality Prediction Using CNN Feature Extraction and Data Balancing

Abstract

Prediction of water quality (WQP) and the safety of drinking water is one of the most difficult tasks given the intricate interrelationship of physicochemical and biological factors, the imbalance in the data, and the other features within the dataset. The proposed approach fuses deep learning techniques and machine learning to establish a hybrid method. In particular, the input features are extracted using convolutional neural networks (CNNs) from the input data with 64 features first, followed by the classification using an RBF kernel as the input on support vector machine (SVM). Pre-processing steps include data cleaning through imputation of both moving mean and global mean. The data is addressed when the data is imbalanced using SMOTE method. This reduces the imbalanced distribution from 7:1 to nearly 1:1 after the algorithm’s operation. Then the predictions made are checked for success with a pair of benchmarks, where the general prediction of water quality (WQP) has 20 predictors as well as the predicted potable water has a total of nine predictors, respectively. For comparison, SMOTE-based techniques greatly enhance baseline algorithms effectiveness. Our proposed hybrid technique applied as combination of CNN-RF and CNN-SVM techniques performs better, which is evident from (the WQP dataset), whose accuracy is 96.7% and 97.25%, while the accuracy of the water potability dataset using the same techniques is 82% and 82.87%.