Water Potability Classification by Using Deep Residual Networks with Sum-Product-Difference Feature Matrices

Tufenkci S., ALAGÖZ B. B.

Water Resources Management, vol.40, no.10, 2026 (SCI-Expanded, Scopus)

  • Publication Type: Article / Article
  • Volume: 40 Issue: 10
  • Publication Date: 2026
  • Doi Number: 10.1007/s11269-026-04853-3
  • Journal Name: Water Resources Management
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, ABI/INFORM, CAB Abstracts, Compendex, Environment Index, Geobase, INSPEC, Natural Science Collection (ProQuest), Biological Science Database (ProQuest), Earth, Atmospheric, & Aquatic Science Collection (ProQuest), Materials Science & Engineering Collection (ProQuest), Technology Collection (ProQuest)
  • Keywords: Classification, Deep learning, Potability, Residual neural networks, Water quality
  • Inonu University Affiliated: Yes

Abstract

Due to the large variety of water contaminants, the accurate classification of water quality necessitates extensive chemical analyses and consideration of multi-contaminant effects on water potability. In this perspective, we propose a ResNet-based deep learning model with three-channel feature augmentation that is designed to improve the potability assessment. Conventional chemical and physical analyses produce only a limited number of features for determining the potability. Evaluating them solely based on their standard threshold values may not be sufficient for accurate classification because feature interactions -both linear and nonlinear- can play a role in the potability. To address this, the features are transformed into Gramian-inspired matrices by using Sum, Product and Subtraction operations of features, which are called Gramian Operation Fields (GOF), and this spatial augmentation can represent inter-feature dependencies within the spatial domain. This enables effective processing and capturing relationships among these dependencies within channels of two-dimensional CNN models and enhances the representation of linear and nonlinear inter-feature relationships. To demonstrate the effectiveness of the proposed method, we conducted numerical experiments on the publicly available Water Potability dataset. Experimental results indicated that the ResNet-18 model with 3-channel feature matrices achieved higher performance compared to other machine learning methods.