Deep Residual Networks Classify Water Potability Using Sum-Product-Difference Feature Matrices

Researchers in Türkiye have devised a way to turn routine water chemistry readings into images, allowing a computer vision network to judge whether water is safe to drink. The approach, described by Sevilay Tufenkci of Malatya Turgut Ozal University and Baris Baykant Alagoz of Inonu University, transforms the handful of numbers produced by standard water analyses into three two-dimensional feature matrices, which are then fed to a deep residual network. Their experiments on a publicly available water potability dataset showed that the resulting ResNet-18 model achieved higher classification performance than other machine learning methods tested for comparison.

The problem the two computer engineers set out to address is one that water utilities and environmental agencies know well. Assessing whether water is potable is not simply a matter of checking each measured parameter against a regulatory threshold. Water quality is determined by a wide variety of contaminants, and the combined, multi-contaminant effects on potability can be subtle. Conventional chemical and physical analyses produce only a limited number of features, such as measurements of dissolved substances and physical properties, and evaluating these features solely by their standard threshold values may not be sufficient for accurate classification. Interactions among the features, both linear and nonlinear, can play a decisive role in whether a sample is safe to drink. Two individually acceptable readings may together signal contamination; conversely, a parameter that slightly exceeds its threshold may be harmless in the presence of other favorable measurements.

Capturing those interactions with classical machine learning is difficult because tabular data with few features offers the algorithms little to work with. The researchers’ insight was to borrow a technique from the deep learning community that has proven successful in other domains: representing non-spatial data as images so that convolutional neural networks, which excel at extracting spatial patterns, can be applied to it. This family of techniques, inspired by Gramian angular fields and related encodings that have been used to convert time series into images for visual inspection and classification, provides the mathematical foundation for the new method.

In the study, the authors transform the feature values of each water sample into Gramian-inspired matrices using three arithmetic operations: the sum of features, the product of features and the subtraction of features. They call the resulting structures Gramian Operation Fields, or GOF. Each operation yields a different two-dimensional field in which every cell encodes a relationship between a pair of features. The sum field highlights additive dependencies, the product field captures multiplicative, or nonlinear interactive, effects, and the difference field represents contrasts between parameters. Stacking the three fields as separate channels produces a three-channel image-like representation of a single water sample, in much the same way that a natural color image consists of red, green and blue channels.

This spatial augmentation is more than a cosmetic trick. By encoding inter-feature dependencies within the spatial domain of an image, the method allows the convolutional layers of a two-dimensional CNN to process and capture relationships among the features directly. Convolutional kernels sweep across the GOF matrices and learn to recognize local patterns that correspond to meaningful combinations of water quality parameters. The representation thereby enhances the expression of both linear and nonlinear inter-feature relationships, which are precisely the effects that simple threshold-based assessment tends to miss. The authors note that the technique builds on ideas from bilinear CNN models for fine-grained visual recognition and from architectures such as factorization machines and deep cross networks, which were developed to model feature interactions in tabular and recommendation settings.

For the classification backbone, the researchers selected a deep residual network, ResNet-18, a convolutional architecture originally developed for image recognition. Residual networks are particularly well suited to this task because their skip connections allow information to flow across layers unchanged, easing the training of deep models and improving generalization. The network processes the three-channel GOF representation of each water sample much as it would a photograph, learning a hierarchy of features that culminates in a potability decision. The choice of ResNet-18, a relatively compact member of the residual network family, also reflects a practical consideration: the model is small enough to be trained efficiently while still being deep enough to capture complex patterns in the augmented feature space.

The team evaluated their approach through numerical experiments on the publicly available Water Potability dataset, a widely used benchmark containing measurements of standard water quality parameters labeled for potability. They compared the ResNet-18 model operating on three-channel GOF inputs against other machine learning methods applied to the same data. The results indicated that the ResNet-18 model with three-channel feature matrices achieved higher performance than the alternatives, supporting the central claim of the study: that augmenting limited water quality features into spatial representations and exploiting the pattern-recognition strengths of convolutional networks leads to more accurate potability classification than conventional approaches.

The significance of the work lies in how it reframes a data-poor problem. Water quality monitoring routinely produces measurements of a modest set of chemical and physical parameters, and datasets built from such measurements are often small and imbalanced, since safe samples outnumber unsafe ones in real monitoring programs. Machine learning models applied directly to such tabular data frequently struggle to outperform simple baselines. By contrast, the GOF augmentation multiplies the information available to the classifier without requiring any additional laboratory analysis: every pairwise interaction among the measured parameters becomes an explicit, spatially organized signal that the network can learn from. In effect, the method lets a nine-parameter water report carry the representational richness of a much larger feature set.

The approach also fits into a broader trend in environmental data science. In recent years, researchers have encoded time series as images for anomaly detection in manufacturing, mapped sensor signals into visual domains for classification, and applied convolutional networks to a growing list of non-spatial problems. The potability study extends this strategy to environmental monitoring, where the underlying data are not sequences but vectors of chemical readings. Related efforts in the literature include hybrid approaches that integrate machine learning with classical analytical methods for water quality monitoring, ensemble and stacking models for stream and groundwater prediction, and explainable AI frameworks for potability prediction. What distinguishes the new work is the specific three-operation construction of the feature matrices, which the authors argue captures a richer set of dependencies than single-operation encodings.

Practical applications could follow in several directions. Automated potability assessment could support treatment plants in screening incoming water, help regulators prioritize samples for full laboratory analysis, and enable faster responses to contamination events when combined with real-time sensor networks. Because the method operates on the same parameters that utilities already measure, adopting it would not require new instrumentation, only a software layer that computes the GOF matrices and runs a trained residual network. The authors also suggest that the framework could generalize beyond potability, since the underlying idea of spatially augmenting feature interactions applies to any classification problem where limited tabular features conceal important nonlinear relationships.

The study, published in the journal Water Resources Management, is the work of Sevilay Tufenkci of the Department of Computer Engineering at Malatya Turgut Ozal University, who performed the material preparation, data collection and analysis, together with Baris Baykant Alagoz of the Department of Computer Engineering at Inonu University. Both authors contributed to the conception and design of the study. The researchers received no external funding for the work, and they note that the datasets generated and analyzed during the study are available from them upon reasonable request. As water scarcity and contamination place growing pressure on drinking water systems worldwide, tools that squeeze more insight from routine measurements, and that can flag the subtle multi-contaminant interactions that fixed thresholds overlook, are likely to attract increasing attention from both the research community and the water industry.