Access to safe drinking water remains a critical public health challenge, with millions of people worldwide relying on contaminated sources that cause severe illnesses. Conventional water quality testing, while accurate, is often slow, labor-intensive, and costly, limiting large-scale implementation. This study presents a multi-modal deep learning framework that integrates structured water quality features—such as pH, hardness, turbidity, sulfate, and chloramine levels—with synthetic hyperspectral image representations to classify water as potable or non-potable. The methodology includes preprocessing steps such as missing value imputation, normalization, dimensionality reduction using PCA, and class balancing through SMOTE. Each sample is converted into a 32 × 32 × 10 grayscale image cube, which is processed by a 3D Convolutional Neural Network (CNN) to capture both spectral and spatial patterns. Among six CNN variants evaluated, the best-performing model achieved a validation accuracy of 72%, demonstrating improved generalization compared to traditional machine learning baselines. These results highlight the potential of combining structured and spectral–spatial data, offering a cost-effective and scalable approach for automated water quality monitoring.