An Attention-Augmented Inception Architecture for Image-Based Water Quality Prediction

Water quality assessment is critical for ensuring safe drinking water and sustainable aquatic ecosystems. Conventional laboratory-based techniques are accurate but time-consuming, expensive, and unsuitable for real-time monitoring. Existing image-processing-based methods often fail to capture complex spatial–spectral dependencies in water surface images, limiting prediction accuracy for parameters such as pH, turbidity, and dissolved oxygen. We propose AttnInceptionNet, a deep learning model integrating Inception modules with multi-head self-attention to extract multi-scale spatial features and selectively emphasize informative regions in water images. Preprocessing involves contrast enhancement, noise reduction, and region-of-interest (ROI) extraction. The model is trained on a dataset of annotated water images with ground-truth physicochemical measurements, using Adam optimizer and early stopping. AttnInceptionNet achieved 96.8% accuracy in water quality classification and outperformed three benchmark models: InceptionV3, ResNet50, and DenseNet121 by margins of 3.4%, 4.2%, and 5.0% respectively. The attention mechanism improved feature discrimination, particularly in images with reflections or low illumination.