Estimativa da cor verdadeira da água em soluções padrão Pt-Co por meio de redes neurais convolucionais

Abstract

True water color is an important parameter for monitoring water quality and controlling industrial processes. In this context, the present study aimed to develop and evaluate a methodology based on Convolutional Neural Networks (CNNs) for the classification of true water color from digital images, using the Platinum-Cobalt (Pt-Co) scale as reference. Standard solutions ranging from 0 to 500 mg Pt-Co·L⁻¹ were prepared in the laboratory, and an image acquisition chamber with controlled lighting conditions was constructed. The images were captured using a smartphone and used to build a database containing 8,248 images distributed across 36 different concentration classes. The neural network model was developed in Python using the TensorFlow and Keras ibraries and evaluated using an independent test set through accuracy, loss function, confusion matrix, F1- score, Cohen's Kappa coefficient, and Mean Absolute Error (MAE). The neural network achieved 67.18% accuracy on the test set, demonstrating its ability to identify visual patterns associated with different Pt-Co scale concentrations. The results showed that classification errors occurred predominantly between adjacent concentration classes, highlighting the inherent difficulty in distinguishing slight color variations. Overall, the proposed methodology demonstrated the feasibility of combining computer vision and deep learning techniques for true water color classification under controlled conditions, indicating its potential for future applications in environmental monitoring and water quality control.