Wastewater quality prediction using real and CGAN-augmented spectral data

This study presents a data-driven methodology for estimating COD, a key wastewater quality indicator, using UV–vis spectroscopy and machine learning. To address the data scarcity present in highly polluted industrial contexts such as tanneries, this study introduces a data augmentation strategy based on CGAN. This model generates synthetic absorbance spectra conditioned on specific COD values, effectively expanding the training set with chemically plausible data. The augmented dataset is used to train regression models, a Multi-Layer Perceptron (MLP) among others, yielding improvements in COD prediction. The quality of synthetic spectra is validated through statistical similarity metrics, Principal Component Analysis, and correlation profiling. Experiments on real wastewater samples from the tanning industry show that the regression models trained on synthetic data generated by CGANs outperform models trained only on real data. These findings underscore the potential of generative modeling in enhancing soft sensing systems for environmental monitoring.