Data imputation of water quality parameters through feed-forward neural networks

ABSTRACT The constant monitoring of water quality is fundamental for the understanding of the aquatic environment, yet it demands great financial investments and is susceptible to inconsistencies and missing values. Using a database composed of 59 sampling campaigns, performed for 12 years, on 10 monitoring stations along the Iguassu River Basin (Southern Brazil), this study presents a model, based on feed-forward neural networks, which imputed 1,370 values for 11 traditional water quality parameters, as well as 3 contaminants of emerging concern (caffeine, estradiol and ethinylestradiol). The model validation errors varied from 0.978 mg L-1 and 0.017 mg L-1 for the traditional parameters, for caffeine the validation error was of 0.212 µg L-1 and for the hormones, the errors were of 0.04 µg L-1 (E1) and 0.044 µg L-1 (EE1). The models underwent two techniques to understand the operations performed within the model (isolation and nullification), which were consistent to those explained by natural processes. The results point to the validity of modeling water quality parameters (especially the concentrations of caffeine) through neural networks, which could lead to better resource allocation in environmental monitoring, as well as improving available datasets and valuing previous monitoring efforts.