Image processing and machine learning approach to estimate turbidity in raw water for treatment plants in Sri Lanka

Abstract

Conventional turbidity monitoring in water treatment plants requires frequent manual sampling and laboratory-based measurements, which can limit the ability to achieve continuous and costeffective monitoring. This research developed an image-based machine learning approach for estimating raw water turbidity using digital images, evaluated as a proof-of-concept alternative monitoring method. A total of 200 raw water samples were collected from the Ambatale Water Treatment Plant laboratory, and corresponding digital images were acquired using a DSLR camera under controlled conditions. The measured turbidity ranged between 6.59 and 10.73 nephelometric turbidity units (NTU). Convolutional neural network (CNN) models were developed and evaluated using image data paired with laboratory-measured turbidity values. The results demonstrated that the improved augmented model achieved the highest prediction performance with a root mean square error (RMSE) of 0.51 and the lowest mean absolute error (MAE) of 0.374, a coefficient of determination of 0.517 as well as a Pearson correlation of 0.740, showing a moderate-to-strong agreement with measured turbidity values while reducing estimation errors compared with the initial models. Furthermore, a conceptual hardware prototype and software monitoring dashboard were developed to evaluate the practical implementation potential of the proposed approach. The study concludes that image-based machine learning can serve as a potential supporting tool for real-time turbidity monitoring in water treatment facilities; however, given the narrow turbidity range and single-plant dataset, these results can only validate a successful early proof of concept; broader datasets must be developed and tested using different plants across wider operational turbidity ranges for islandwide real-time deployment.