Seasonal evaluation of a convolutional neural network for coagulation assessment using floc images at operating water...

Floc images may enable faster and more objective interpretation of jar tests, but the reliability of image-based classification under year-round changes in raw-water quality remains uncertain. We developed a convolutional neural network (CNN) to assign floc images to three supernatant-turbidity classes and evaluated it using jar tests conducted twice weekly over ten months at a Japanese water purification plant, where raw-water temperature ranged from 5 to 27°C and turbidity from 2 to 425 NTU, including during two typhoons. A ResNet-50 model trained on background-normalized images collected 4-6 min after coagulant addition achieved 80.1% accuracy per image and 83.2% per jar test in jar-test-grouped five-fold cross-validation of 428 jar tests. Model selection repeated with jar-test-grouped inner validation gave nested estimates of 80.0% and 82.5%, and weekly campaign-blocked cross-validation gave 79.7% and 81.1%. In a preliminary external test at a downstream plant (16 jar tests), image-level accuracy was 68.7-77.8% without retraining. Errors were concentrated at low feed-water turbidity and low temperature, under which image descriptors showed smaller and fainter flocs, and in the intermediate class, in which misclassified jar tests had floc sizes similar to those of the neighboring class. Duplicate jar tests differed by a median of 0.11 NTU, and excluding boundary-ambiguous labels (36% of jar tests) increased jar-test accuracy to 94.9%. The model rejected 89.8% of inadequately coagulated jar tests; the false-acceptance rate of 10.2% fell to 2.1% when the least confident 37% of jar tests were referred to an operator. Image-based assessment is best used as a confidence-gated screening aid alongside the conventional jar test.