Deep learning-based morphology classification of activated sludge flocs in wastewater treatment plants

Abstract

Microscopy inspection of the morphology of activated sludge (AS) flocs can provide important information regarding the AS properties, which strongly affect the performance of AS systems. However, the acquisition of such information from microscopy inspection results requires skilled and specialized expertise. In this study, we aimed to develop two deep learning-based two-label classifiers for recognizing aggregated or dispersed flocs (classifier-1) and the presence or absence of filamentous bacteria (classifier-2). To achieve this, we used a convolutional neural network (CNN)-based method and selected the pre-trained Inception v3 as the CNN architecture. We developed an automatic microscopy image acquisition system, enabling us to obtain 154 images for 7 min. Over 12 000 images of aggregated and dispersed flocs were obtained from wastewater treatment plant (WWTP)-S and -E over 15 weeks. Classifier-1 was retrained using these images. Classifier-1 distinguished the aggregated and dispersed flocs with a training accuracy of approximately 95% and recognized a 20% morphological change in the aggregated flocs. Classifier-1 also recognized the morphology of AS flocs obtained from other WWTPs, the AS from which was used for retraining. Classifier-2 quantitatively recognized an abundance of filamentous bacteria in the AS flocs. These results clearly indicated that the developed image classification model could serve as a useful warning system for the settleability deterioration and abundance of filamentous bacteria in the aeration tank of a full-scale AS system.