Advancing Microbial Activity Inference from Machine Vision Framework Towards Intelligent Control of Environmental Biotechnology for Wastewater Treatment

Pioneering biotechnologies hold significant promise for enhancing sustainable wastewater treatment. However, their widespread implementation usually hinges critically on achieving stable and effective control. The emerging deep learning offers a reliable approach to achieve the real-time and accurate monitoring of microbial activity and paving way for the intelligent control of environmental biotechnology. In this context, anaerobic ammonia oxidation (anammox) process, a cutting-edge biotechnology in the past decades, was investigated, and an end-to-end machine vision (MV) framework was developed to directly discern the activity states of sludge based on their distinct visual features presented in thousands of images. The vision model could identify activity states of anammox sludge spanning a wide range with a high accuracy of 84%, and the maximum relative error (RE) of prediction was below 3% on independent tests. This microbial activity prediction performance was further validated in a continuous reactor beyond the model development dataset with RE <5%. Gradient-weighted class activation mapping visualization revealed that autonomous selection of key visual features while ignorance of noise information of microbial aggregates contributed to the superior performance of the vision model for high accuracy recognition of microbial activity. Moving forward, the developed MV framework was applied to predict the microbial activity of anammox sludge sourced from various laboratory and engineering scenarios, and high accuracies of 84.74% and 86.73% were achieved, demonstrating a robust generalization capacity. This study offers a conceptual validation of intelligent identification of microbial activity, and open a new avenue for intelligent control of environmental biotechnology for sustainable wastewater treatment.