Cross-resolution learning for scalable detection of filamentous bacteria in activated sludge

Accurate detection and classification of filamentous and floc-forming bacteria in microscopic images provide a comprehensive understanding of sludge settleability and reliable prediction of sludge bulking events in activated sludge systems. Recent studies demonstrated artificial intelligence (AI)’s potential in processing microscopic images, depending on image resolution. Specifically, higher resolutions tend to capture finer morphological details of microorganisms but substantially increase computational cost and machine training effort for pixel-level annotations, particularly when distinguishing bacteria type. In this study, we proposed a novel AI-based strategy where convolutional neural network (CNN) models were trained on low-resolution images (396 × 266 and 788 × 530 pixels) and then tested to predict bacterial populations in higher-resolution images (1544 × 1038 and 3088 × 2076 pixels). The results indicated that models trained with low-resolution images could effectively predict higher resolution images and capture fine details not otherwise visible in the training inputs. This approach significantly improved filament detection and was validated monitoring filament length and floc area. Additionally, computational cost analysis revealed an optimal combination of training/testing resolutions based on the expected number of testing images. Overall, the proposed framework substantially reduced computation and annotation demands while maintaining reliable classification accuracy. It also provided an efficient and scalable solution for large-scale bacterial monitoring in wastewater treatment systems.