Rapid estimation of biosludge polymer demand for dewatering via image classification

Applying the optimal polymer dose to biosludge prior to dewatering can help minimize the economic and environmental costs of disposal. However, measuring the polymer demand is a time-intensive process for wastewater treatment plants, and biosludge characteristics vary over time. A novel, rapid image-based polymer demand estimation tool was developed by designing a simple and accessible image capture station and applying transfer learning of an open-source image classifier. Images were captured with a smartphone fixed in an enclosed structure above a biosludge sample dosed with polymer in a 150 mm petri dish. To facilitate collecting the required image database, the entire structure was fastened to an orbital shaker to intermittently apply shake-steps between image capturing that simulate replicates. Using TensorFlow in Python, image databases were constructed and used to retrain EfficientNetV2 to classify images of dosed biosludge as either underdosed, well-dosed, and overdosed. Two image-based models were constructed, one for pulp and paper mill biosludge, and the other for municipal biosludge, which achieved 86 % and 83 % accuracy against an unseen set of images. • New image-based method to predict polymer demand was successfully demonstrated • A 10,000-image library of dosed biosludge images was captured using a smartphone • Transfer learning of open-source image classifier used to predict dosing efficacy • Overall model accuracies were 86 % for pulp and paper mill, 83 % for municipal sludge