Transfer learning for quantitative image analysis of biosolids

This study explores how quantitative image analysis can enhance dewatering efficiency of stabilized biosolids at a major industrial wastewater treatment plant in Northern Europe. The aim is to develop a transparent and systematic analysis workflow encompassing data integration from various sources to predict decanter organic solids recovery. During two campaigns, data were collected from operational conditions and laboratory measurements. In addition, data were collected from image analysis and generated by transfer learning techniques using a readily available online database. Partial Least Squares (PLS) and Random Forest (RF) models were tested using different combinations of data sources. The best recovery prediction was obtained using a RF model utilizing both process and laboratory data in combination with transfer learning, improving the prediction by 14% as compared to baseline prediction (using average values). In addition, clustering of segmented particle images and RF-based recovery prediction revealed a strong dependency on specific crystalline particles. In general, the RF model outperformed the PLS model in predicting recovery, although both models lack consistency in prediction across the organic solids concentration range. This study offers operators insight into factors affecting dewatering efficiency and provides a diagnostic workflow transferable to other systems with heterogeneous particle mixtures.