Prediction of compost physicochemical properties using image-based feature extraction and machine learning

This study evaluated standard digital photographs, combined with automated image segmentation and machine learning, to rapidly and non-destructively predict compost physicochemical properties. A total of 230 compost samples were collected from retail stores and commercial composting facilities across Georgia, South Carolina, and Tennessee, USA. Images were acquired using an iPhone 13 Pro Max under controlled lighting conditions, and compost regions were automatically isolated using a U-Net convolutional neural network. Image-derived colour, texture, and morphological features were then used to predict pH, electrical conductivity (EC), volatile solids (VS), ash content, and bulk density (BD). Ten regression algorithms were compared; K-nearest neighbours (KNN) regression provided optimal predictive performance. Internal validation of predictive model performance was strong: EC (R 2 = 0.97, RMSE = 490 µS cm −1 ), pH (R 2 = 0.95, RMSE = 0.21), VS (R 2 = 0.92, RMSE = 5.5%), ash content (R 2 = 0.94, RMSE = 23.0%), and BD (R 2 = 0.87, RMSE = 26.2 kg m −3 ). External independent blind validation performance via KNN models remained encouraging: EC (R 2 = 0.69, RMSE = 134 µS cm −1 ), pH (R 2 = 0.75, RMSE = 0.27), VS (R 2 = 0.81, RMSE = 11.5%), ash content (R 2 = 0.70, RMSE = 14.4%), and BD (R 2 = 0.63, RMSE = 96.2 kg m −3 ). Smartphone-based imaging with U-Net segmentation and KNN regression, can provide a rapid and low-cost screening approach for selected, visually linked compost-quality indicators. These findings encourage broader validation across diverse feedstocks, composting systems, and conditions in support of field-scale deployment.