IHSS&IWA26 / BRNO / CZECHIA / 23–28 August 2026 BOOK OF ABSTRACTS 36 Analysis and Characterization Monday, 24 August 2026 / Hall B+C SL12 Fluorescence-Based Machine Learning for Compost Maturity Prediction and Identification of Key Humification Mechanisms Gyumin Kim1, Han Saem Lee2, Hyun Sang Shin1,* 1 Department of Environmental Engineering, Seoul University of Science and Technology, Seoul 01811, Republic of Korea, rbals9677@seoultech.ac.kr 2 Department of Mechanical and Manufacturing Engineering, University of Calgary, Calgary, Alberta T2N 1N4, Canada * Corresponding author: Hyun Sang Shin (hyuns@seoultech.ac.kr) Rapid and reliable assessment of compost maturity is essential for ensuring compost quality and safe agricultural use. Conventional maturity indicators, such as the germination index (GI) and C/N ratio, are often timeconsuming, labor-intensive, and sensitive to analytical conditions, limiting their applicability for routine monitoring [1]. Fluorescence spectroscopy has been widely investigated as a rapid tool for compost maturity assessment. However, quantitative interpretation remains challenging because compost comprises complex mixtures of organic compounds with substantial spectral overlap [2]. Although multivariate approaches such as principal component analysis (PCA) and parallel factor analysis (PARAFAC) have been applied to address this issue, their performance can be affected by variations in composting conditions, feedstock characteristics, and experimental environments [3]. In this study, we developed a fluorescence-based machine learning approach for predicting compost maturity from three-dimensional excitation-emission matrix (3D-EEM) data. A total of 325 EEM-GI datasets were compiled from 33 published studies covering multiple maturity stages of livestock-manure composting. To improve comparability among literature-derived datasets, all EEM images were subjected to a standardized preprocessing workflow, including resizing, denoising, normalization to a 0-1 range, and conversion into numerical matrices for model input. An Extremely Randomized Trees (Extra Tree, ET) model was constructed using GI as the target variable, and model performance was evaluated using a hold-out validation strategy with training, validation, and test sets. The ET model showed strong predictive performance for GI estimation, with a coefficient of determination of R2 = 0.828 on the test set. Furthermore, occlusion heatmap analysis was performed to identify excitationemission regions that contributed most strongly to GI prediction, providing interpretable insights into fluorescence features associated with compost
RkJQdWJsaXNoZXIy NDA4Mjc=