Linking fluorescence spectral to machine learning predicts the emissions fates of greenhouse gas during composting

Abstract

Dissolved organic matter plays a complex and crucial role in regulating microbial activity and greenhouse gas (GHG) emissions. However, the relationship between dissolved organic matter and GHG emissions to enable intelligent prediction remains limited. Therefore, the variations in GHG emissions and dissolved organic matter characteristics were assessed across different composting scenarios in this study, including various raw materials, auxiliary materials, and composting processes. After that, three machine learning models of varying depths—Gradient Boosting Regression, Random Forest, and Deep Neural Network—were established based on dissolved organic matter characteristics to accurately predict the dynamics of GHG emissions during composting. The results indicated that the Deep Neural Network model performed best in predicting CH₄ emissions (R² = 0.96), while the Random Forest model excelled in predicting N₂O and CO₂ emissions (R² = 0.93 and R² = 0.76, respectively). Meantime, further feature analysis revealed that soluble microbial by-products in raw materials, the degree of organic matter degradation, and microbial activity are crucial factors influencing the emissions of CH₄, CO₂ and N₂O, respectively. This study successfully achieved accurate predictions of GHG emissions, identified key dissolved organic matter components driving gas emissions, offered a new perspective for future research on GHG dynamics, and provided scientific guidance for GHG management during composting.