Process-Aware Deep Learning for Low-Cost Greenhouse Gas Sensing: Insights from Composting toward Scalable...

Greenhouse gases (GHGs) monitoring is essential for mitigating emissions from anthropogenic activities, yet the deployment of high-precision analyzers remains constrained by cost and operational complexity. Low-cost GHGs sensors often suffer from signal drift, cross-sensitivity, and unstable responses under extreme temperature and humidity conditions. To address these challenges, GHGsNet was developed as a low-cost, process-aware deep learning framework for predicting CO2, CH4, and N2O emissions during composting, a representative high temperature and high humidity anthropogenic activity. GHGsNet integrates low-cost gas sensor signals with gas-state variables and key process parameters using gas-specific deep learning architectures. Compared with a sensor-only baseline model (TriGasNetSensor, TGNS), GHGsNet significantly improved prediction accuracy for CH4 (R2 = 0.9268) and N2O (R2 = 0.9310) by embedding causal emission drivers rather than relying on intergas correlations only. Model interpretability analysis based on SHapley Additive exPlanations-Partial Dependence Plot (SHAP-PDP), together with 16S rRNA microbial evidence, demonstrated that the identified drivers are consistent with established biogeochemical pathways governing methanogenesis and nitrification-denitrification. Independent validation using large-scale composting data confirmed strong generalization, while cross-domain tests of TGNS highlighted the necessity of incorporating process-aware information. Overall, GHGsNet provides a robust and interpretable framework for low-cost GHGs monitoring in high-temperature and high-humidity composting systems, with potential for future extension toward other process-driven anthropogenic activities.