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

Area of Science:

• Environmental Science
• Data Science
• Biogeochemistry

Background:

• Greenhouse gas (GHG) monitoring is crucial for climate change mitigation but faces challenges with high-cost, complex equipment.
• Low-cost sensors often exhibit poor accuracy due to signal drift and environmental sensitivity, especially in high-temperature, high-humidity conditions like composting.

Purpose of the Study:

• To develop GHGsNet, a novel, low-cost, process-aware deep learning framework for accurate prediction of CO2, CH4, and N2O emissions.
• To address the limitations of existing low-cost sensors by integrating process parameters and causal drivers into the prediction model.

Main Methods:

• GHGsNet utilizes gas-specific deep learning architectures, integrating low-cost sensor signals with gas-state variables and key process parameters.
• Model interpretability was assessed using SHapley Additive exPlanations-Partial Dependence Plot (SHAP-PDP) and validated with 16S rRNA microbial analysis.
• The framework was compared against a sensor-only baseline (TriGasNetSensor, TGNS) and validated on large-scale composting data.

Main Results:

• GHGsNet significantly improved prediction accuracy for CH4 (R² = 0.9268) and N2O (R² = 0.9310) compared to the sensor-only model.
• Interpretability analysis confirmed that identified drivers align with known biogeochemical pathways of methanogenesis and nitrification-denitrification.
• Independent validation demonstrated strong generalization capabilities, highlighting the importance of process-aware information.

Conclusions:

• GHGsNet provides a robust, interpretable, and low-cost solution for GHG monitoring in challenging composting environments.
• The framework's success underscores the necessity of incorporating process-aware information for accurate emission prediction.
• GHGsNet has potential for broader application in monitoring other process-driven anthropogenic activities.