Abstract

The growing prevalence of air pollution poses significant risks to human health and ecological stability. Conventional air quality monitoring systems, while accurate, are expensive and geographically limited, restricting their deployment in large-scale sensing networks. Recent advancements in Complementary Metal-Oxide Semiconductor (CMOS) sensor technologies offer a promising pathway for developing cost-effective and miniaturized air monitoring platforms. However, these sensors often face limitations in calibration stability, data drift, and environmental noise interference, which compromise the reliability of pollutant concentration measurements. The major challenge lies in enhancing the accuracy and spatial scalability of low-cost CMOS-based air pollution sensors. Traditional machine learning models fail to capture the complex spatial-temporal dependencies between sensing nodes and environmental factors such as humidity, temperature, and wind dispersion patterns. This study proposes a Graph Neural Network (GNN)-enhanced environmental sensing framework that integrates CMOS-based gas and particulate matter sensors with a distributed graph learning model. The GNN architecture models inter-node relationships and spatial correlations across sensor networks, allowing real-time inference and adaptive recalibration. Data collected from multiple low-cost sensor nodes were processed through graph convolutional layers to estimate pollutant levels (PM2.5, NO2, CO, and O3) with high precision. The system was implemented on a resource-efficient embedded platform to ensure scalability and low energy consumption. The proposed framework demonstrates high predictive accuracy, achieving a Mean Absolute Error (MAE) of 3.2 µg/m³ for PM2.5, Root Mean Squared Error (RMSE) of 4.2, and R² of 0.93, significantly outperforming Random Forest, CNN regression, and Graph Attention Network baselines. The Calibration Drift Reduction (CDR) reached 42%, validating the effectiveness of adaptive recalibration. Computational efficiency remained within 30 ms per node, ensuring feasibility for real-time, large-scale deployment. The results confirm that moderate graph correlation weights (0.4–0.5) and EMA smoothing coefficient of 0.7 provide optimal performance, which shows the robustness, reliability, and scalability of the proposed GNN-enhanced CMOS sensor network for urban air quality monitoring.

Authors

B. Guruprakash¹, Prithviraj Singh Chouhan²
Sethu Institute of Technology, India¹, Medicaps University, India²

Keywords

Air Pollution Monitoring, CMOS Sensors, Graph Neural Networks, Environmental Sensing, Smart Cities