Optimizing Air Quality Sensor Networks Using Gaussian Process Regression and Global Optimization Techniques

Despite recent advancements in technology and purification techniques, industrial pollution continues to pose significant challenges in terms of human exposure and monitoring of air quality close to sources. Optimizing air quality networks and integrating them with advanced spatiotemporal statistical methods is thus essential for effective monitoring of atmospheric contamination.

This study addresses the problem of optimizing the placement of sensors for measuring air pollution at urban and regional scales. Several global optimization techniques, including the GlobalSearch Algorithm, Genetic Algorithm, and Particle Swarm Optimization, are applied to this problem.

Two interpolation methods are used to estimate contamination levels at control points: the standard triangulation-based Natural Neighbour interpolation method for scattered data and Gaussian Process Regression (GPR), which employs covariances derived from a dynamic pollution transfer model. The GPR technique is particularly suitable for simulating smoke-like, narrowly directed industrial pollution at distances of less than a few tens of kilometres from the source.

Numerical experiments were conducted using two pollution datasets: aerosol (PM10) concentrations simulated by the ADMS model for the Dunkirk region in northern France and sulphur dioxide (SO2) concentrations simulated by the CALPUFF model for the Dnipropetrovsk region in Ukraine. The first dataset involves diffuse pollution from multiple anthropogenic and natural sources, while the second involves emissions from industrial point sources.

Optimal sensor placements are identified, and estimation errors are evaluated for the interpolation methods and datasets. The described method could allow the construction of effective air quality networks for different types of atmospheric pollution and provide a means to estimate their effectiveness.