Abstract
Potential source density function (PSDF) is developed to identify, that is, locate and quantify, source areas of ambient trace species based on Gaussian process regression (GPR), a machine-learning technique. The PSDF model requires backward trajectories and sampling data at a receptor site in the calculation as in the conventional model to locate source areas of ambient trace species, such as the potential source contribution function (PSCF). The PSDF model can identify source areas quantitatively and provide information on the reliability of the estimation, while the PSCF model cannot. To verify and evaluate the capability of the PSDF model, tests are carried out using three scenarios based on ambient trajectory analysis data and simulated source distributions. The test results demonstrate that the PSDF model can identify the sources of ambient trace species more accurately than the PSCF model. The PSDF model can quantify the size of the source contaminating the air parcels passing through it, and the model can detect the variation of source intensity. Also, in the test, we evaluate reliability of the information provided by the PSDF model. In addition, future works are recommended to improve the model and increase its applicability.
Highlights
Combination of ambient measurements and models is an essential approach to understand the atmospheric environment
The potential source density function (PSDF) model is developed to identify sources of ambient trace species based on Gaussian process regression (GPR)
Algorithms of PSDF are improved by structured kernel interpolation (SKI)
Summary
Combination of ambient measurements and models is an essential approach to understand the atmospheric environment. We introduce a new model for identifying source areas of ambient trace species, called potential source density function (PSDF). This model can estimate the source distribution, that is, location and intensity, influencing the ambient concentrations at a receptor site based on Gaussian process regression (GPR), a machine-learning technique. Source distribution estimated by the PSDF model helps one understand the intensity and consistency of each area's influence on the concentrations of ambient trace species at a receptor site. The sampled air parcel contains gradually accumulated pollutants collected while traveling along its trajectory ξξii This process of accumulation is modeled as an integral over a source density function, which is denoted here as ff. We propose an efficient method for evaluation of covariances, which is based on the concept of structured kernel interpolation (SKI) (Wilson and Nickisch, 2015)
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
