1Institut für Mathematik, University of Potsdam, Potsdam, Germany
2MPI for Chemistry, Mainz, Germany
3Interdisciplinary Center for Scientific Computing, University of Heidelberg, Heidelberg, Germany
4Institut für Umweltphysik, University of Heidelberg, Heidelberg, Germany
5Computer Science and Artificial Intelligence Lab, Massachusetts Institute of Technology, Cambridge/MA, USA
Abstract. With the increasing availability of observations from different space-borne sensors, the joint analysis of observational data from multiple sources becomes more and more attractive. For such an analysis – oftentimes with little prior knowledge about local and global interactions between the different observational variables available – an explorative data-driven analysis of the remote sensing data may be of particular relevance.
In the present work we used generalized additive models (GAM) in this task, in an exemplary study of spatio-temporal patterns in the tropospheric NO2-distribution derived from GOME satellite observations (1996 to 2001) at global scale. We modelled different temporal trends in the time series of the observed NO2, but focused on identifying correlations between NO2 and local wind fields. Here, our nonparametric modelling approach had several advantages over standard parametric models: While the model-based analysis allowed to test predefined hypotheses (assuming, for example, sinusoidal seasonal trends) only, the GAM allowed to learn functional relations between different observational variables directly from the data. This was of particular interest in the present task, as little was known about relations between the observed NO2 distribution and transport processes by local wind fields, and the formulation of general functional relationships to be tested remained difficult.
We found the observed temporal trends – weekly, seasonal and linear changes – to be in overall good agreement with previous studies and alternative ways of data analysis. However, NO2 observations showed to be affected by wind-dominated processes over several areas, world wide. Here we were able to estimate the extent of areas affected by specific NO2 emission sources, and to highlight likely atmospheric transport pathways. Overall, using a nonparametric model provided favourable means for a rapid inspection of this large spatio-temporal data set,with less bias than parametric approaches, and allowing to visualize dynamical processes of the NO2 distribution at a global scale.