Data fusion

Our objective is to examine, test and apply the combination of in-situ measurements, chemistry transport model results, and Earth Observation (satellite) data using data fusion, namely residual kriging (Denby et al., 2008) resp. regression – interpolation – merging mapping at different time steps.

A major obstacle in the use of the satellite data is a frequent occurrence of spatial gaps. In order to fill spatial data gaps, we use an algorithm employed in “gapfill” package (Gerber et al., 2016). This algorithm uses available surrounding spatial and temporal data for the estimation of missing data.

Data_fusion_Fig1

The picture above shows NO2 satellite data in consequent days and years prior (left) and next to (right) the processing of the gapfill algorithm. (Data source: OMI resp. OMNO2d.)

SO2 satellite data show larger spatial gaps compared to NO2. Therefore we decided to merge data of different source (i.e. GOME2 and OMI resp. OMSO2), in order to get product including more valid data within the domain. Then we apply the gapfill algorithm on the merged data.

The gapfilled satellite data can be further used in the data fusion mapping. Mapping methodology without and with the use of the satellite data has been compared in several variants, using cross-validation. Under the cross-validation, the data fusion estimate is calculated for each in-situ measurement point from all available information except from that point. This procedure is repeated for all measurement points in the available set. With help of statistical indicators (like RMSE and bias) the quality of the predictions is demonstrated objectively. The picture below shows comparison by RMSE (left) and bias (right) between NO2 daily European-wide maps for September 2014, created without (red) and with (blue) the satellite date.

Data_fusion_Fig2

The last pictures show the mapping results for NO2 (left) and SO2 (right) daily data for 25. 2. 2014, for the Czech Republic, without (top) and with (bottom) the use of the satellite data.

Data_fusion_Fig3

References:

Gerber F, Furrer R, Schaepman-strub G, Jong R de, Schaepman ME, 2016. Predicting missing values in spatio-temporal satellite data.

Denby B, Schaap M, Segers A, Builtjes P, Horálek J, 2008. Comparison of two data assimilation methods for assessing PM10 exceedances on the European scale. Atmospheric Environment 42, 7122–7134.