22 Overview of the Change in NO 2 Assessment Maps During the Last …
141
network of the Flanders Environment Agency (VMM). The other were low-resolution
model data (15 × 15 km
2 ) provided by a chemical transport model (CTM) BelEUROS. The raw CTM model data was (still is) quite poor. This was partly due to the
fact that the spatial scale of the CTM model is not adequate to represent the high
degree of spatial variability in a densely populated area such as Flanders. The model
also exhibited time- and spatial-dependent biases. Therefore, the first course of action
was to improve the quality of the background concentration map.
22.3 Introducing a Better Background Map
In order to improve the quality of the low-resolution background map, and in order
to improve the resolution, an hourly land-use regression model was introduced [2,
3]. This model combines, in an intelligent way, the measured data with the land
use, resulting in hourly 4 × 4 km
2 maps of the concentration field for ozone, PM 10 ,
PM 2.5 , SO 2 and NO 2 . More recently, NH 3 and BC were added to the model. The
core of the interpolation model consists of a land use indicator, the β-parameter.
This parameter is based on the local land use around the measurement locations and
is then used to detrend spatially the measurements. This detrending eliminates the
local character of the measurements (at a scale of 4 × 4 km
2 ). The resulting field
can then be interpolated over the area (using kriging interpolation). Afterwards, for
every location, the results are retrended again, using the β-parameter to add the local
characteristics into the results.
The direct use of measurements results in low biases and it has been extensively
shown by leaving-one-out validation and independent validation sets that the resulting RIO-model delivers high quality results, with low biases, low RMSE values and
high correlations (Fig. 22.1).
22.4 Increasing the Resolution
In general, it is known that land-use regression models (LUR) such as RIO offer
a cost-effective methodology for air quality assessment and, if the available proxy
data is at high resolution, can be used at high resolution. However, they are prone
to overfitting [9] which can be reduced by increasing the number of measurement
data points in the region. However, in order to obtain good results, an extensive
monitoring network is needed covering the resolutions for which the resulting map
is made. In other words, for the 4 × 4 km
2 map that was constructed above, the
current telemetric network is sufficient. However, for high resolution maps for the
region of Flanders, the number of stations would need to increase quite strongly. This
is feasible if low-cost measurement systems, such as passive samplers, can be used,
but if hourly data is needed the cost would be very high. In addition the quality of
141
network of the Flanders Environment Agency (VMM). The other were low-resolution
model data (15 × 15 km
2 ) provided by a chemical transport model (CTM) BelEUROS. The raw CTM model data was (still is) quite poor. This was partly due to the
fact that the spatial scale of the CTM model is not adequate to represent the high
degree of spatial variability in a densely populated area such as Flanders. The model
also exhibited time- and spatial-dependent biases. Therefore, the first course of action
was to improve the quality of the background concentration map.
22.3 Introducing a Better Background Map
In order to improve the quality of the low-resolution background map, and in order
to improve the resolution, an hourly land-use regression model was introduced [2,
3]. This model combines, in an intelligent way, the measured data with the land
use, resulting in hourly 4 × 4 km
2 maps of the concentration field for ozone, PM 10 ,
PM 2.5 , SO 2 and NO 2 . More recently, NH 3 and BC were added to the model. The
core of the interpolation model consists of a land use indicator, the β-parameter.
This parameter is based on the local land use around the measurement locations and
is then used to detrend spatially the measurements. This detrending eliminates the
local character of the measurements (at a scale of 4 × 4 km
2 ). The resulting field
can then be interpolated over the area (using kriging interpolation). Afterwards, for
every location, the results are retrended again, using the β-parameter to add the local
characteristics into the results.
The direct use of measurements results in low biases and it has been extensively
shown by leaving-one-out validation and independent validation sets that the resulting RIO-model delivers high quality results, with low biases, low RMSE values and
high correlations (Fig. 22.1).
22.4 Increasing the Resolution
In general, it is known that land-use regression models (LUR) such as RIO offer
a cost-effective methodology for air quality assessment and, if the available proxy
data is at high resolution, can be used at high resolution. However, they are prone
to overfitting [9] which can be reduced by increasing the number of measurement
data points in the region. However, in order to obtain good results, an extensive
monitoring network is needed covering the resolutions for which the resulting map
is made. In other words, for the 4 × 4 km
2 map that was constructed above, the
current telemetric network is sufficient. However, for high resolution maps for the
region of Flanders, the number of stations would need to increase quite strongly. This
is feasible if low-cost measurement systems, such as passive samplers, can be used,
but if hourly data is needed the cost would be very high. In addition the quality of