The starting point is a building dataset for the WB, from the Global ML Building Footprints project , denoted MLBuildings onwards in this report. For the WB region, the dataset contains about 30 million building polygons. For parts of the Republic of North Macedonia and southeastern Serbia, there are gaps in the MLBuildings dataset. The gaps were replaced by building polygons from the OpenStreetMap project . It is worth noting that for those areas, the OpenStreetMap has a fairly low level of coverage of buildings and therefore does not compensate fully for the lack of data in the MLBuildings dataset. An example of building polygons for Novi Travnik (Bosnia and Herzegovina) can be seen in Figs. –.

CORINE land cover (CLC) is a geographic land-use dataset covering Europe, in vector format, with a minimal mapping unit (MMU) of 25 ha. Urban Atlas (UA) is a detailed land-use dataset covering European municipalities having at least 100 000 residents, with an MMU of 0.25 ha. The CLC dataset was cut out for the WB region, and the attributes for land-use classes were adjusted to match those of the UA dataset. Further, cutouts were made in the CLC datasets for all cities where UA data were available. Finally, the CLC and UA datasets were merged into a single land-use dataset.

The following fuels have been included:

wood

Starting from the MLBuildings + OpenStreetMap dataset, described in Sect. , the first step was to remove all buildings that are not one- to three-family housing units. The buildings are all polygons, allowing for calculation of the building footprint area for every building. Incorrect and duplicate geometries were detected and corrected. Then, all buildings having a footprint area less than 50 m² or more than 200 m² were removed.

The next step was to determine which of the remaining buildings are located on land that can have housing. The land-use dataset (CLC and UA combined, as described above) was combined with the buildings dataset so that each building polygon also had attributes for land use. This means that each building has a land-use category associated with it, such as “Industrial or commercial units” or “Continuous urban fabric”. Using this information (combined with the filtering of the footprint area described above) for each building, it was then possible to filter out buildings unlikely to be one- to three-family housing units.

The following land-use classes

The classes listed here are a combination of the UA and CLC datasets.

All buildings belonging to land-use classes other than the above were removed from the dataset. While some of the above land-use classes are mainly agricultural lands, it can be seen, from random sampling of the dataset, that those land-use classes may contain housing, in particular for areas close to smaller settlements, though the extent of this is different in different subregions of the WB region. An example image of the buildings dataset after filtering is shown in Fig. .

Finally, because the area of each individual house is hard to estimate correctly from the MLBuildings dataset, the polygon geometries are not needed, and therefore we convert the polygon layer into centroid points. The outline area size of each individual building is not used to calculate the heating energy needs.

At each building point, energy consumption for the different fuels was assigned based on available statistical data. The data provide the energy usage per fuel per one- to three-family housing unit. The level of detail, in terms of geographical regions and fuels, varies between the different data sources. Each building point in the dataset was processed to calculate the energy usage.

Albania. Region-wise energy usage data (three climate zones), from .

Once the energy use for each fuel has been assigned to all building points, the resulting emissions for the relevant pollutants may be calculated. The methodology follows the EEA guidebook for emission factors as well as the composition of heating appliances. For each building point, emissions were calculated for each fuel, based on the composition of heating appliances and the emission factors for each appliance. The following pollutants were included in the dataset: PM₁₀, PM_2.5, NO_x, and SO_x.

As an example to illustrate the emission calculations, we consider wood fuel in Serbia. The EEA guidebook lists the appliance composition as 94 % stove, 5 % single house boiler, automatic feed (denoted shba below), and 1 % single house boiler, manual feed (denoted shbm below). For each substance, the effective emission factor for a single house is then the weighted average of the emission factors for the three appliances: 1EFeff=0.94⋅EFstove+0.05⋅EFshba+0.01⋅EFshbm. For each combination of fuel and appliance type, the EEA guidebook contains a suggested emission factor for each substance, which was used in this work.

The MLBuildings dataset contains data based on satellite imagery captured during multiple years, from 2014 and onwards. Therefore, the buildings dataset in itself does not correspond to a single, well-defined year. Similarly, the energy usage data represent multiple years, i.e., 2015–2020, while the appliance composition data in the EEA guidebook are for the year 2010. The yearly change in heating emissions from one- to three-family housing units is not expected to vary considerably between the years in the period 2010–2020. We therefore set the year to be 2019, as it is the most recent year in the period that was not affected by the COVID pandemic.

Finally, the point-wise emissions were rasterized into grids for each substance, with 500×500 m horizontal resolution, in the ETRS89-extended/LAEA Europe projection (EPSG:3035). While the calculations of energy and subsequently emissions were done for centroid points for all buildings, the data were rasterized into cells because they might otherwise seem to be more detailed and accurate than they actually are. Further, when using the resulting dataset for atmospheric dispersion modeling or similar, it is computationally more efficient to treat the data as rasters.

Table  list the number of buildings together with estimates from other sources. Also listed is the percentage of urban buildings, calculated from the number of buildings in all urban land-use categories (land-use code ≤11240 in Sect. ). The other sources (references are listed in the table) are different statistical reports and not from national databases of building permits or similar. It is also worth noting that in the different references, the categorization of buildings may differ. As far as possible, we have compared the number of buildings in this work to one- to three-family housing units in the references.

For several countries, the agreement is fairly good, especially for Albania, the Republic of North Macedonia, and Serbia. For Kosovo and especially for Bosnia and Herzegovina, the methodology used in this work leads to considerably more buildings, 9 % and 30 % more, respectively. The resulting number of buildings for Bosnia and Herzegovina may be overestimated, though this is uncertain because there is a lack of independent housing data. It is also uncertain what could cause such overestimation.

Inspection of a random selection of locations suggests that the high number of buildings for Bosnia and Herzegovina is due to the dataset having too many buildings in rural areas where there are smaller farms. This consideration is also supported by the percentage of urban buildings in the dataset, which is considerably lower for Bosnia and Herzegovina (and Montenegro) than for the other countries. This may have a larger effect in Bosnia and Herzegovina than in the other countries. However, the percentage of urban buildings may very well be different in the different WB countries. There are also uncertainties in the references for the number of residential buildings. In order to determine the accuracy of the number of buildings, more independent datasets are required.

In Fig. , the total energy is shown. All building points have been rasterized to a 500×500 m grid. It is worth noting some of the different patterns that can be seen in this image, particularly one in Serbia in which the northern part lies on a plateau and has large agricultural areas with settlements in fairly concentrated villages, leading to a “dotted” pattern in the image. The southwestern part of Serbia, on the other hand, is more mountainous and has smaller agricultural areas scattered between the mountains; the settlements are also more scattered into smaller pieces. This leads to a more continuous pattern in the image.

When comparing the countries in the Western Balkans in this image, it is worth noting that the calculated result in Fig.  depends on both the number of buildings and the statistical data for heating energy needs.

An uncertainty not included in this study is the fraction of residential buildings that are not in use. This has not been included due to a lack of available data.

Maps of emission data for PM_2.5, NO_x, and SO_x are shown in Figs. , , and . Note that the scales in the figures are different. PM₁₀ is not shown, as it is very similar to PM_2.5, which is expected for heating emissions due to the fact that most of the particles from combustion in PM₁₀ are fine particles (PM_2.5). Both PM_2.5 and NO_x show similar patterns, while for SO_x, the emissions in Montenegro, Albania, and the Republic of Northern Macedonia are much lower compared to the other countries. This is an effect of the heating energy data for different fuels, which are very low or zero for coal in these three countries, leading to very low SO_x emissions.

The resulting emission data for NO_x and PM_2.5 were compared to CAMS-reg 4.2, sector C , as well as EDGAR, sector buildings (accessed through https://edgar.jrc.ec.europa.eu/country_profile/, on 10 September 2024). The CAMS and EDGAR datasets are regional and therefore have a coarser resolution (0.05×0.1 and 0.1×0.1°

0.1×0.1° corresponds to about 8×11 km for the WB region.

The CAMS emissions were reprojected to the ETRS89-extended/LAEA Europe projection (EPSG:3035) and resampled to a 1×1 km grid size. Then, the total emissions for the different regions were calculated using the zonal statistics in QGIS. The resulting national statistics are shown in Table .

For NO_x, the overall agreement between the two regional datasets and that between this work and the regional datasets are similar, though the results in this work are closer to the EDGAR dataset than to CAMS-reg. The largest difference is found for Bosnia and Herzegovina, where this work has more than twice the emissions compared to CAMS-reg and EDGAR.

While the EDGAR dataset has higher emissions than CAMS-reg for NO_x, the case is the opposite for PM_2.5, except for Serbia and Montenegro, where EDGAR is about a third higher. For PM_2.5, the emissions in this work are considerably higher than both regional datasets. Because this was not the case for NO_x, the discrepancy is not only a result of the number of buildings but must also be due to different fuel usage and/or emission factors.

Finally, for SO_x, the data in this work are rather non-consistent when compared to the regional datasets. While this is much lower for some countries, it is considerably higher for other countries. This is a direct effect of the fuel data, as mentioned above. For Albania, Montenegro, and the Republic of North Macedonia, there is no coal usage reported (see Sect. ), which leads to very low SO_x emissions, while for Bosnia and Herzegovina and for Serbia, the coal usage leads to high emissions of SO_x, in this case much higher than for the regional datasets.