Influence of land cover on noise simulation output – A case study in Malmö, Sweden

1 Introduction

Urban densifications are becoming increasingly more common as a result of the dual effect of global population growth [1] and urbanization [2]. Urban densifications are considered more sustainable compared to urban sprawls [3], since more people get to use the same infrastructure while valuable agricultural land, often found on the outskirts of cities, is preserved [4]. However, planning an urban densification is a complex task, with one of the major challenges being the mitigation of noise pollution [5,6].

Long-term exposure to noise has been associated with sleep disturbance [7], cardiovascular disease [8,9], metabolic disease [8], stress [10], depression [11], annoyance [12], as well as reduced cognitive performance [13,14] and productivity in the workplace [15]. In the EU, there are estimations of about 12,000 people dying prematurely every year due to exposure to environmental noise, with road traffic as the main source [16]. Noise has also been identified as the causal effect behind changes in the physiology (e.g. reduced reproductive success and increased mortality) and behaviour (temporary or permanent departure from noisy areas) of urban wildlife, thus threatening the biodiversity and balance of urban ecosystems [16,17].

To prevent negative health impacts, the United Nations’ World Health Organization (WHO) and European Union (EU) have issued noise guidelines while individual countries have introduced noise-regulating laws and recommendations. Since road traffic is the most widespread noise source in Europe and the main contributor to high-level noise exposure in urban populations [18], WHO’s guidelines recommend a reduction of road traffic noise levels below 53 dB(A)^[1] for the entire day L _den ^[2] and below 45 dB(A) for night-time L _night ^[3] [19]. EU has issued the European Noise Directive (END) to monitor and address the exposure of citizens to noise pollution. According to END, municipalities (population > 100,000 inhabitants) of every member state are obliged to create and publish noise maps as well as noise management action plans in 5-year intervals for locations with estimated noise levels above 55 dB(A) L _den and 50 dB(A) L _night [20]. In Sweden, it is mandatory to check for noise pollution when planning new buildings [21,22,23,24]. This is controlled at various stages of the urban planning process (e.g. detailed development plan, building permit).

Exposure to noise pollution is assessed by performing regular real-world noise measurements or by conducting computer-based noise simulations. Computer-based noise simulations are less expensive and have the potential to cover larger geographic areas compared to high-quality noise measurements [25]. Consequently, noise analyses required by the EU or for urban planning purposes are usually conducted by executing noise simulations [26,27].

However, noise simulations are time-consuming, and the increasing pressure of constructing residential buildings more rapidly in Sweden has led to initiatives aiming to repeal the laws and cancel the recommendations related to noise pollution in urban planning [28,29]. Albeit, in the context of urban densifications, even if modern sound-insulating techniques could potentially largely prevent outdoor noise from entering buildings [30,31], this still leaves the issue of population and wildlife exposure to noise in outdoor areas open. Hence, excluding noise analyses from the planning of urban densifications would jeopardize urban sustainability. To avoid that, it is imperative to reduce the time required for conducting noise analyses. Common factors responsible for the prolonged duration of noise analyses (besides available computational power) are (1) the size of the study area, (2) the level of detail of the input datasets, (3) the time required for finding and pre-processing input datasets, and (4) the choice of noise propagation model (NPM) [32].

Noise simulations are based on a NPM, which is a mathematical model describing the physical processes that disperse noise emitted by a source in space and time. By using the NPM, it is possible to estimate how noise propagates from one or several sources to target points (receivers) often placed on building façades. Several NPMs have been developed. The END recommends the implementation of the CNOSSOS-EU NPM [33] when conducting city-scale (1–5,000 km²) noise analyses. Currently, in Sweden, it is mandatory to execute noise simulations for urban planning using the RTN-96 NPM. RTN-96 is soon to be replaced by Nord2000. Nord2000 is more detailed compared to CNOSSOS-EU and RTN-96 [34,35,36,37,38] and is, therefore, deemed more suitable for simulating noise propagation on local or street scale (1 m² to 1 km²) for, e.g. urban planning.

Nord2000 requires information on the characteristics (geometry and material) of ground surfaces (vegetation, soil, rock, etc.) and man-made objects (buildings, roads, etc.) including their flow resistivity (reflecting the air permeability through porous materials), which is related, among others, to the specific acoustic impedance and whose units are N s/m⁴ or Pa s/m²). Specific Acoustic Impedance (Z) is defined as the complex ratio of the effective sound pressure at a point of an acoustic medium to the effective particle velocity at that point and is measured in units of N s/m³. Land cover (LC) information describes the material of a ground surface (e.g. grass) and is not to be confused with land use (LU), which refers to how a ground surface is utilized (e.g. park or football court). Nord2000 uses LC information to identify the noise absorbance of different ground surfaces and man-made objects [39]. LC information is available in varying formats (raster, vector) and quality, but producing it is a non-trivial, time-demanding task often requiring the processing of multiple datasets from different sources [40].

CORINE is a pan-European open-access LC/LU product featuring 44 thematic classes with a 6-yearly updating frequency, a minimum mapping unit of 25 ha for aerial phenomena and a minimum width of 100 m for linear phenomena [41]. The European Environment Agency (EEA) in collaboration with Copernicus EU and the European Space Agency have created the Urban Atlas, an open-access product that includes LC information on a selected number of urban areas in Europe [42]. The Urban Atlas LC information is updated every 6 years and has a spatial resolution of 10 m [42]. Many countries also provide open-access LC datasets with national coverage, usually derived from satellite remote sensing data of a higher spatial resolution than CORINE [40,41,43]. Municipalities collect information on how urban ground surfaces are utilized as a primary resource for urban planning. This information is produced by processing aerial images, very-high-resolution satellite images [44], LiDAR data [45], and sometimes even terrestrial observations. The final output dataset is stored in vector format (lines) and is characterized by high spatial accuracy. Though the (LC-related) dataset is frequently updated, producing surfaces (polygons) from line layers is a complex and time-consuming task.

Cities often have very detailed LC data for some LC classes such as transportation, water, and park areas, but their data seldomly cover the whole city. A possible solution for producing complete LC datasets for cities is to utilize artificial intelligence (AI) techniques such as convolutional neural networks (CNN). CNNs can be used to automatically derive LC information from aerial and satellite images [46,47,48,49,50,51,52,53]. CNNs are the cornerstone of most semantic segmentation models (i.e. a computer vision task in which the goal is to categorize each pixel in an image into a class or object), due to their ability to learn spatial hierarchies and capture localized patterns [54]. Though aerial and satellite imagery are openly available from multiple sources, this is not the case for LC training datasets. Some international initiatives exist for LC classification, but the number of available datasets for training and evaluation is still significantly lower compared to other domains (e.g. autonomous driving) [55,56].

To minimize the preprocessing time of LC input data to noise simulations, it is important to automate the mapping of flow resistivity value to every LC class. There are examples of such mappings [57,58,39,59,60], but whilst there is no LC classification standard detailed enough for urban areas, additional work is often required [61].

Semantic 3D city models could be utilised to minimize the time required for finding and accessing input datasets for noise simulations. Semantic 3D city models are digital 3D representations of the urban physical environment including the option to store semantic information for every geometry (building address, material of building façade, etc.) [62]. The open geospatial consortium has released a semantic 3D city model standard called CityGML [63,64]. CityGML includes several themes (Building, Transportation, City Furniture, Vegetation, LandUse, etc.) supporting the storage of LC data and other input data to NPM. Furthermore, CityGML can be extended to support the storage of additional information required by certain applications using application domain extensions (ADE) [63]. Noise ADE is an example of a CityGML ADE created to support noise simulations including acoustic impedance value per ground surface object [65].

As noted above, the most common factors responsible for the prolonged duration of noise analyses are the size of the study area, the input data level of detail, the time required for finding and pre-processing input data, and the choice of NPM. From an urban densification planning perspective, the size of the study area is preset to local or street-level while the choice of NPM has already been predetermined by national directives, leaving only the two remaining factors open to interpretation. Here the focus is set on LC. LC datasets have the potential to become accessible to users by being stored in the LandUse theme of semantic 3D city models. However, it is crucial to determine the minimum requirements an LC layer must meet to be suitable as input to noise simulations before storing it in that manner. As previously mentioned, LC data varies depending on spatial resolution, temporal updating frequency as well as on the observation data and classification system used to produce it. Another factor affecting the suitability of raster LC datasets in noise simulation analyses is the priority rules applied. For example, for raster cells containing more than one LC class, a majority rule could be applied; alternatively, the LC class could be set based on LC class priority. The aim of this study is to examine the effect of different LC input data on noise simulations. The following research questions are addressed:

• Is noise simulation output affected by different priority rules applied to produce an LC dataset?

• What is the influence of the LC data spatial resolution on noise simulation output?

• From a noise simulation perspective, what should the specification requirements for the LC classification system producing LC datasets be?

2 Methods

This section is dedicated to presenting the study area along with the materials and methods used for the implementation of our case studies. It should be mentioned that all case studies consider car traffic as the only noise source.

2.1 Study area

The study area encompasses the new Bellevue and Lorensborg neighbourhoods of Malmö municipality in southern Sweden and covers an area of approximately 945,000 m² (Figure 1). The location was chosen because it was recently deemed suitable for densification. The area is dominated by residential buildings (villas as well as 3–6 story high buildings) with some commercial and educational buildings also being present. Moreover, the study area is characterized by varying land cover comprising both vegetated areas (parks, gardens, traffic islands, etc.) and various built-up surfaces (roads, pavements, car parks, cycle paths, walking paths, sports facilities, school yards, refuge islands, pools, ponds, etc.) close to the city centre.

Figure 1

Study area – Lorensborg and new Bellevue neighbourhoods in Malmö municipality, southern Sweden.

The area includes some urban roads that are highly trafficked (Figure 2) with the highest loads being observed on Lorensborgsgatan, Stadiongatan, and Ärtholmsvägen (marked in orange and red colours correspondingly). These roads are used for transportation in and out of the city centre.

Figure 2

Map displaying road traffic density as the average number of vehicles per day for 2020 (Malmö municipality, 2020).

Figure 1 highlights the buildings that were used as input to the noise simulations in red and white colours. The red buildings represent the buildings hosting receivers, while all buildings (white and red) function as noise-obstructing surfaces. In the context of this study, receivers are defined as point locations placed at a certain height above ground (1.5 and 4.0 m) and at 0.1 m distance from the building façade for registering the noise levels computed by the noise simulation software. The red buildings were chosen based on how close they were to highly trafficked parts of urban roads (sound source), how varying the land cover between the sound source and the building in questions was, the building typology (shape and height) as well as the use of the building (residential, commercial, educational, retirement home, etc.).

Trial simulations for all LC layers were performed using receivers from a subset of the study area (i.e. buildings identified in BS1 in Figure 1), positioned at heights of 1.5 and 4.0 m above ground (recommended receiver heights by Kragh et al. [39] and Kephalopoulos et al. [33]). As the largest differences in dB(A) L _den were observed for receivers placed 1.5 m above ground, we decided to execute noise simulations for the remaining buildings only at that height. An additional reason for this was that the recommended receiver height for sensing noise pollution for one-storey buildings was 1.5 m [39] and that we wanted to explore the worst-case noise exposure scenario for day-care centres, schools, retirement homes, residential apartments, and offices located on the ground floor. When comparing the noise simulation outputs with different LC layers as inputs, the most noticeable differences are observed for receivers located on buildings in BS6 (Figure 1).

2.3 Other input data to noise simulations

To get a detailed representation of the ground surface, we imported elevation points from a Digital Elevation Model DEM (downloaded from Lantmäteriet) at a high spatial resolution (1 m).

Line vector geometries for roads were provided by Malmö municipality’s division for property and roadwork administration (Fastighet och Gatukontoret – FGK), while traffic count information (i.e. total number of vehicles per vehicle category per lane per day) for every road segment for 2020 was provided by different sources (Swedish Transport Administration (Trafikverket), FGK, and Malmö municipality’s division for environmental administration (Miljöförvaltningen -– MF)). The speed for light vehicles (category 1) and medium vehicles (category 2) was set to the registered speed limit (30 km/h or 40 km/h) of the corresponding road segment (as defined by the Swedish National Road database – NVDB) and was the same for all time periods (daytime 06:00–18:00, evening 18:00–22:00, night- time 22:00–06:00). The distribution of the traffic per vehicle category (light, medium) for different time periods of the day (day, evening, and night) follows Malmö municipality’s rule of thumb for light vehicles and the recommended composition for residential roads as stated in Kragh et al. [39] for medium vehicles (Table 1). The recommended traffic distribution of light vehicles on residential roads stated in Kragh et al. [39] (day: 0.80, evening: 0.10, night: 0.10) is not much different from Malmö municipality’s rule of thumb, but the latter was preferred since it is based on years of experience measuring traffic in the area. No such information was available for heavy vehicles.

Table 1

Traffic composition per vehicle category and time of day

Vehicle category	Traffic composition for Day (06:00–18:00)	Traffic composition for Evening (18:00–22:00)	Traffic composition for Night (22:00–06:00)	Source
Light vehicles (category 1)	0.76	0.17	0.07	Malmö municipality
Medium vehicles (category 2)	0.85	0.05	0.10	Table 5 in Kragh et al. [39]

Information on traffic intensity (total number of vehicles) per vehicle category and time of day was estimated based on the following equations:

(1) LV day = 0.76 × 100 −   HV 2020 100 × MVD 2020 ,

(2) LV evening = 0.17 × 100 − HV 2020 100 × MVD 2020 ,

(3) LV night = 0.07 × 100 − HV 2020 100 × MVD 2020 ,

(4) MV day = 0.85   × HV 2020 100 × MVD 2020 ,

(5) MV evening = 0.05 × HV 2020 100 × MVD 2020 ,

(6) MV night = 0.10 × HV 2020 100 × MVD 2020 ,

where LV_day, LV_evening, and LV_night correspond to the total number of light vehicles per a certain time period of a day (daytime, evening, and night), while MV_day, MV_evening, and MV_night refer to the corresponding values for medium vehicles. HV₂₀₂₀ represents the percentage of traffic consisting of medium heavy vehicles and MVD₂₀₂₀ represents the total number of vehicles per weekday for 2020.

Noise simulation software requires information on noise-obstructing objects (buildings, noise barriers, etc.). In our study, we only consider 3D buildings available as CityGML LOD2 geometries (Figure 3) retrieved from Malmö municipality’s 3CIM data. As most noise simulation software can only handle CityGML LOD1 buildings [66,71,72], the dataset was transformed to CityGML LOD1 with the height of every building set to the highest point of its equivalent CityGML LOD2 version.

Figure 3

The five LODs of CityGML 2.0 From “An improved LOD specification for 3D building models,” by Biljecki et al. [73].

Since noise simulations are computationally demanding and time-consuming, an effort was made to limit the number of receivers and simulations without compromising the quality of the analysis. For this purpose, noise simulations were executed for receivers placed at 1.5 m above ground and at 0.1 m from the building façade with a horizontal distribution in accordance with the CNOSSOS-EU model [74].

3 Results

Section 3.1 presents the result of the comparison of the LC datasets while Section 3.2 presents the results of the noise simulations.

3.1 Comparison of LC data

The donut plots show the percentage of the study area classified as low vegetation, building surface, road and railway surface, water, or other built-up surfaces in accordance with NS-LC L2 for the four LC datasets (Figure 4). The largest percentage of vegetation-classified area is observed in NMDa (61.1%), with NMDm (45%) and PKO (44.3%) showcasing similar percentages, and AIO having the smallest percentage (40.6%). Areas classified as building are relatively similar in all LC datasets, except for a slight overrepresentation in NMDm and a corresponding underrepresentation in NMDa. Road surfaces are overrepresented in both NMDa and NMDm compared to PKO and AIO. The percentage of areas classified as other built-up surfaces varies substantially between the datasets. Water is only included in the NMDa and PKO datasets covering the same percentage of the study area.

Figure 4

Donut plot presenting the percentage of the study area covered by every NS-LC L2 thematic class. The maps provide a descriptive illustration of the study area’s LC classification results per dataset.

Figure 4 depicts the maps presenting LC information for all four LC datasets according to NS-LC L2. The AIO and PKO maps are the ones that look most alike and resemble the orthophoto, while the NMDa and NMDm maps (having a 10-m spatial resolution) look pixelated and fail to accurately represent hard surfaces that do not correspond to buildings or main roads (e.g. pavements, paved inner yards, walking paths, bicycle lanes, etc.). In comparison to the other maps, vegetation is clearly overrepresented in the NMDa map.

Figure 5 displays the LC differences observed when comparing one LC dataset to another. The columns present changes in LC per NS-LC L2 class as percentage bar plots for all LC changes occurring when comparing a specific pair of LC datasets (the sum of every column is 100%).

Figure 5

Percentage area of the study area that has retained or changed its LC classification when comparing two LC datasets. Columns correspond to LC changes occurring when comparing two LC datasets (the total sum of percentage areas in a column is 100%). Rows show the results per NS-LC L2 class for all LC dataset pair comparisons.

From Figure 5 it becomes evident that the most similar LC datasets are PKO and AIO with 87.07% of their areas having been classified with the same LC class. The least similar datasets are the NMDa and AIO (58.27% commonly classified areas), and NMDa and PKO (60.86% commonly classified areas). NMDa and NMDm are 79.23% similar. Overall, the largest LC changes occur in the vegetation class particularly when converting areas of this class to other built-up areas. A smaller percentage of areas are also experiencing LC changes transforming building surfaces to vegetation or other built-up areas. A similar pattern is observed for road surfaces being converted to vegetation or other built-up areas. A small percentage of other built-up areas are converted to vegetation, road surfaces, and building surfaces. The smallest area percentage LC changes are observed for the water LC class.

3.2 Comparison of noise simulation output

Table 2 shows the min, max, median, mean, and standard deviation of the noise simulation output (L _den, L _day, and L _night) difference between pairs of LC datasets. The difference is expressed in dB(A).

Table 2

Statistical results of difference in simulation noise output produced by the four LC datasets

LC datasets (difference)	L den (dB(A))					L day (dB(A))					L night (dB(A))
	Min	Max	Median	Mean	Std	Min	Max	Median	Mean	Std	Min	Max	Median	Mean	Std
NMDm-NMDa	−0.1	3.0	0.6	0.8	0.5	−0.1	3.3	0.7	0.9	0.6	−0.1	2.6	0.6	0.7	0.5
PKO-NMDa	−0.8	2.2	0.6	0.6	0.4	−0.8	2.3	0.6	0.7	0.5	−0.8	2.1	0.5	0.5	0.4
PKO-NMDm	−1.8	1.5	−0.1	−0.2	0.5	−2.1	1.5	−0.1	−0.2	0.6	−1.5	1.3	0.0	−0.1	0.5
AIO-PKO	−0.9	1.3	−0.1	0.0	0.3	−0.9	1.5	−0.1	−0.1	0.3	−0.7	1.3	0.0	0.0	0.2
AIO-NMDa	−1.0	2.5	0.5	0.6	0.5	−1.1	2.9	0.6	0.6	0.5	−0.8	2.1	0.5	0.6	0.4
AIO-NMDm	−2.4	1.6	−0.2	−0.2	0.5	−2.7	1.8	−0.2	−0.3	0.5	−2.0	1.3	−0.1	−0.1	0.4

In general, noise simulation output produced using different LC datasets as input differs from –2.7 to +3.3 dB(A) (Table 2). The largest noise simulation output differences are spotted for L _day (0–3.3 dB(A)) while the differences are somewhat smaller for L _den (0–3.0 dB(A)) and smallest for L _night (0–2.6 dB(A)). The LC dataset generating the greatest differences in noise simulation output compared to the others is NMDa. In this case, the average difference lies between 0.5 and 0.9 dB(A), which with the addition of the standard deviation, amounts to a difference ranging from 0.1 to 1.5 dB(A).

Figure 6 presents two relative cumulative frequency plots of the total number of receivers having a certain L _den dB(A) value for simulations executed using each one of the LC datasets. The left plot includes simulation output from receivers placed on buildings in a subpart of the study area (BS6 in Figure 1), while the right plot presents results from all the receivers in the study area. Figures 7 and 8 present the same information as Figure 6 but for L _day and L _night correspondingly.

Figure 6

Relative cumulative frequency graph of noise simulation output in L _den (left: receivers placed on buildings in BS6, right: all receivers in the study area) for the four different LC datasets.

Figure 7

Relative cumulative frequency graph of noise simulation output in L _day left: receivers placed on buildings in BS6, right: all receivers in the study area) for the four different LC datasets.

Figure 8

Relative cumulative frequency graph of noise simulation output in L _night (left: receivers placed on buildings in BS6, right: all receivers in the study area) for the four different LC datasets.

Simulated noise levels for all LC datasets seem to follow similar patterns when looking at the L _den, L _day, and L _night relative cumulative frequency graphs for all receivers in the study area (the right plot in Figures 6–8). The only exceptions being NMDa showcasing more receivers with lower dB(A) values and NMDm showcasing more receivers with slightly higher dB(A) values. This pattern becomes more prominent when focusing on subparts of the study area (left plot in Figures 6–8).

The histogram in Figure 9 shows the distribution of L _den noise levels produced by noise simulations using different LC input datasets. The noise simulation distribution can be approximated with a normal distribution with means and standard deviations that are almost identical for the outputs produced using the AIO, and PKO LC datasets as input. The distribution curve for noise simulation output generated using NMDa as input deviates from the others, indicating more receivers with lower dB(A) values, while the corresponding distribution curve for NMDm indicates the existence of more receivers with higher dB(A) values.

Figure 9

Histograms and distributions of simulated L _den noise levels produced using the four LC input datasets.

To investigate whether the spotted differences are significant, we implemented a t-test where the null hypothesis was that L _den, L _day, and L _night noise levels do not differ significantly between noise simulation outputs generated using different LC datasets. The results of the t-test (t-value and p-value) performed on pairs of noise simulation outputs are presented in Table 3.

Table 3

T-test results examining the significance of the difference between simulated L _den, L _day, and L _night noise levels produced by a pair of different LC datasets. Bold values rows correspond to statistically significant noise level dB(A) differences

Comparisons	L den			L day			L night
	Mean difference (dB(A))	t-Value	p-Value	Mean difference (dB(A))	t-Value	p-Value	Mean difference (dB(A))	t-Value	p-Value
NMDm-NMDa	0.8	3.07	0.002	0.9	3.32	0.001	0.7	2.75	0.006
PKO-NMDa	0.6	2.44	0.015	0.7	2.61	0.009	0.5	2.23	0.026
PKO-NMDm	−0.2	0.65	0.518	−0.2	0.74	0.462	−0.1	0.54	0.589
AIO-PKO	0.0	0.16	0.874	−0.1	0.25	0.804	0.0	0.03	0.974
AIO-NMDa	0.6	2.30	0.022	0.6	2.38	0.018	0.6	2.21	0.027
AIO-NMDm	−0.2	0.81	0.420	−0.3	0.99	0.325	−0.1	0.58	0.565

The difference in noise levels produced by simulations using different LC datasets as input is statistically significant for all L _den, L _day, and L _night noise levels when comparing NMDa to the other LC datasets (see rows with values in bold format in Table 3).

The largest mean difference in dB(A) values between noise simulations was observed for L _den between NMDa and the other LC datasets. To analyse this, Figure 10 was created to explore the spatial distribution of LC changes (shown as difference in flow resistivity) in relation to L _den dB(A) noise-level differences. Every LC geometry was assigned with an LC class and corresponding flow resistivity value (Table A4). The background maps (Figure 10) depict differences in flow resistivity between NMDa and the other LC datasets as well as a selection of receivers whose L _den (dB(A)) difference was equal or greater than 2 dB(A) when comparing NMDa simulation output to simulation output generated using other LC datasets. It should be mentioned that all highlighted receivers have an L _den value greater or equal to 53 dB(A). When comparing PKO with NMDa noise levels, 11/1,468 receivers (0.8%) showcase a L _den diff >2 dB(A) while having a L _den > 53 dB(A). When comparing AIO with NMDa noise levels, 17/1,468 receivers (1.2%) meet those requirements and, finally, when comparing NMDm to NMDa noise levels, 71/1,468 receivers (4.8%) do the same. The largest differences in L _den noise levels are observed for receivers being placed on buildings close to the most trafficked roads (Figures 2 and 10). Examining the background maps in Figure 10, it is becoming obvious that the other LC datasets have more surfaces with higher flow resistivity values (purple-shaded areas) compared to NMDa.

Figure 10

Maps presenting noise simulation output based on the 24 h L _den metric, highlighting receivers with L _den >= 53 dB(a) in both LC datasets, and whose corresponding difference is at least 2 dB(A). A background map depicting changes in flow resistivity values when subtracting NMDa flow resistivity values from the flow resistivity values of another LC dataset is used to show the spatial relevance between changes in flow resistivity and changes in simulated L _den noise levels.

During the manual digitization of the LC datasets NMDm and PKO, several LC types that were not included in the NS-LC emerged. Typical examples are surfaces covered by wood (e.g. patios, staircases), glass (greenhouses), grass concrete (grass concrete paver), metal (temporary platforms, metal roofs), and grass (green roofs) (Figure 11). In PKO wooden surfaces cover 0.25% (2,342 m²), grass concrete pavers 0.16% (1,551 m²), and glass roofs 0.009% (80 m²) of the study area. In NMDm green roofs cover 0.07% (700 m²) of the study area.

Figure 11

Examples of wooden surfaces (left), grass concrete pavers (middle-left), and green roofs (middle right and right) found in urban settings.

4 Discussion

According to our LC comparison, NMDa overestimates vegetation-covered areas. This is likely because NMD was initially produced to facilitate environmental monitoring and therefore applied rules that prioritized vegetation (cf. 2.2.1). Another factor contributing to the overrepresentation of vegetation in NMDa is that the satellite imagery does not capture the ground surface beneath vegetation. Besides tall trees with dense foliage covering hard surfaces like asphalt roads (Figure A1), there are cases where vegetation on balconies of tall buildings and green roofs have been classified as forest (Figures A2 and A3), vegetation on building facades having been classified as forest or vegetation (Figure A4) and climbing plants growing on external walls being classified as vegetation (Figure A5).

It is worth mentioning that, at its current state, NS-LC does not allow distinguishing buildings with different types of roofs (clay tiles, green roof, solar tiles, metal roofing, wood roofing, asphalt shingle, etc.). Adding roof material to the NS-LC classification at L4 (A. Persson, Lund University, Department of Physical Geography and Ecosystem Science, Lund, personal communication, September 17, 2024) could be significant from a noise propagation perspective since green roofs have totally different flow resistivity values compared to other roof materials [77].

Vegetation is a bit underrepresented in AIO compared to the other LC datasets, but this could be attributed to some vegetated surfaces (e.g. grass lawns) being shaded by tall trees, bushes or buildings and the AI classification algorithm classifying those pixels as other built-up areas instead (Figures 4 and 5). Kuras et al. [78] have implemented a technique incorporating LiDAR with hyperspectral data to eliminate the effect of shadows based on the difference in height that is detectable with LiDAR data.

Ground surface areas classified as buildings are almost identical in AIO and PKO with the 0.4% difference possibly being attributed to the resolution of the orthophoto used in the AI LC-classification method. NMDa and PKO have similar percentages of building-covered surfaces, despite the fact of the Swedish Environmental Protection Agency implementing the algorithm that only classifies a cell as “building” if its centroid is covered by a building footprint. The NMDm shows a greater number of areas covered by buildings compared to the rest of the LC datasets. This is caused by implementing the majority rule on raster cells of coarse resolution (10 m).

Road areas are greatly overestimated in both NMD datasets due to their coarse spatial resolution. An additional reason in the NMDa LC dataset’s case, is the implementation of the priority rule ensuring that main roads are represented by adjacent raster cells. AIO and PKO produce the same percentages of areas classified as road.

Areas classified as other built-up areas are underrepresented in both NMD datasets, more so in NMDa as it favours vegetation over this LC class. However, this class is overrepresented in AIO mainly because shaded areas or areas not been deemed suitable for any other LC class, were classified as other built-up areas.

Areas covered by water are only present in NMDm and PKO. AIO does not include any areas classified as water due to a lack of training data for water-covered urban areas (e.g. swimming pools, ponds). NMDa does only classify a surface as water if it is larger than 100 m², which affects its representation of swimming pools, canals, and ponds in cities.

From Figures 4 and 5, it is evident that PKO and AIO have the highest level of resemblance with almost 90% of their areas being covered by the same hard or soft surface followed by NMDa and NMDm with 79.23% of resemblance, while NMDa and AIO are the most different with only 58.27% of their common area maintaining its initial surface classification. When comparing AIO’s and PKO’s differences from NMDa, the surfaces whose LC classification has changed from hard to soft or the contrary are scattered evenly over the study area (Figure 10). This indicates that the differences are occurring randomly and could therefore be a consequence of the coarser spatial resolution (10 m) of NMDa. The surfaces changing LC classification in the PKO and AIO difference map (Figure 10) are predominantly doing so from soft to hard and are relatively scattered. When looking at the maps depicting the difference between PKO, AIO, and NMDm from NMDa, the areas maintaining their original LC classification are fewer while those that change LC include both larger areas at certain locations as well as smaller areas scattered all over the study area. Changes in LC classification in this case could be attributed to both the coarser spatial resolution of NMDa (smaller areas changing LC classification) but also to NMD’s priority rules (larger areas changing LC classification). Another possible reason for changes covering larger areas when comparing PKO to the other LC datasets and the main reason behind differences between PKO and AIO is the fact that the municipality basemap used for producing it is more recent than the orthophoto with some areas having been densified (green areas turned into buildings or parking lots). This is an example where the updating frequency of remotely sensed imagery is too low, and raising this frequency becomes increasingly important in urban areas undergoing rapid development.

Though the quality and spatial resolution of LC has been proven to have a significant influence on noise propagation from airborne traffic in urban areas [79], this is not as prominent when studying noise propagation from car traffic in our study area. Noticeable differences are spotted on a local scale (BS6), but these differences are somewhat smoothed out when switching the focus to the entire study area (Figures 6–8). Generally, simulated outputs differ by about ±3 dB(A), as shown in Table 2. The largest differences in dB(A) values are observed for simulated noise values during daytime (L _day) when comparing the NMDa to the NMDm, PKO and AIO LC datasets (Table 2). Differences between L _den, L _day, and L _night simulation output produced by using NMDa compared to simulation output produced using the other LC datasets are on average close to 1 dB(A) with a max difference of 3.0 dB(A) L _den and statistically significant (Table 3), indicating that NMDa is not the best choice of LC dataset for noise analyses in urban settings. This becomes even more important when considering that the largest differences occur on locations close to noise sources (most trafficked roads) where NMDa surfaces have changed from hard to soft and the observed noise level is greater or equal to the WHO recommended threshold of 53 dB(A) L _den (Figure 10).

Sound travels faster during daytime due to high air temperature with humidity increasing that speed even further. Additionally, during daytime sound waves bend away from the ground, but typically, during night-time, because of more stable atmospheric conditions (or even temperature inversion), sound waves bend towards the ground. Consequently, at night, distant sounds can be heard as the sound waves get refracted and reach receivers from distant sources. In cities like Malmö, where humid or windy cold nights are not unusual for a great part of the year (air temperature varies from −12 to 29°C [80] while the annual average relative humidity is 81% [81] and wind speed varies from 0.0 to 13 m/s with the wind predominantly blowing from SW and NW directions [82]), living in ground floor apartments of buildings close to larger urban roads may significantly disturb sleep quality. This is due to the combined effect of the noise source being so close (road), ground surfaces close to the receivers being hard (e.g. pavement) and because of stable atmospheric conditions (which typically occur at night) making sound waves from car traffic pointing downwards hitting those hard surfaces close to the windows (located at 1.5 m above ground, i.e. same as receiver height). The same goes for the early morning hours that are cold and humid and would lead to higher noise levels disturbing people in schools, day-cares, and offices located on the ground floor of buildings close to busy roads. In both cases, using the NMDa LC dataset (a dataset that overrepresents vegetation and underrepresents hard surfaces) in the noise simulation would underestimate the noise at the building façade with true L _den values exceeding the WHO recommended threshold of 53 dB(A) for the highlighted receivers (Figure 10).

NMDa is the LC dataset requiring the least amount of pre-processing by the user before importing it to noise simulation software, but since it yields significantly different noise level estimations for both daytime and night-time, when maintaining low noise levels is so important for health and productivity, it could be worth looking at other alternatives. Noise simulations utilizing NMDm produce somewhat higher noise levels compared to AIO and PKO. However, those outputs are still quite similar, indicating that the priority rules applied in the LC classification algorithm can be more influential than the spatial resolution. This observation is especially important considering our study area is characterized by a mixed LC. Therefore, a spatial resolution of 10 m could potentially be enough for representing LC in urban areas depending on the choice of priority rules. However, considering the difference in simulated noise between NMDm and PKO or AIO, it is doubtful whether a coarser resolution would be good enough. Consequently, the suitability of other open-access LC datasets such as CORINE having an even coarser resolution is not favourable. Additionally, NMD, Urban Atlas, and CORINE have very low temporal updating frequencies, with Urban Atlas only providing LC information for a subset of urban areas in Europe.

AI-generated LC layers are a promising alternative for rapid LC dataset production, under the condition that aerial and satellite imagery are accessible and updated frequently, adequate training dataset are openly available, and AI-algorithms and workflows are easy to access or implement. Manually digitized LC datasets may be more representative but are time-consuming and expensive to produce. Regardless of what LC data is chosen, storing it in a semantic 3D city model may assist in sharing the updating tasks among municipality divisions and avoid overlaps and redundancy, ultimately contributing to an up-to-date and accessible LC layer for all actors involved in the urban planning process.

NS-LC should include thematic classes corresponding to all materials (including wood and concrete grass pavers) often found in urban settings. For NS-LC to be suitable for noise simulations, it is important that when aggregating NS-LC levels of detail then the flow resistivity value associated with the aggregated LC class is representative for all the LC classes included in the higher NS-LC levels.

Currently, this is achieved for most of the themes, the only exceptions being the buildings (e.g. depending on the material used on the building’s exterior or roof, this can affect the flow resistivity) and other built-up areas where excavated ground NS-LC L4 is supposed to have the same flow resistivity value as concrete NS-LC L4 when aggregated to built-up areas in NS-LC L1-L3. Additionally, as stated by Gustafson and Genell [61], new noise measurements are required to determine the flow resistivity of all urban surfaces included in the NS-LC. As flow resistivity values are greatly influenced by climate [83], it might be advisable for countries with non-uniform climates to include multiple flow resistivity tables depending on what climate zone each of their regions belongs to.

We have chosen to execute the noise simulations using SoundPLAN v.9.0 as it is considered one of the most used noise simulation software [84,85] and because it has implemented NPMs such as Nord2000 [84]. We did not try to execute the same simulations in any other noise simulation software.

The results found in this study correspond to a local or street scale area characterized by a mixed LC composition in a climate typical for a coastal city in southern Sweden. It would be interesting to repeat the experiments for more and larger areas with varying LC compositions under similar climate conditions. Our study area, though it includes many LC classes, is quite flat. Examining the influence of LC on locations with more varying topography would be interesting. Also, investigating the influence of LC on urban areas of different climates would be advisable [83].

Only simulation results for receivers placed at 1.5 m above ground were included in the study. Noise simulations for receivers placed at 4.0 m above ground were also performed, but only for a smaller area (see BS1 in Figure 1). The produced results were similar for all LC datasets. Examining if and to what extent different LC datasets influence the noise detected at receivers placed on different horizontal levels (heights) of building facades was not part of this study. This would however be interesting to study in more densely built-up areas with taller buildings or in areas with varying topography.

We used a small search radius (500 m), even though 1–1.5 km is the usual recommended search radius when conducting environmental noise simulations [37]. This was done to limit the execution time. Additionally, as the extent of our study area was approximately 1,000,000 km² and because most of our receivers were placed on buildings close to its centre, this meant that we got the influence of noise sources (road traffic) from the entire study area for most receivers. However, it should be noted that the purpose of this study was not to provide the most accurate noise simulation output, but rather to detect changes in noise simulation output caused by using different LC datasets as input.

5 Conclusion

In this study, we have examined the effect of using different land cover (LC) datasets as input to noise simulations. The largest effect, resulting in up to 3.3 dB(A) difference in simulated L _day values, stems from a combination of priority rules implemented when producing the LC datasets and what data were used to produce the LC datasets. The satellite-based (Sentinel-2; 10 m spatial resolution) NMDa LC dataset whose production rules prioritize vegetation, could not adequately identify LC under canopies. This dataset contained a substantially larger proportion of vegetation compared to the study’s other three LC datasets, resulting in significantly lower simulated L _den, L _day, and L _night noise values. From this perspective, it can be considered suboptimal for urban noise analyses. The second satellite-based (Sentinel-2; 10 m spatial resolution) LC dataset, NMDm, that implemented a majority rule to determine the LC class of every raster cell, produced simulated noise levels closer to the AI-generated and manually digitized LC datasets, indicating that priority rules in LC classification are more influential than spatial resolution.

When comparing the output between noise simulations utilizing LC datasets (NMDm) with lower spatial resolution (10 m) to higher spatial resolution (0.08 m) AI-generated LC datasets (AIO) and manually digitized vector LC datasets (PKO), the differences are small (on average less or equal to 1 dB(A)). That is, even though the datasets were produced using different techniques, resulting in non-identical LC representations, they generate similar noise simulation output. However, because of the fine-scale, mixed LC nature of the study area, utilizing LC datasets of lower spatial resolution than 10 m (e.g. CORINE) would not be advisable, since this would fail to represent objects of the urban fabric that substantially influence noise propagation. Nevertheless, a too-fine spatial resolution significantly increases the execution time of noise simulations, in our case leading to a more than doubled execution time. However, this result does not imply that it is acceptable to disregard small 3D barriers (e.g. sound barrier walls) as input to noise simulations; their effect has not been evaluated in this study.

The specifications of LC datasets should ensure that every type of material often found on ground surfaces in urban environments is represented in the corresponding LC classification system. Taking into consideration that most LC datasets are not produced with the objective to represent surfaces of urban environments (e.g. CORINE, NMD), it is natural that some materials (e.g. wood or glass) or structures combining different materials (e.g. grass concrete pavers) are not represented in the classification system. As these materials or structures may have a significant impact on simulations conducted in the urban planning process (including noise as shown in this study), we propose that national agencies take these into account when creating their national specifications for LC. Additionally, if a multiple-level LC representation schema is to be implemented, it is important that higher level LC thematic classes (concrete, excavated ground, artificial grass, etc.) have reasonably similar flow resistivity values with their corresponding aggregated (lower level) thematic class (e.g. other built-up areas). Finally, to speed up noise analyses, it is paramount to associate every thematic class in the NS-LC with its corresponding flow resistivity based on updated real-world noise measurements inside urban areas.

Moreover, many datasets focus on LU instead of LC, mix the two terms, or are not detailed enough to point to the material of the ground surface (e.g. CORINE, Urban Atlas). Since it is the material of a ground surface and not its use that has different properties (acoustic, spectral, and water absorption), we recommend that LC is prioritized. This is particularly important for the creation of open-access LC datasets that can be used for training AI tools.