Attribution of local climate zones using a multitemporal land use/land cover classification scheme

Andreas Wicki; Eberhard Parlow

doi:10.1117/1.JRS.11.026001

3 April 2017 Attribution of local climate zones using a multitemporal land use/land cover classification scheme

Andreas Wicki, Eberhard Parlow

Author Affiliations +

Journal of Applied Remote Sensing, Vol. 11, Issue 2, 026001 (April 2017). https://doi.org/10.1117/1.JRS.11.026001

Abstract

Worldwide, the number of people living in an urban environment exceeds the rural population with increasing tendency. Especially in relation to global climate change, cities play a major role considering the impacts of extreme heat waves on the population. For urban planners, it is important to know which types of urban structures are beneficial for a comfortable urban climate and which actions can be taken to improve urban climate conditions. Therefore, it is essential to differ between not only urban and rural environments, but also between different levels of urban densification. To compare these built-up types within different cities worldwide, Stewart and Oke developed the concept of local climate zones (LCZ) defined by morphological characteristics. The original LCZ scheme often has considerable problems when adapted to European cities with historical city centers, including narrow streets and irregular patterns. In this study, a method to bridge the gap between a classical land use/land cover (LULC) classification and the LCZ scheme is presented. Multitemporal Landsat 8 data are used to create a high accuracy LULC map, which is linked to the LCZ by morphological parameters derived from a high-resolution digital surface model and cadastral data. A bijective combination of the different classification schemes could not be achieved completely due to overlapping threshold values and the spatially homogeneous distribution of morphological parameters, but the attribution of LCZ to the LULC classification was successful.

1. Introduction

Before the use of air conditioning to counteract unpleasant effects of enhanced heat stress, people used to construct their housing adapted to the local climate conditions. This kind of urban architecture can still be observed in old Arabic medina districts¹ or in the historical centers of European cities. Modern urban design, however, often ignored its role in the urban climate system due to improved heating and cooling systems. During periods of rapid climate change,² these adaptations have to be reinvented, especially in the moderate and temperate climate zones.

Historical centers define usually the core of most European or North African cities with different morphology due to climatic differences, topographic adaption, demographic and cultural development, and destruction through totalitarian regimes³ or wars. Comparing buildings outside the old medieval core built over the last 200 years during rapid urban growth (i.e., Wilhelminian or Gruenderzeit era) with suburban neighborhoods, urban morphological characteristics like street width, mean building height, or building density differ substantially. These differences are well observable in satellite imagery with subkilometer resolution or in orthophotos. The major challenge of land surface analysis is to translate these visible differences into comprehensible quantities. Although satellite sensors usually measure data in a spectral range beyond the visible wavelengths, and therefore see many more differences than the human eye, they cannot classify surface structures automatically. Different approaches of characterizing these measurements have been developed: One possibility is to work with different indices to identify and quantify vegetation,⁴^,⁵ soil,⁶ buildings,⁷^,⁸ and water bodies.⁹ Spectral mixture analysis¹⁰^–¹² and support vector regression,¹³ to estimate the composition of satellite imagery, are other popular methods. In this study, a classical approach using a maximum-likelihood classification based on ground truth data [regions of interest (ROIs)] is used to subdivide the city into different levels of densification with various characteristics concerning thermal behavior, air movement, morphological parameters, vegetation cover, and many more. This information is crucial for urban development and should be available to planners in a useful manner to enable the distribution of knowledge about basic urban climate aspects among the different disciplines.¹⁴^–¹⁶ Furthermore, LULC analysis data are important for urban climate scientists as an input for models and urban climate studies.

Another issue of land surface analysis is the comparability of different classification schemes throughout different cities around the world.¹⁷^–²³ Stewart and Oke¹⁷ have tried to overcome this heterogeneity concerning land cover analyses with their local climate zones (LCZ) classification scheme. The major thinking behind this approach is the definition of general zones within cities in different countries worldwide to compare the climatic behavior. This can be very useful, for example, to characterize the urban environment of a meteorological measurement station. The idea of subdividing cities into “climate zones” has a long tradition in urban climatology. Weischet introduced his Baukörperklimatologie (building-complex climatology) to group similar urban structures and investigate their influence on the local climate of Freiburg im Breisgau in the 1970s.²⁴^,²⁵ Scherer et al.²³ defined “climatopes” and Fehrenbach²⁶ so-called “areal types,” which already considered fractions of different land cover in a satellite pixel.

The LCZ approach claims to be universal and applicable in cities all over the world. It is based on morphological parameters like imperviousness (impervious surface fraction and pervious surface fraction), sky view factor, height of roughness elements (buildings and vegetation, here referred to as mean building height), building to surface fraction, and aspect ratio. This scheme is used by many scientists in current research as a common reference frame and has a certain impact on urban climate science, as it is frequently applied by the urban climate community.¹⁸^,²⁷^–²⁹ But many people overestimated the fact that the scheme does not cover the heterogeneity in an old European city as easily as the chessboard-structure of a planned North American city. Using the thresholds of the LCZ scheme in a strict manner results in many unclassified areas and only in a few different LCZ classes.²⁹ Altering the thresholds on the other hand would reduce the comparability, which is one of the major advantages of the LCZ classification. Often, the only way to generate comprehensive LCZ is to map them manually. Nevertheless, the LCZ scheme offers the possibility to compare different parts of different cities with trenchant distinctions representing the heterogeneous thermal behavior within an urban environment.

To use the benefits of the LCZ scheme and apply them to a classical LULC classification, a method to bridge the gap between those different approaches has been developed. The following sections describe the workflow of this simple way to merge different classification schemes.

2. Materials and Methods

2.1.

Site Description

The city of Basel is located in the southern part of the Upper Rhine Valley in the trinational border region France–Germany–Switzerland and is the capital of the Canton Basel-Stadt (BS). The city structure is characterized by the river Rhine, whose channel undergoes an almost 90 deg bend during its passage through the city to enter the upper Rhine valley in the north (Fig. 1).

Fig. 1

Overview and location of the study area (map extent) with a topographic map based on the Aster GDEM data set including the borders of the Canton Basel-Stadt (white) and the surrounding countries (black) as a vector overlay. The overview map is created with Natural Earth data.³⁰

Basel is well known for its historical old town center located on the south bank of the river with a minor part on the north bank. Large industrial areas hosting big pharmaceutical/chemical companies along the river are sited at the eastern and northwestern borders of the city. Compared to other Swiss cities, Basel (250 m a.s.l.) has a relatively mild climate with record high summer temperatures. The annual mean temperature is 10.5°C with an annual precipitation of 842 mm according to the suburban weather station Basel/Binningen maintained by the Federal Office of Meteorology and Climatology (MeteoSwiss) for the reference period 1981 to 2010.³¹ Although the city is located between the mountains of Jura, Vosges, and Black Forest providing the city with fresh air,²³ the distinct nocturnal urban heat island causes severe heat stress during major heat waves.³²^,³³

The investigation area contains the metropolitan area of Basel with $\sim 730,000$ inhabitants in total including the bordering suburbs located in the Canton Basel-Land (BL), the German cities Weil am Rhein and Lörrach, the French city Saint-Louis and the city of Basel.¹⁵ The north to south and west to east extent reaches 21 by $20 {km}^{2}$ , respectively, spanning from the upper left 47°38’35”N 7°29’17”E to the lower right 47°27’25”N 7°45’32”E.

2.2.

Data

For this study, Landsat 8 data acquired between June and August 2013 to 2015 with a minimum cloud and haze cover is used from the Landsat paths $195 / 196$ and row 27. Seven cloud free scenes with a comparable sun elevation were selected as an input for the classification algorithm.

Morphological parameters are estimated by rasterizing a digital surface model (DSM) provided by the administration of BS³⁴ updated with LIDAR derived tree information and vector data of the official cadastral survey to a 1-m raster grid.³⁴ Typical for an international border zone, these data are only available for BS (white polygon in Fig. 1), omitting the French, the German, and the BL parts (Table 1).

Table 1

Raster and original vector datasets used in this study with corresponding acquisition or update time.

	Dataset	Acquisition date/last update
Raster	Landsat 8 OLI	05 June 2013, 08 June 2014, 17 July 2014, 11 June 2015, 04 July 2015, 05 August 2015, 30 August 2015
Vector (origin)	Administrative boundaries	Downloaded 23 February 2016
	Official cadastral survey	Downloaded 04 March 2016
	DSM	Validity 20 March 2009
	LIDAR-derived trees	Validity 19 November 2011
	Communal boundaries	Downloaded 26 January 2016

2.3.

Methods

2.3.1.

Land use/land cover

The LULC map is created using multitemporal Landsat 8 images in the VIS, NIR, and SWIR range. Additionally, the data are resampled to 15 m using the Gram–Schmidt pan-sharpening algorithm. Therefore, a higher spatial resolution panchromatic image is merged with several lower resolution bands using a statistical procedure.³⁵ Especially considering the sharpness, Gram–Schmidt pan-sharpening is known for generally high quality results.³⁶ Although theoretically pixel-based classifications are not applicable on pan-sharpened images, particular tests have shown higher accuracy compared to the 30-m original data (see Sec. 4.1 for more information).

The LULC analysis is based on a supervised classification using ROIs as training areas for a maximum-likelihood algorithm.³⁷^,³⁸ After several runs and refining of the training areas to reduce the unclassified and false classified pixels, the same ROIs are used to calculate the maximum-likelihood for six other Landsat 8 images covering the same extent and satisfying the criteria (no clouds, no haze, high sun elevation, and comparable vegetation cover). In a next step, the classification images are stacked and combined expecting that not all pixels in the various Landsat 8 images were classified in the same way. Thereafter, every individual pixel in the final product is the result of the most frequently occurring class (i.e., modal value) of the specific $x - y$ -location in the different classification layers among this multilayer stack. A similar approach was used before to successfully create complete coverage land cover maps with improved accuracy of areas with high cloud frequency.³⁹

Unclassified pixels are omitted and filled with the class-majority of the neighboring pixels. This is done first in a $3 \times 3$ surrounding and, if unclassified pixels still occur, in a $5 \times 5$ surrounding.

This procedure is applied only for the built-up classes and the water class, where no—or only minor—changes are expected within the two-and-a-half-year period. Due to the heterogeneity of the surrounding landscapes and the urban material, 25 different ROIs and resulting classes are determined. Finally, these classes are grouped reasonably in five different nonurban and six built-up classes due to the structures of the city and their natural surroundings (Table 2). Most of the grouping concerned the classes forest, water, agriculture green, and agriculture yellow, which represented a large number of the 25 ROIs.

Table 2

Description of the final LULC class compositions considering real world conditions after combination and filtering.

Class	Description
Dense Urban	Preindustrial old town development, small houses, roof tiles, crooked and narrow streets, minimum vegetation cover
Urban	Wilhelminian housing with regular three to five story buildings and backyards with greenspace
Suburban	Detached/semidetached houses with gardens, two to three story buildings
Urban Garden	Allotment gardens, cemeteries
Rail/Road/Concrete	Large roads, open concrete space, train stations, bridges, large railways, railway sidings
Industry/Commercial	Manufactories, metal roofs, industrial neighborhoods, port, construction sites, artificial sports ground, pharmaceutical/chemical industry
Agriculture Yellow	Mature crops, rape
Agriculture Green	Meadows, growing crops, maize, urban greenspace
Vineyard/Shrub	Vineyards, shrubs, small urban trees, bushes, medium-sized vegetation
Forest/Plantation	Deciduous and coniferous forests, agricultural plantation, urban trees
Water	Lakes, ponds, pools, rivers, dredging lake

2.3.2.

Morphological parameters

The morphological parameters are estimated using GIS data provided by the administration of BS. Therefore, the data are spatially limited by the borders of BS (see Fig. 1 for orientation).

The sky view factor is calculated using the urban multiscale environmental predictor (UMEP), an extension to QGIS provided by Lindberg et al.⁴⁰ To reduce computation time, the building layer of the vector DSM is rasterized to 3 m. Urban trees, received from LIDAR data provided by the administration of BS, are added for the sky view factor calculations as well. The values of roof and vegetation pixels are masked to obtain the ground-level sky view factor product.

Mean building height is computed using the vector DSM rasterized to 1-m cell size. Averages omitting the ground pixels, but including trees, are calculated on a 15-m grid. The same DSM is used to calculate the building to surface fraction, which represents the ratio of building to nonbuilding pixels in a 15-m grid environment.

Aspect ratio is computed using an approach described in detail by Lindberg et al.⁴¹ Using the 1-m raster DSM, the Euclidean distance from each pixel to the nearest wall is calculated. A local maximum algorithm filters the values representing the half-street width (HSW). To receive the corresponding building height to the particular street canyon, the 1-m raster DSM is adjusted using Voronoï polygons ( ${MBH}_{V}$ ) and divided by the HSW as follows: $a spect ratio = \frac{{MBH}_{V}}{2 * HSW}$ .

Impervious surface fraction and pervious surface fraction are estimated using GIS vector data containing the surface coverage based on a cadastral map of BS. Therefore, percentages of sealed surfaces (including streets, railways, pedestrian ways, motorways, etc.) and nonsealed surfaces (bare soil, parks, forests, water, etc.) are estimated on a 1-m grid. Similar to the other urban morphology parameters, the impervious surface fraction and pervious surface fraction are resampled to 15 m. All rasterizing is done using the Landsat classification image as a geometrical reference to ensure spatial congruence. An overview showing the calculated morphological parameters within the extent of BS is presented in Fig. 2.

Fig. 2

Morphological parameters [i.e., (a) sky view factor, (b) aspect ratio, (c) building to surface fraction, (d) mean building height, (e) impervious surface fraction and (f) pervious surface fraction] of Basel-Stadt with 15-m resolution and the coordinates in UTM-32N.

2.3.3.

Combination of land use/land cover and morphological parameters

LCZ are defined by morphological characteristics and the corresponding thresholds (Table 3). Ideally, a supervised classification based on remote sensing data should result in LULC classes with different morphological behavior throughout different classes.

Table 3

LCZ thresholds by Stewart and Oke.17 Only zones 1 to 10 are shown (built-up zones) and the terrain roughness class criteria is omitted.

Local climate zone	Sky view factor	Mean building height	Building to surface fraction	Aspect ratio	Impervious surface fraction	Pervious surface fraction
LCZ 1a	0.2 to 0.4	$> 25$	40 to 60	$> 2$	40 to 60	$< 10$
LCZ 2a	0.3 to 0.6	10 to 25	40 to 70	0.75 to 2	30 to 50	$< 20$
LCZ 3a	0.2 to 0.6	3 to 10	40 to 70	0.75 to 1.5	20 to 50	$< 30$
LCZ 4b	0.5 to 0.7	$> 25$	20 to 40	0.75 to 1.25	30 to 40	30 to 40
LCZ 5b	0.5 to 0.8	10 to 25	20 to 40	0.3 to 0.75	30 to 50	20 to 40
LCZ 6b	0.6 to 0.9	1 to 10	20 to 40	0.3 to 0.75	20 to 50	30 to 60
LCZ 7c	0.2 to 0.5	2 to 4	60 to 90	1 to 2	$< 20$	$< 30$
LCZ 8c	$> 0.7$	3 to 10	30 to 50	0.1 to 0.3	40 to 50	$< 20$
LCZ 9c	$> 0.8$	3 to 10	10 to 20	0.1 to 0.25	$< 20$	60 to 80
LCZ 10c	0.6 to 0.9	5 to 15	20 to 30	0.2 to 0.5	20 to 40	40 to 50

^aLCZ 1-3: compact high-rise, compact midrise, compact low-rise.

^bLCZ 4-6: open high-rise, open midrise, open low-rise.

^cLCZ 7-10: lightweight low-rise, large low-rise, sparsely built-up, heavy industry.

With the combination of the morphology and LULC classes, the LCZ thresholds, as defined by Stewart and Oke,¹⁷ can be used to characterize the land cover classification.

It needs to be mentioned that LCZ are not intended to be applied on a per pixel scale. Stewart and Oke describe a minimum radius of 300 m as reasonable.¹⁷ Due to the small-scale characteristics of a typical European city, the proposed radius is not able to depict this heterogeneity. Therefore, the 15-m resolution is retained until potential aggregations are presented at the end of Sec. 4.4.

The workflow and basic steps of the approach are summarized in Fig. 3 as a simple flow chart.

Fig. 3

Flow chart including the basic steps during the data processing.

3. Results

3.1.

Accuracy Assessment

The overall accuracy of the classification after the combination and the frequency analysis, estimated using the input ROIs within a confusion matrix, is 83%. The highest amount of incorrectly classified pixels within the ROIs is found in the Urban class. This can be explained with the heterogeneous building types between the historic city center and the suburbs or construction dynamics. The other urban classes reveal accuracies of 76% (Suburban) to 91% (Urban Garden). The two separated industrial classes show substantial overlap with only 56% of Industry/Commercial class pixels and 29% of Rail/Road/Concrete class pixels in the Industry/Commercial ROI field. The latter showed high purity within its ROI with 93% overlap between ground truth ROI and classification. These two classes are nevertheless separated due to the large differences in building to surface fraction and aspect ratio, which is important for roughness class estimation and urban climate studies. The nonurban classes (agricultural, vineyards, forests, and water bodies) show accuracies of almost 100% with Vineyard/Shrub as the exception. This can be explained with the large number of limestone walls and narrow streets within vineyards, misclassified as built-up areas.

Besides the occurrence matrix, visual interpretation using expert knowledge was performed carefully through the whole scene.

To test the benefit of the multitemporal method, the same quality assessment is performed for all individual classification layers and for a second test run without the pan-sharpening. All seven input classifications showed a lower overall accuracy compared to the combined classification with a mean value of 79%. Comparing the pan-sharpening approach with the original 30-m resolution classification, an improvement of 1% in overall accuracy is achieved using the sharpened images. Nonetheless, in most of the built-up classes, the accuracy is improved by up to 4%. Only the class Industry/Commercial revealed better results without the pan-sharpening approach.

Different ways of using the multitemporal data and panchromatic channel (without pan-sharpening) were tested as well but revealed many more unclassified pixels and are, therefore, not further investigated. The Kappa coefficient is not mentioned in this accuracy assessment due to the findings of Pontius and Millones.⁴²

3.2.

Land Use/Land Cover Classification

The LULC map shows high levels of detail and, due to the multilayer approach, high robustness. Detailed structures like bridges, airport runways, motorways, or small urban parks are clearly visible (Fig. 4). The distinction between built-up, water, forest, and other natural surfaces worked almost perfectly due to the use of multiple classification layers and therefore a minimization of misclassification. The inner-urban discrimination is much more difficult, due to mixed pixels and the large heterogeneity within urban environments. Nevertheless, the classification map represents the distribution of different urban built-up types with high accuracy.

Fig. 4

LULC map of Basel and its surroundings using the modal value of seven individual classifications from 2013 to 2015. The coordinate system is UTM-32N and the pixel resolution is 15 m. The national borders are represented by black lines and the BS border by a white line; a shaded relief overlay using resampled SRTM data depicts the topography. The vegetation represents an overpass of 30 August 2015.

3.3.

Morphological Parameters

Histograms describing the frequency distribution of the morphological parameters for every LULC class are shown for the sky view factor [Fig. 5(a)] and the mean building height [Fig. 5(b)]. The histogram displays comparable sky view factors in the urban classes, with increasing tendencies among decreasing urban density. In contrast to the other urban classes, almost all Dense Urban (i.e., old town) pixels have sky view factors below 0.75. These results have been anticipated, but they also show that many of the classes in reality cover a wide range of morphological parameters, which lead to ambiguous assignments of classes.

Fig. 5

(a) sky view factor including vegetation and omitting roofs and (b) mean building height including vegetation and omitting streets based on a vector DSM of Basel-Stadt for each built-up class.

The mean building height distribution represents the urban structures with medium high buildings in the Urban Dense class, a bimodal distribution due to the typically three to five story buildings in the Urban class and lower mean building height in the Suburban class. The heterogeneous structures within the classes Industry/Commercial and Rail/Road/Concrete with large streets and tower buildings alternated by flat construction halls are represented by the sky view factor (widespread values between 0.2 and 0.9) and the bimodal distribution in the mean building height histogram. The Rail/Road/Concrete class also includes train stations, bridges, and large concrete roofs, such as parking flats. Therefore, the building height distribution also contains values far above street level for this class, which might be unexpected.

Spatial extent, plus mean and standard deviation ( $σ$ ) of sky view factor, mean building height, building to surface fraction, aspect ratio, and impervious surface fraction are shown in Table 4. The spatial distribution is dominated by the Forest/Plantation class with almost 29%. This class is located mainly on the surrounding hills and provides the city and agglomeration with fresh and cool air due to nocturnal cooling and therefore has an important function for the urban climate system.⁴³

Table 4

Description of the LULC analysis with spatial extent (km2 and %) and class specific morphological surface parameters (built-up classes only). The sky view factor is calculated including trees and omitting roofs and the mean building height contains only buildings and trees, without ground pixels. All parameters are based on a resampled DSM (1 m and 3 m for sky view factor) including the Canton Basel-Stadt only.

Class	Total area		Sky view factor (±σ)	Mean building height (±σ) [m]	Building to surface fraction (±σ) [%]	Aspect ratio (±σ)	Imperv. surface fraction (±σ) [%]
Class	[km2]	[%]	Sky view factor (±σ)	Mean building height (±σ) [m]	Building to surface fraction (±σ) [%]	Aspect ratio (±σ)	Imperv. surface fraction (±σ) [%]
Dense Urban	3.1	0.7	$0.29 \pm 0.17$	$15.3 \pm 6.5$	$46.9 \pm 34.0$	$1.22 \pm 0.86$	$46.2 \pm 33.4$
Urban	15.4	3.6	$0.31 \pm 0.16$	$12.8 \pm 5.5$	$32.3 \pm 32.3$	$0.76 \pm 0.64$	$41.2 \pm 33.8$
Suburban	63.8	15.2	$0.36 \pm 0.18$	$10.3 \pm 5.2$	$22.0 \pm 28.2$	$0.46 \pm 0.43$	$26.3 \pm 29.5$
Urban Garden	14.7	3.5	$0.57 \pm 0.21$	$7.6 \pm 5.2$	$9.3 \pm 17.2$	$0.21 \pm 0.26$	$18.6 \pm 25.5$
Rail/Road/Concrete	31.7	7.5	$0.54 \pm 0.24$	$14.8 \pm 8.0$	$22.4 \pm 32.5$	$0.71 \pm 0.77$	$60.5 \pm 37.6$
Industry/Commercial	17.3	4.1	$0.54 \pm 0.22$	$14.4 \pm 9.4$	$38.5 \pm 40.3$	$0.86 \pm 1.06$	$50.1 \pm 39.8$
Agriculture Yellow	45.9	10.9	—	—	—	—	—
Agriculture Green	74.1	17.6	—	—	—	—	—
Vineyard/Shrub	27.3	6.5	—	—	—	—	—
Forest/Plantation	120.6	28.6	—	—	—	—	—
Water	7.0	1.7	—	—	—	—	—

Although essential for urban dwellers and for the cityscape with the historic old town as an important venue, Dense Urban as the smallest class covers only $3.1 {km}^{2}$ of the whole study area (0.7%). Suburbs dominate the built-up classes, but the industry classes—due to large industrial zones and extensive railway tracks—are also well represented in the investigation area.

3.4.

Combination of Land Use/Land Cover and Morphological Parameters

The combination of the different classification schemes was done using the average morphology of every class and the threshold values given by the definition of Stewart and Oke¹⁷ introduced in Table 3. Most class values fit in several LCZ classes, due to overlapping threshold values, as shown in Table 5.

Table 5

Relation between the LULC classification and LCZ from Stewart and Oke by applying the mean morphological parameters for each class to the LCZ thresholds.17 The numbers are the respective LCZ types that fit to the given thresholds.

Class	Sky view factor	Mean building height	Building to surface fraction	Aspect ratio	Impervious surface fraction
Dense Urban	1, 3, 7	2, 5, 10	1 to 3, 8	2 to4, 7	1 to 6, 8
Urban	1 to 3, 7	2, 5, 10	4 to 6, 8	2 to 4	2 to 6, 10
Suburban	1 to 3, 7	2, 5, 10	4 to 6, 10	5 to 6, 10	2 to 6, 10
Urban Garden	4 to 5	3, 6, 8 to 10	$\sim 9$ a	8 to 10	5, 7, 9
Rail/Road/Concrete	4 to 5	2, 5, 10	4 to 6, 10	5 to 6	1 to 6, 8
Industry/Commercial	4 to 5	2, 5, 10	4 to 6, 8	2 to 4	1 to 6, 8

^aNearest LCZ value applied.

In the next step, the above table is summarized and evaluated. Figure 6 shows this evaluation with values ranging from 0 (= no matches) to 5 (= fitting to all criteria). Only one class (i.e., Rail/Road/Concrete) matched with exactly one LCZ (i.e., open midrise or LCZ 5) in all criteria. Urban and Urban Garden are the only other classes with explicit matching (four matches) to compact midrise (LCZ 2) and sparsely built (LCZ 9), respectively. The other classes have two matching zones: Dense Urban fits to compact midrise and compact low-rise (LCZ 3), Suburban to open midrise and heavy industry (LCZ 10), and Industry/Commercial to open high-rise (LCZ 4) and open midrise (LCZ 5).

Fig. 6

Affiliation of LULC classes to the LCZ by Stewart and Oke¹⁷ according to the LCZ thresholds. All matches are summed up ( $maximum = 5$ , dark gray) and the individual boxes are colored according to the number of matches.

4. Discussion