Determination of Aggregated Areal Types from a Landsat-TM and ERS-1 based Land-use Classification for the Agglomeration of Basel/Switzerland

D. Scherer, U. Fehrenbach, E. Parlow, H.-D. Beha

In: Parlow, E. (ed.) Progress in Environmental Research & Applications. Balkema Publishers. Rotterdam

wb01626_.gif (272 bytes) Abstract
The main objective of this application-oriented project is to present maps that reveal the functional relations between climate and spatial structure of a heterogeneous region as represented by the Swiss Cantons Basel-Stadt and Basel-Land. In the scope of this project, the concept of areal types was developed to overcome the limitations of land-use classifications directly derived from satellite data by means of maximum-likelihood classifiers. It could be validated that areal types give a more realistic description of the complex structure of a highly urbanized region like Basel. Moreover, the concept allows an objective definition of the areal types, a prerequisite of an operational application.

wb01626_.gif (272 bytes) Introduction
In the framework of a regional climate study called KABA (KlimaAnalyse der Region BAsel), which started in January 1995, detailed information on the spatial structure of a region of approximately 50 x 30 km² has to be obtained both from satellite data and by means of a digital terrain analysis based on a DEM (Digital Elevation Model). Surface characteristics, mainly derived from remote sensing data, are to a great extend represented by areal types reflecting the physical and functional structure units of the region. All distributed layers of information are then combined with climatic and meteorological time series for further climatic analysis in order to generate maps displaying the functional relations between surface structure and elements of the regional climate.
In this paper, the methodology of the determination of aggregated areal types from a Landsat-TM and ERS-1 based land-use classification is demonstrated for the agglomeration of Basel in Switzerland. In this context, remote sensing operates as a highly significant source of information since no other method would have been able to provide spatially homogeneous data sets in that high qualitity prerequisite to the KABA project.

wb01626_.gif (272 bytes) The Concept of Areal Types
Land-use classification, based on remote sensing data, results in surface structure classes for each pixel of the satellite data. Landsat-TM or ERS-1 provide land-use information for pixels of 30 x 30 m². Despite of the problem of mixed pixels, which frequently lead to misclassifications, there are many situations that do not allow a direct determination of land-use by physical surface properties, where the satellite-based classification depends on. This holds particularly in urbanized regions, where complex patterns of building structures, construction works and areas with vegetation like village greens or allotments are present, to mention only a few examples.
To overcome this serious problem of incorporating satellite data into land-use studies, the concept of areal types was developed. A sharp distinction between pixel-based classes, briefly termed pixel classes, and aggregated areal types is introduced. Despite of their similar class names, they represent different forms of areal structure and function. Areal types are complex aggregates usually consisting of several pixel classes, which contribute to them in characteristic proportions, while the pixel classes themselves should be regarded as pure and homogeneous representatives of single 'traditional' land-use classes.


No. KABA pixel classes Characteristics
Urban and rural
Urban peripheral zones
High density settlements                   
Village cores                              
Building blocks with inner yards     
Serial houses                        
Urban and rural
Building complexes                   
Urban, partially in rural zones
Industrial buildings                 
Urban, partially in rural zones
Rural, partially urban
Meadows, pastures and orchards       
Rural, partially in urban peripheral zones
Arable lands                         
Rural, partially in urban peripheral zones
Water areas                          
Urban and rural
Asphalt and concrete surfaces        
Urban and rural
Sports fields, parks and village greens          
Urban and rural
Detached houses                      
Rural and in urban peripheral zones
Table 1. Definition and characteristics of the pixel classes, derived from a Landsat-TM and
	 ERS-1-based multi-sensor land-use classification (Beha et al. 1995) 

wb01626_.gif (272 bytes) Data Preprocessing
Beha et al. (1995) present a land-use classification for the agglomeration Basel/Switzerland. This data set, one of the results of the ESA-Pilotstudy ERSCLIP (ERS-1 CLImate Project) project, is based on one Landsat-TM and three ERS- 1 scenes from 1991 and 1992, processed by a multi-sensorial approach to combine the advantages of both systems while avoiding their specific drawbacks (for further information cf. to Beha et al. 1995). For the KABA test site, which is not completely covered by this classification, land-use data from a classification study for the whole REKLIP (REgio-KLIma-Projekt) region (Scherer et al. 1994) was combined with the ERSCLIP data set.
The combined land-use data layer is geocoded using the Swiss National Coordinate System with a grid resolution of 30 m. Subclasses, which had been introduced in ERSCLIP and REKLIP to overcome statistical problems of the maximum-likelihood approach, were put together in one pixel class since there are no physical or functional reasons of treating them individually within KABA.

wb01626_.gif (272 bytes) Areal Percentages of Pixel Classes
The nominal character of pixel classes restricts their computational usage seriously. Misclassifications caused by mixed pixels are an additional source of errors in subsequent calculations. Depending on the objectives of a project, the grid resolution of 30 m may not be required for an adequate treatment of land-use information. In such cases, it is extremely useful to compute areal percentages of pixel classes as input in further processing methods, since these data can be handled arithmetrically. Within KABA, the original 30 m satellite resolution is transformed into a grid resolution of 100 m by means of the areal percentages of each pixel class.

The areal percentage p of pixel class k in the grid element (m,n) of 100 m grid resolution can be determined from the original 30 m pixel classes c(i,j) by the following method:


where is sigma[c(i,j)k] is 1, if c(i,j) equals k, or 0 otherwise; (i,j)w(m,n) is the overlapping area between pixel (i,j) and grid element (m,n). This computation is carried out for each grid element and each pixel class for the Basel study site. This method finally results in 15 information layers with the additional advantage of a significant reduction of misclassifications. E.g. in case of mixed pixels caused by diagonal borders of land-use classes, the errors show the tendency to cancel each other, when integrated in the spatial domain.

wb01626_.gif (272 bytes) Rule-Based Classification of Areal Types
The next step was to develop a classification scheme starting with areal percentages of pixel classes and resulting in areal types. The concept of the determination of areal types was also designed to avoid subjective criteria as far as possible and therefore to enable a transfer of this method to other regions. In a supervised classification, the selection of training areas providing the statistical information for the classification algorithm is highly subjective. A better solution is to set-up a set of rules describing the typical composition of pixel classes within a certain areal type. These rules may solely use the areal percentages of single pixel classes, e.g. a 50 % threshold to ensure absolute majority of a certain pixel class (forestial areas are dominated by forests). Thresholds may be chosen with respect to the aims of the project and the desired accuracy demand. But rules may also combine the percentages of two or more pixel classes by 'AND' or 'OR' conditions, as it is necessary for most of the settlement types. This requirement is due to their higher degree of differentation, and usually results in a complex combination of pixel classes.
In general, the areal type a(m,n) is a function of the areal percentages (k)p(m,n) of the involved pixel classes:


The areal types have been validated by various methods, including maps, aerial photographs and field checks, not only by the authors themselves but also by regional planners not involved in the classification procedure.

No. Code Areal type Spatial characteristics
1 10 Forestial areas
  • Rural, partly urban.
2 20 Grasslands
  • Rural, partly in urban peripheral zones.
3 30 Water areas
  • Urban and rural.
4 40 Areas of arable land
  • Rural, partly in urban peripheral zones.
5 50 Sports fields, parks and urban greens
  • Urban and rural.
6 60 Horticultural areas
  • Urban peripheral zones.
7 70 Extensive railway areas
  • Urban, partly in the center of larger rural towns.
8 80 Low density housing
  • Rural and in urban peripheral zones.
9 90 Settlements of mixed structure
  • Urban and in rural village centers.
10 100 Urban housing
  • Urban.
11 110 Dense urban housing
  • Urban, mainly close to the city center.
12 120 Combined housing and industrial areas
  • Urban, partly in dedicated rural zones.
13 130 Commercial and industrial areas
  • Urban, partly in dedicated rural zones.
14 140 High density urban areas
  • Urban, mainly in the city center.
15 150 Extensive traffic areas
  • Urban and rural.

Table 2. Definition and characteristics of areal types, derived from the pixel classes by means of a rule-based classification.

wb01626_.gif (272 bytes) Conclusions
The following major conclusions can be drawn from this study:
Areal types provide a more realistic description of the structural setting of a region than pixel classes, since not only physical surface properties as seen from satellite, but also functional aspects of land-use are taken into consideration.
Starting from satellite-based land-use data, the methodology is operationally applicable for various regions due to the objective definition of areal types by rules not depending on a specific test site. The method may further be used for a more detailed differentiation of rural land-use classes.
Since areal percentages of all pixel classes serve as input in the determination of areal types, these percentages may also be used for other purposes relevant to environmental research or for regional planning affairs. Both aspects are addressed in this project.
Within the KABA project, areal types will be combined with DTM-based information to derive a com- prehensive set of regions with specific interrelations between the atmosphere and themselves in order to reveal their climatic significance. Regional climate maps with additional planning remarks (e.g. concerning ventilation and air pollution in critical parts of the region) will then be designed for further use in regional planning affairs.

wb01626_.gif (272 bytes) Acknowledgements
This study is part of the KABA project, which is funded by the Swiss Cantons Basel-Stadt and Basel-Land. Thanks are given to O. Schaub and K. Kamber from the swiss company Suiselectra, and to Dr. H. R. Moser from the Lufthygieneamt beider Basel.

wb01626_.gif (272 bytes) References
Allan, T.D. 1983. Satellite microwave remote sensing. New York: Halsted Press.

Beha, H.-D., E. Parlow & D. Scherer 1995. Land-use classification of the agglomeration of Basel/ Switzerland combining multi-sensorial satellite data from Landsat-TM and ERS-1. Proc. 15th EARSeL symposium and workshop Basel 1995. Rotterdam: Balkema.

Parlow, E. 1992. Klimarelevante Zusatzinformationen aus Satellitendaten. Ann. Meteorol. 28: 165- 167.

Parlow, E. 1994. Faktoren und Modelle für das Klima am Oberrhein - Ansätze und Ergebnisse des Regio- Klima-Projektes REKLIP. Geogr. Rundschau 3/1994: 160-167.

Parlow, E. (1996). The regional climate project REKLIP - an overview. Theor. Appl. Climat.: Vol. 53 (1-3), 3-7

Richards, J.A. 1986. Remote sensing digital image analysis: an introduction. New York: Springer Verlag.

Scherer, D., G. Bing, U. Heinzmann, E. Parlow & M. Wüthrich 1994. Problems and applications of a LANDSAT-TM based land use classification of the trinational region of the Upper Rhine Valley (Germany, France and Switzerland). In Vaughan, R. (ed.), Remote sensing from research to operational applications in the new Europe: 3-10. Budapest: Springer-Hungarica.

Scherer, D., E. Parlow & H.-D. Beha 1995. Roughness parameter derivation from ERS-1 and Landsat- TM satellite for the agglomeration of Basel/ Switzerland. In: Parlow, E. (Ed.): Progress in Environmental Research & Application. Rotterdam: Balkema.

Wüthrich, M. 1991. Modelling the thermal structures of Basle (Switzerland): A combined approach using satellite data and GIS-techniques. Proc. 11th EARSeL symposium Graz 1991: 298-304.

wb01337_.gif (904 bytes)