| GISdevelopment.net ---> AARS ---> ACRS 1991 ---> Poster Session 2 |
Modelling the development of
Asian cities using Airphotos and Remote Sensing data
ir C.A. de
Bruijn
ITC - Urban Survey Division, Netherlands
ITC - Urban Survey Division, Netherlands
Abstract
Urban models are used to simulate the development of cities in order to anticipate the needs for land, infrastructure and facilities. Modelling is an important planning tool but depends on the availability of accurate and detailed datasets on socio-economic aspects and interactions. Not in all countries such data are readily collected from air photos or remote sensing images, particularly urban landuse at various points in time.
A GIS enables analysis of the development process as an interaction between landuse changes, infrastructure and distribution of major activity nodes. In this way it is possible to establish and quantify relationships that model the spatial development rather well. Case studies on the development of Hyderabad in India and Bangkok in Thailand will be described to illustrate the potential of landuse based urban modelling.
Using existing knowledge from GIS systems and model - derived probabilities will make it possible to create time series of comparable urban development data based on routine remote sensing processing.
1 Introduction
As a contribution to rational procedures in the planning process models are important. Models quantify assumptions how an urban area develops and grows, not in isolation but in an integrated way: the assumptions mutually influence each other which may reinforce or reduce certain expected effects.
The arrival of PC's at planners' desks increase the accessibility of models as a planning tool. FOOT ( 1984) discussing some practical examples of modelling states that since computers are now readily available urban models can easily be developed and can provide good useful information to assist the decision maker.
Performance and spatial resolution of the model depend on quality and accuracy of the basic data. In many developing countries small area statistics may not be available or they could be highly unreliable and incomplete. Administrative data will not ( yet/) be accessible for modeling exercises and they may be also far from comprehensive.
Development of a class of models that is based on observable physical changes that can be interpreted from airphotos and/or satellites could be a way out of the data problem.
2. Development Probability, a landuse driven approach
The absence of good datasets in many cities and the difficulty in defining quantitatively "utility" when there is a large informal sector, has led to study possibilities for an approach based on landuse data as primary data source: the development probability model.
While most urban model are demand - side driven and use landuse at most as a capacity constraint development probability models are supply-side oriented. They derive future development mostly from geographical and locational aspects of the existing site. An early version of the model, developed for Limburg (Natherlands), is described in DE BRUIJN ( 1981).
Factor that influence suitability and hence development probability often have a strong distance decay function; the influence of proximity to a road or existing built-up area diminishes quickly. Approaches based on district data can describe only average values and usually do not have sufficient spatial resolution to capture the real interaction between the factors concerned. GIS systems offer spatial overlay capabilities and gridcell processing methods that enable to analyze spatial interaction.
3.The coeficient of prediction ( CP)
To compare the performance of different models the coefficient of prediction (CP) has been developed. The coefficient of prediction indicates how much of the actual landuse change has been predicted by the model and is based on how much of the study area is required to predict a certain amount of change. Assume a study area of 10,000 HA, including in 1970 a town of 2,000 HA. The town has grown to 3,000 Ha by the year 1980. There has been a change from rural to urban of 1,000 HA or 12.5% of (10,000-2,000) HA. A perfect prediction would label exactly 12.5% of the non-urban land as sensitive to change and only those areas that really changed. The CP should be 1. A prediction that would no predict anything would catch 10% of the change with 10% of the land and 50% of the changes with 50% of the land etc. In such a case the CP should be 0. Computation of the CP is described in details in DE BRUIJN ( 1991).
The following CP's have been obtained for various models included in this research:
Coefficient of prediction (CP) for rural/urban landuse changes
4. Model for Hyderabad, India
An extensive analysis for the city of Hyderabad has been undertaken by BHIDE ( 1988). He studied structure and growth pattern of the urban fringe in Hyderabad over a period of 15 years in order to develop a predictive model which should be useful to decision makers to formulate appropriate policies on the desired future growth. The study was carried out the USEMAP 4.0 software. The 1971 built-up area was delineated from the 1:20,000 topographical map. For 1985 air photos scale 1:20,000 have been interpreted.
The study area was 1600 sq. km. No digitizing facilities were available at the time of data collection and manual gridcell coding had to be used. A cell size of 500 m* 500 m was selected (the total area would be 90 cells). For each cell the percentage built ( developed ) and the majority landuse were recorded. The percentage development per 25 HA cell was measured and the analysis included (1) increase in the total number of cells with any development and (2) changes in percentage built of those cells that were already developed. Cells with 4 or more percent built area increased from 1984 to 2106, (+27%). The total built area in these cells increased from 17500 HA to 26800 HA, (+53%).
Based on available data, local knowledge and assumptions derived from questionnaires to local planners and decision makers a descriptive model was attempted, that includes 4 aspects:
- distance to core area
- distance to main road and rail
- areas served by water and sewage
- proximity to already developed area ( adjoining gridcells )
The selected model shows following results:
Low + very low probability areas contain about 65% of the study area but only 29% of all development. The high + very high probability is about 9% of the area and gets 30% of all development.
Work on the Hyderabad model was continued was to see if a more rigorous method could be used to arrive at an acceptable model. Relationship that Bhide considered as important were statistically tested in two ways (1) by comparing average values e.g. of all cells within a certain distance interval of a road, (2) by analyzing at the gridcell level, considering each gridcell as an independent observation.
The basic data "built 71" and "built 85" and the derived "change 71-85" were compared with (1) the concentration index for cells that are > 60% developed, (2) the distance to roads and ( 3) the distance to the core of the town.
A new model has been developed based on those factors ( C8)
The model can be formulated as
CLASS X,Y = ( DEV 71X,Y + niax ( COREDIST XY' DIST71BIGX,Y DISTHWX,Y) * EXCL X,Y
Where
CLASSX,Y dev probability class of cell x,y
COREDIST distance to core
DIST71BIG distance to cells > 60% developed
DISTHW proximity to highway.
DEV71 already devpd in 1971 or not
EXCL exclusion when not possible to develop
The new model had a CP of .58 showing a considerable improvement compared with the CP of 0.42 of the earlier model. Performance is good in core and first fringe. The second fringe ( particularly class 5) shows an underestimation. In this fringe there is quite a number of vacant cells that start development, but the model can not predict which cells and makes a lesser error by neglecting them all. In the periphery there is underestimation due to some large new planned industrial developments which are only partly predicted by the trend-based model.
The compactness of the development may be related to the transportation situation. The amount of private cars in Hyderabad is low and the availability of formal and informal sector public transport is limited. Such a situation puts a bonus on proximity to other development particularly for the informal sector.
5. Model for Bangkok, Thailand
Impact of spatial factors on the development of Greater Bangkok has been studied by PRASERTPUNT ( 1988 ) and DEN HAAN ( 1990 ) using delineations of the developed area in 1974 and 1985 obtained by photo interpretation. Presertpunt's study concentrates on estimating the residual capacity of Bankgoks districts given the existing plans for infrastructure development. DEN HAAN focuses more on the empirical relation between development and location of infrastructures. He concludes that there is a significant statistical relationship between the observed development of the urban area and the road network. USEMAP GIS-data base from these studies has been used to develop a probability model for Bangkok along similar lines as the model for Hyderabad. .
Data for Bangkok is less detailed than for Hyderabad. A 500 m* 500 m gridcell can either be developed or not, there is no indication for the intensity of development. The urban core has been defined in this case by using a measure for the amount of clustering of developed cells, i.e. the concentration index. All cells the core area. The concentration index is further used as a measure for proximity to already developed areas. The Bangkok development probability model has been formulated as
CLASS X,Y = max (CORPROXX,Y , RDPROX X,Y' GTOM[TPCX,Y
where
CLASS x,y dev probability class of cell x,y
CORPROX proximity to core
RDPROX proximity to roads
FRINPROX proximity to development fringes.
( RDPROX has only 2 classes: < 250 m and 250-750 m. This factor only plays a role in the peripheral zones when the other 2 factors are less than 2).
The CP obtained with the model is 0.78. Several other models have been tested e.g. using the airport and the harbour-industrial zone as major development attractions. As no better CP's were obtained they were not continued.
The performance of the model has been checked at the district level. All cells in a given probability class were given a % value equal to the observed development for that class between 74 and 85. The resulting total development was computed per district and compared with the actual total development in 85. Only probability classes 3,4,5,6 were considered. For the total area the amount of development was about 13% underestimated ( as the development that occurred in the probability classes 0.1.2 was not predicted).
The correlation between model development and actual development in 85 has a r2 of 0.98., between development 74 and development 85 r2 was also 0.98. The main deviations are concentrated in the three large district east of the city, containing the airport bangkapi etc. The remaining 23 districts have an r2 (model.actual 85)= 0.943 while r2 (actual 74/actual 85)= 081. For those districts the model is a better predictor than the previous development state.
ACP of 0.78 is rather high and suggest that the development of Bangkok is a process of continuous agglomeration, not influenced much by planning or other outside factors. Two thirds of all changes can be explained by agglomeration and core fringe effects. In the 10 years of the study-period 52% of This area has been changed from undeveloped to developed. Outside the fringe areas the proximity to roads explains 75% of the remaining changes. However, only 11% of the areas close to roads had been developed during the study period.
The model still may be improved when a map could be made of areas to be excluded because of constraints to development. However, the site suitability obstacles (flooding) hardly seem a deterrent to development in Bangkok and planning seems too weak to have much impact. Information on land property might hold some clues. The influence of roads is less than had been expected. While some roads show a lot of roadside development, others, particularly on the weastbank does not seem to trigger the same effect. Data on traffic intensity and traveltiems may lead to a better fit but are less readily available.
6. Conclusion: Use of Urban Models in Remote Sensing
The studies on Hyderabad and Bangkok show that effective predictive models are possible with datasets obtainable by Remote Sensing using the development probability method described in this paper. The models may have many applications in planning (see e.g. DE BRUIJN, 1990). They can be useful also in computerized delineation of urban areas from remote sensing images.
Classification of urban areas is difficult as there is no direct relation with spectral classes, given the mixture of textures that all belong to what should be delineated as urban. When there are a priori indications of the probability that an area may have changed to urban the accuracy of the classification process can be improved (STRAHLER, 1980).
Using existing knowledge from GIS systems and model derived probabilities along the lines described here will make it possible to create time series of comparable data based on routine remote seing processing. Provision should be made to check apparent inverse changes (urban to non-urban) and the older data set should be corrected when necessary. A limited number of doubtful cases may have to be field-checked on the ground or by visual inspection from the air, complemented with do-it-yourself oblique airphotos.
Based partly on probability models the data may not quality for an "early warning" system. The humble beginnings of new development that sometimes form the start of new spatial trends are unlikely to be observed as the model will only look for development where it is likely to occur. Additional information collection in combination with regular aerial photography remains necessary for this purpose (DE BRUIJN, 1983).
The described modeling method holds promises for the routine production of more frequent time series about the spatial growth of urban areas. The complex process of urbanization requires systematic and regular data to monitor and analyze urban development. Good time series form remote sensing sources that enable comparative studies between cities may make a valuable contribution.
References
- BHIDE, A.V. (1988)
Spatial structure of urban fringe: Case study of Hyderabad, India.
Unpublished MSc thesis, ITC, Enschede, The Netherlands. - BRUIJN C.A. DE (1981)
Analyzing urban development issues using sequential landuse data and a geodatabase: A case study from Limburg, The Netherlands
Paper Harvard Comp Graphics Week, Boston, july 26 - 31 - BRUIJN C.A. DE (1983)
Urban Ariphoto Interpretaiton in a Geodataprocessing Environment.
In : Proc 4th ACRS, Colombo 1983, p Q-1-1: Q1-10 - BRUIJN C.A. DE (1990)
Modelling urban landuse changes in datapoor situations.
Paper UNCRD Expert Meeting Regional Planning, Kuala Lumpur 24- 28 September 1990 - BRUIJN C.A. DE (1991)
Spatial factors in urban growth: towards GIS models for cities in developing counties.
To be published in ITC Journal 1991/4. - FOOT, D. (1984)
Computer models of cities; an introduction.
In Cities, August 1984 p. 469-479. - HAAN, R.N. DEN (1990)
Influence of infrastructure on spatial development. A case study on Bangkok, Thailand.
In : Proc EGIS 90 (Vol. 1) p. 418-427. - PRASERTPUNT, V (1988)
Using Geographical Information System (GIS) in Analyzing the influence of Infrastructure on the pattern of development in the greater Bangkok Area; In: Proc ACRS (1988), Bangkok p S-8-1: S-8-9
- STRAHLER, A.H. (1980)
The use of prior probabilities in maximum likelihood classification on remotely sensed data.
In : Rem Sens of Env (10) p 135: 163