GISdevelopment.net ---> AARS ---> ACRS 1989 ---> Land Use

Detection of Land Use Changes using Remote Sensing and GIS

Kam, Suan Pheng and Foo, Lay Kuan
School of Biological Sciences,
University Sains Malaysia, 11800 Penang, Malaysia



Abstract: A thematic map produced from digital classification of a SPOT multispectral subscene over a portion of south Johor, Malaysia, was brought into a GIS to assess classification accuracy using a land use map produced from aerial photo interpretation and ground checks. Percentage accuracy’s were found to be generally lower (by 5.3 % to 51.2 %) compared to accuracy’s assessed based on homogeneous areas test. Changes of land use were analyzed by GIS overlays of the 12974 and 1986 land use maps rather than with the interpreted SPOT data because of the generally low classification accuracies. The problems arising from analyzing changes between maps of different land use and land cover categories were pointed out.

Introduction
Remote Sensing techniques have proven useful for gathering information about the natural resources on a large-scale basis, such as for a whole country. The Malaysian Ministry of Agriculture has been using aerial photo interpretation for land use mapping for Peninsular Malaysia. With its aggressive land development policy, there is great need for timely information on the agricultural and forestry resources of the country. With the increasing availability of high resolution satellite imagery, and as the land information handling capabilities within the country is modernized, it is opportune to explore the potential of remote sensing and geographical information systems (GIS) in information gathering and updating of its resources. This paper reports an attempt to bring over a land cover map from classification of SPOT imagery into a GIS database and discusses a number of issues relating to reconciliation of land use maps of different times from different remote sensing sources.

Background
The land use ,maps of Peninsular Malaysia produced in 1966 and again in 1974 by the Malaysian Ministry of Agriculture were based on visual interpretation of aerial photos at 1:25,000 scale (Siew,1969: Wong,1976). Color maps as 1:126,720 scale were published for the two surveys. In 1980/81, aerial photos were flown at 1:40,000 to update the 1974 land use maps. The 1974 manuscripts were used as the base for land use changes exceeding 40 ha in size. By then, the Ministry had acquired a COMARC GIS for digitizing the interpreted maps and to serve as a data retrieval facility whereby maps of any selected area can be plotted on demand with area summaries.

It would be expected that the use of computer-assisted mapping would accompany or be accompanied by the use of computer-assisted techniques for processing of remote-sensed data. However, the use of remote sensing techniques other than aerialk photography for land use mapping is at present still experimental. A number of issues would have to be faced in making the transition from the conventional methods to the new, particularly in the context of the Malaysian situation. These Include :
  1. The utility of remotely-sensed data such as satellite imagery for land cover mapping, Particularly with respect to specificity and classification accuracy.
  2. Availability and timeliness of remotely-sensed data for a complete coverage of Peninsular Malaysia, if not the whole country.
  3. Registration with former maps for detection of land use changes and for related spatial analyses on the GIS.
The first of these issues is being actively researched by the various institutions in the country which have recently acquired image processing capabilities. The second issue would hopefully be facilitated with the impending functioning of the Malaysian Center for Remote Sensing for centralized acquisition of remotely-sensed data. The third issue, though minor in comparison to the first two in terms of research and data acquisition effort, is nevertheless important in view of a potentially growing user community with data needs for a wide range of GIS applications. This was borne out in the present study by the authors as part of the ASEAN-USAID Coastal Resources Management Project when faced with the task of reconciling land use maps of two different times at different levels of interpretation.

Objectives
The main objectives of this paper are :
  1. To determine the true classification accuracy of a land cover map produced by digital classification of a SPOT imagery by comparing it with a land use map interpreted from aerial photographs.
  2. To determine the extent to which the land use changes can be detected by comparison with an earlier (1974) and use map.
Study Site Description
The Study Site is located in the Kota Tinggi District of the state of Johor, the southernmost state of Peninsular Malaysia. The subscene used for this study is part of the large estuary of the Johor River which flows into the Straits of Johor separating Peninsular Malaysia from Singapore. The area now faces rapid development and land use changes. This study site is chosen because it is within the coastal zone which is currently under study in the ASEAN-USAID Resources Management Project (CRMP), in which the authors are involved. Hence there is a ready information base about the past landaus patterns and the future development plans for the area.

Materials and Methods
A thematic map showing the land cover of the Belungkor area was produced by digital image classification of a 1986 SPOT subscene using the ERDAS image processing system (IPS), as reported by Kam and Koay (1989). The SPOT scene was acquired on 19 June 1986. The image had been rectified and resample prior to classification so that it could be brought into a geographic database for the purpose of this study. The thematic map, which was in raster format, was imported into aspens (Spatial Analysis System), a microcomputer-based GIS (TYDAC, 1988). A number of transformations were carried out to convert the raster map file to the quad tree structure of SPANS and also to convert from the Geographical coordinate system of the ERDAS output to the Rectified skew Orthomorphic (RSO) map projection which is used in all Malaysian topographic (RSO) map projection which is used in all Malaysian topographic maps. The fit between the imported map and the base map with in the GIS was found to be good.

As part of the AEAN-USAID CRMP, a 1986 land use map was prepared by interpreting aerial photos flown in May, 1986 and through ground checks. A simplified land use classification scheme from that used by the Ministry of Agriculture (Wong, 1976) was employed, whereby some of the agricultural crop types with low occurrence were omitted. The land use map was drawn on a topographic base eat 1:63,360 scale. The map was digitized into the GIS, with 9 land use categories.

The 1974 land use map at 1:126,720 scale which was published by the Ministry of Agriculture (Wong, 1976) was also digitized into the GIS. For purpose of comparison with the later maps, the legend for this map was simplified by collapsing several crop type categories, leaving 8 land use classes within the study area.

These three map layers, i.e. the 1974 land use map, the 1986 land use map and the 1986 interpreted SPOT data (henceforth referred to as the “land cover map”), were used for subsequent overlay and other GIS analyses.

An area cross-tabulation between the 1986 land use map and use map and the 1986 land cover map was carried out to assess the true classification accuracy of the latter map. In order to display the agreement and disagreement in mapping between the two maps for the major land cover types, a matrix overlay was carried out.

Area cross-tabulation and matrix overlay were also carried out on the 1974 land use map and the 1986 land use map to determine changes in land use over 1974-1986 period.

Results and Discussion
The mismatch between land use and land cover categories for the three maps used presents a problem in comparing maps for change detection. Certain land use types cannot be readily interpreted using digital classification of SPOT data for two main reasons. One is because poor spectral reparability, such as between mining and cleared land; and between mixed horticulture and urban on the one hand and between mixed horticulture and rubber on the other. The land use classification used by the Ministry of Agriculture (Wong, 1976) emphasizes agricultural use, with numerous crops types which may not be spectrally separable for low spectral resolution data like SPOT and Landsite MSS. The other reason is the absence of spatial and contextual information in conventional digital classification techniques. Water as detected as a land cover category can be either sea, river, reservoir or aqua culture pond, the distinction of which can only be done based on context, shape and size.

On the other hand, certain land cover categories detected from image classification are not usually mapped in conventional land use maps, such as roads, cloud and cloud shadow . Hence some of the disagreements between maps are due to the lack of concordance of land use and land cover categories.

Table 1 shows an area cross-tabulation of the 1986 land cover map against the 1986 land use map. It is also a contingency matrix of the true classification accuracy of the SPOT data, based on the supervised, maximum likelihood classification which was carried out on the subscene. The values in the table are expressed in square kilometer rather than in pixels to give an idea of areas involved in each category. The “settlements” land cover category includes “urban” as well as villages, which are classed as mixed horticulture” in the 1986 land use map. The land use and land cover categories which correspond directly are those of rubber, oil palm, forest, cleared land, aqua culture and mangrove. The result of the matrix overlay illustrating the thematic correspondence between the two maps is show in Figure 1.

Table1 . Area cross tabulation of SPOT interpreted map us 1986 land use map
Row : Land cover from SPOT interpretation
Col : Land use from aerial photo interpretation
Area (km sq) Col % Urban Mining Mixed hort Rubber Oil palm Aqua culture Cleared Forest Mangrove Total
Road 0.0107 0.0090 0.0056 0.0310 0.0743 0.0006 0.0169 0.0231 0.0169 0.1009
3.42 16.67 0.06 0.40 0.62 0.05 0.40 0.31 0.07  
Cloud Shadow 0.0000 0.0000 0.0023 0.2168 0.2956 0.0231 0.1233 0.3308 2.1491 3.1407
0.00 0.00 0.26 2.80 2.48 1.92 2.93 4.43 9.25  
Settlements 0.1599 0.0231 0.2624 2.3431 2.1330 0.1284 1.7735 0.7971 2.0770 9.6925
51.08 42.71 30.78 30.23 17.87 10.66 42.08 10.68 8.94  
Cloud 0.0000 0.0079 0.0011 0.0642 0.0394 0.0186 0.0938 0.0310 0.1250 0.3345
0.00 14.58 0.13 0.83 0.33 1.54 2.22 0.41 0.54  
Rubber 0.0676 0.0073 0.2421 2.9073 2.5489 0.0327 0.4685 0.9927 0.8492 8.1112
21.58 13.54 28.40 37.51 21.36 2.71 11.12 13.30 3.65  
Aquaculture 0.0034 0.0017 0.0366 0.1045 0.2083 0.5882 0.1734 0.0180 0.9207 2.0541
1.08 3.13 4.29 1.35 1.75 48.85 4.11 0.24 3.96  
Water body 0.0000 0.0000 0.0963 0.0146 0.0000 0.0191 0.0028 0.0191 0.7151 0.8127
0.00 0.00 11.29 0.19 0.00 1.59 0.07 0.26 3.08  
Cleared land 0.0023 0.0039 0.0265 0.1244 0.1720 0.1616 0.9393 0.0912 0.2388 1.1628
0.72 7.29 3.10 1.61 1.44 13.42 22.28 1.22 1.03  
Forest 0.0017 0.0000 0.1481 1.1614 1.6555 0.0760 0.4049 2.5801 3.6534 9.6543
0.54 0.00 17.37 14.99 13.87 6.31 9.61 34.57 15.72  
Mangrove 0.0011 0.0000 0.0253 0.3353 0.4415 0.1557 0.1557 0.1824 2.5069 16.0211
0.36 0.00 2.97 4.33 3.70 12.93 4.33 33.58 53.27  
Total 0.3131 0.0541 0.8525 7.7500 11.9341 1.2039 4.2148 7.4645 23.2370 57.0240

The most serious misclassifications are the assignment of the mangrove class to forest pixels, and the assignment of forest and settlements classes to rubber and oil palm pixels. The mangrove-forest confusion is most evident in the transition zone between dry land mangrove and upland forest (see figure 1) where the species mix might have caused poor spectral separability of these vegetation types. The confusion among oil palm, rubber, forest and settlements shows that there is much spectral variability of these vegetation types over be scene which were not adequately accounted for by the homogeneous area training sets. The settlements category is a mix of both built up areas and village-type sundry cultivation of a variety of food and fruit trees which tend to have broad and mixed spectral characteristics.

The misclassifications of non-vegetation classes occur mainly among areas of overall high spectral reflectance. These features, such as dry aquaculture ponds, cleared areas and high density settlements, are not readily distinguishable by their spectral characteristics but would be easily differentiated based on context and shape, which are criteria not used in the digital image classification techniques employed for the processing of the SPOT data in this instance.

In general, there are much lower percentages of coincidence, indicating lower percent accuracy’s, than those determined using homogeneous test areas (Table 2), as this accuracy assessment covers the entire population of pixels in the scene, thereby accounting for the full spectral variability of land cover features. The added presence of cloud and shadow tends tot reduce accuracy’s either through confusion with “bright” features or by masking the actual underlying land use type.

Although the accuracies can still be improved upon by more rigorous classification methods like refinement of training sets, probability thresholding, and smoothing of the output map, this initial attempt shows that classification accuracies from homogeneous area tests can be deceptively impressive.

Table 2. Comparison of classification accuracies from two methods of assessment
Method of Assessment Land cover class
Rubber Oil palm Cleared Forest Aqua Mangrove
Homogeneous Area test* 63.8 87.8 20.4 39.9 82.7 94.6
Overlay with 1986 Land use map 37.5 36.6 22.3 34.6 48.9 53.3

Values in percentages
*From Kam and Loay (2989)

Figure 2 is the land use change detection map produced by overlay analysis of the 1974 and 1986 land use maps. The major changes have been from forest to rubber and from rubber to cleared land and to oil palm. The cleared land category is generally a transition to oil palm planting. There are no extensive conversion of forest land to oil palm within the study area., This indicates that most of the oil palm planting over the 1974-86 period had been replanting of rubber areas, with few remaining forest areas available for direct clearing and new oil palm planting. The map also indicates conversion of mangrove areas for aqua culture and for agriculture.

Because of the low classification accuracy of the SPOT interpretation, land use change detection was not carried out using the 1986 land cover map at this stage.


Figure 1. Thermatic Correspondence between 1986 Land Cover Map and 1986 Land Use Map



Figure 2. Land Use Changes, 1974 - 1986

Conclusions
The following are our main conclusions within the limited Arial coverage of this study:
  • There was a generally good registration of the rectified output from the satellite imagery on to the topographical map base over the study area despite the fact that the rectification was done using planar coordinates and the map base was in the Malayan RSO projection. This is perhaps helped by the equatorial position of this region where distortions of projecting the earth’s spherical surface on a plane are least.
  • The comparison of the land cover map from SPOT interpretation and the land use map from aerial photo interpretation allows an assessment of how well the classification represents reality/ The nature of confusion of the classification data suggest strongly the potential for improvement of computer-assisted techniques were combined with visual interpretation to take advantage of the high spatial resolution of SPOT data. While improved classification techniques will undoubtedly improve classification accuracy, our initial findings show that accuracy assessment using homogeneous areas tested can be over-optimistic about the accuracy of classification results.
  • Although the land use change detection was not done with the SPOT interpreted map, there are some issues which can be anticipated. One is the lack of correspondence between the a number of land cover. Classes which can be detected given the spectral characteristics of satellite data and the land use classes of the maps produced from aerial photo interpretation.
If satellite imagery were more widely used for land use monitoring and mapping in Malaysia in the future, there is a need to consider a level of land use/land cover classification scheme which is commensurate with the specificity of the satellite data and yet has sufficient correspondence with previous classification schemes to allow the integration of data from the past and the future for purposes like change detection.

Acknowledgement
We would like to acknowledge the ASEAN-USAID Coastal Resources Management Project for enabling us to use the data and the SPANS GIS for this study. Our thanks also go to the School of Biological Sciences for the use of the computer facilities and to University Sains Malaysia for permitting us to present this paper.

References Cited
  • Kam, S.P., koay. 1989. Evaluation of classification accuracy for land cover using SPOT multi spectral imagery. Paper presented at the 10th Asian Conference on Remote Sensing. November 23-29, 1989. Kuala Lumpur, Malaysia.
  • Siew K.Y.1969. Present land use survey of West Malaysia. Reports 1-9. Soil Science Division, Ministry of Agriculture. Kuala Lumpur.
  • TYDAC 1988. Spatial Analysis System Version 4.0. Volumes 1-3. Unpublished manuals.
  • Wond, I.F.T. 1976. The present land use of Peninsular Malaysia. Vol. 1. Ministry of Agriculture, Peninsular Malaysia.