Land Use/Land Cover Change
Detection in the Chiang Mai Area using Landsat TM Somporn
Sangavongse Abstract Department of Geography and Environmental Science. Monash Unversity, Clayton, Victoria 3168, Australia The rapid growth of Chiang Mai city in Northern Thailand has resulted very rapidly and have had an adverse effect on the environment. and therefore multi-temporal Landsat TM imagery for monitoring land cover change has proven to be the best tool in this study. Suitable change detection techniques were developed for the Chiang Mai area by taking into account its physical and cultural conditions, there by optimising use of information in the land cover maps Results showed that forested areas have decreased by about 29% during 1988 to 1991. while agricultural lands and built-up areas have increased by about 5% and 26% respectively. The image processing system used in this study was micro BRIAN. 1. Introduction Chiang Mai. the second largest city in Thailand, is located approximately 700 kilometers to the north of Bangkok. Its surroundings are experiencing much change in land use and land cover. Due to the rapid growth urban and suburban/industrial land is sharply increasing. The expansion of the city has resulted not only in the depletion of natural resources, but the deterioration of the environment (Surarerks, 1992). Agriculturally productive land and forest lands have been converted into residential and other uses Evidence of the rate of change is found in tie series of Landsat TM imagery and can be documented using various change detection techniques. This paper reports on the application of remote sensing techniques to map land use/land cover patterns as well as to detect areas of change using dry season image data of the Chiang Mai area as an example. 2. The Study Area Chiang mai city and its surroundings was selected for this study. The study site lies approximately between 18****40 and 19****00 N latitude and between 98**55 and 99** 10'E longitude (Figure 1). It encompasses the administrative districts of Mae Rim, Muang, Hang Dong, Saraphi, Sansai, Sankamphaeng and Doi Saket. The size of the study area corresponds to the 1002 by 1236 pixels of the Landsat TM image subscene which covers four 1:50000 scale topographic mapsheets (47461-II and 4846III-IV) of the Royal Thai Survey Department (1969). The total area accounts for about 111,460. Ha. 3. Methodology Landsat-5 TM digital data, path/row 131/47, of February 9, 1988, and February 17. 1991 were employed in this study. This data set was supplied by. Thailand. Criteria to the selection of the multi-temporal Landsat data set involved assessment of cloud cover percentage, time of acquisition, and sensor type so that change detection scope was optimized. In order to obtain the maximum information from Landsat TM used for image processing for land use/land cover change study, the selection of optimum Landsat TM bands is necessary. There are many different ways to select the optimum bands, such as the use of entropy criterion (Chen et al. 1986), calculation of transformed divergence between classes (Swain & Davis. 1978 Mausel et al, 1990), optimum index factor (OIF) (Chavez, et al, 1984. analysis of variance and covariance of the image scene (Sheffield, 1985) and image differencing of a suitable band, and Principal Components Analysis (PCA) of bands (Horler and Ahern, 1986. Byrne et al., 1980) In this study, the selection of TM bands was based on PCA and an OIF technique. PCA is regarded as one of the most effective ways to reduce the redundant information (of highly correlated bands) contained in Landsat TM data, while retaining the useful information. The OIF offers the optimum combination of triplet bands by order, based on the standard deviation and the coefficient correlation of the image statistics (ILWIS User's Manual 1993). Procedures followed the application of PCA and OIF have been extensively detailed in Sangawongse (1995). To ensure overlay precision for change detection purposes, a geometric rectification was conducted (on microBRIAN version 3.01-PC based of software) using image-to-map and image-to-techniques. More detailed descriptions regarding geometric rectification and resampling can be seen in Sangawongse (1995). Various change detection techniques were applied on both TM images individually. The change detection techniques used here have been developed to take into account the physical and cultural conditions within the study area, as well as the characteristics of any change detection algorithms that are appropriate to the image data (Jensen, 1981), Two major approaches were found to be best suited to the study area in terms of either ground conditions and the algorithms used. These approaches are: 1. Change detection based on two-date classification; and 2. change detection based on image ratioing. 3.1 Change detection based on two-date classification. This method has been developed from a number of previous studies (Ahmad, 1992, Monkolsawat and Thirangoon 1990, Wara-Aswapatti, 1990; Singh, (1986). Using this method, both actual and temporal changes in land use/land cover between different dates can be detected. In order to extract the maximum information on land use/land cover in the study area, classification using the best band combinations of Landsat TM imagery was emphasized. As the classified images ware to be used as inputs for change detection analysis later, they had to be classified as accurately as possible. monkolsawat and Thirangoon (1990) and Ahmad (1912) have pointed out that the accuracy of the output image is greatly dependent upon the accuracy of the input images used for a classification. Classifications procedures were applied on each Landsat TM scene separately by using the supervised classification method. For landsat TM scene 19988, some possible combinations (as indicated by OIF) such as 2,3,4, and 3,4,5, were chosen for the supervised classification using minimum distance to means, and maximum likelihood classifiers. The supervised classification technique is preferred, because the data of the study area is available and the author has a prior knowledge of the study area. There are 24 training area (2030 samples) used for the supervised classification of this image. These training areas were delineated from a false color composite image generated from two combinations (2,3,4 and 3,4,5). study areas , e.g. forests, agricultural field and orchards. To avoid misclassification, these tainting areas must be as homogeneous as possible. Ancillary data such as soil map and vegetation map were considered during the selection of training areas in order to obtain the greatest accuracy of the classification results (Hutchinson, 1982). For Landsat TM scene 1991, it was found that the visibility of this scene is not clear due to atmospheric effects such as haze and the increased air pollution in the city mainly caused by increased number of motor vehicles. In order to remove/reduce atmospheric effects, Normalized Difference Vegetation Index (NDVI) image was considered as another channel of information for extracting land use and land cover patterns from this scene. Some possible combinations of different bands together with NDVI were tested at different times in order to obtain the best result. In the end, the combination of all bands was chosen for the classification of the 1988 scene. whereas the combination of 2,3,4,5 and NDVI was applied to the 1991 Landsat scene. The inclusion of NDVI as another channel in this classification has offered considerable benefit, not only for extracting the maximum information topography and sun angle (Holben and Justice. 1980) Twenty five training areas (1458 samples) were used for classifying this image into 15 different classes using a maximum likelihood classifier. These training areas ware also chosen from the same band combinations as the 1988 one. Actual change can be obtained by a direct comparison between classification result from one date with that from the other date. Temporal changes that have occurred between the two dates can be measured by performing a change matrix (Howarth and Wickware. 1981) The change matrix was created in three stages: (1) to combine the classification channels of two Landsat scenes together; (2) to calculate changes between two dates using an appropriate formula; and (3) to crosstabulate two image channels as a means of portraying pixels in the classified images into a matrix (CSIRO) and MPA, 1993). The change image was created by using the following formula. (Jupp 1993-pers comm) where i is the number of elements in the matrix by row: j is the number of elements in the matrix by column; ci provides the row and has N1 levels (i =1.......N1); cj provides the columns and has N2 levels (j =1...........N2); N1 and N2 are the number of levels in the rows and columns respectively. The output image was rescaled into the range of 1 to 225 (15 * 15 matrix) digital values, which can be painted to enhance where changes occur. These values contain the number of pixels that have been changed from February 1988 to February 1991. 3.2 Change detection based on image ratioing This technique has been successfully applied by many authors for the study of land use/land cover change (e.g. Howarth & Wickweare, 1981 and Howarth & Boasson, 1983). Because of its simplicity and efficacy, it was selected as one of the techniques for examining changes in the study area. In this study, band 3 (visible red) from the February 9, 1988 scene was divided by band 3 from the February 17, 1991 on a pixel-by-pixel basis. Band 3 for Landsat Tm (similar to Landsat MSS band 5) which has band width between 0.63-0.69 micrometer was chosen because it is useful for discriminating between many plant species as well as for determining soil boundary, geological boundary delineation and cultural features (Howarth and Boasson, 1983). Also, bands 4 from both Landsat TM scenes, which have similar wavelength to Landsat MSS band 7 were ratioed in the same manner as bands 3. Areas of change in the study area can be enhanced by producing a False Color Composite (FCC) image (Howarth and Wickware 1981). The FCC image can be created by two ways : (1) assigning band 3 ratio to the red gun, band 4 ratio to the green and blue guns respectively and (2) assigning TM 3 or TM 4 of date one the red gun, TM3 or TM4 of date 2on the green gun and either ratioed 3 or 4 image on the blue gun for examining localized change. 5. Results and Discussion 5.1 Results obtained from two-date classification for change detection Land use/land cover in the 1988 and 1991 image scenes was classified into 15 categories equally. The accuracy assesment was conducted for both classified data (Congalton, 1991) to obtain an overall accuracy level of 82% and 85% for the 1988 and 1991 respectively. The land use/land cover statistics derived by a direct comparison between the two classified image data are represented in Table 1. It was found that the use of some possible combinations for classifying the 1988 image did not provide a satisfactory result as indicated by difference between the classified image and the ground data. For example, the 234 band combination did not provide accurate information of built-up areas. A paddy field to the south-west of the Chiang Mai international airport was mis-classified as belonging to a high density built-up area. Nor did it provide good enhancement of water bodies. For this reason, the overall TM bands (except the thermal band) was chosen for extracting information on land use from the 1988 scene so that, regardless o date handling problem, all information contained in these original bands would be assessed. Another reason is that this data has a good quality in terms of stability and is less subject to noise (Jupp, 1994-personal communication). Classification of the 1991 Landsat TM scene using the combination of bands 1-5 and 7 to give correspondence to the 1988 image yielded poor results. In particular, some shadows in the mountains were misclassified as belonging to water bodies. The change image constructed by performing change matrix yielded 225 possible land cover changes in the study area (see Table 2). In Table 2, the number of unchanged pixels are represented by values along the diagonal of the matrix, while the number of changed pixels are represented by values off the diagonal. It was found that 350398 pixels (31536 ha) had not changed, while the number of changed pixels totaled 888074 (79927 ha). Only 14 major changes were considered meaningful in the comparison to the ground truth information, so that they were maintained in the change map as shown in Figure 2. The biggest change was found from dipterocarp forest to low density built-up areas, which accounts for about 10422 hectares. In a descending order, the changing from dry paddy field to low density built-up areas accounts for about 7537 ha; change from dipterocarp forest to vacant land use accounts for 6996 ha; change form dipterocarp forest to dry paddy field accounts for about 5205 ha, and change from dipterocarp forest to mixed field crops accounts for about 4127 ha (see Table 3) 5.2 Change results obtained from the ratio image The result of applying this technique shows that changes have occurred between 1988 and 1991 Based on the false color composite image created by combination of bands 3 from two dates, and its ratio (Figure 3), green and blue colors obviously highlight areas of change, especially from dry paddy field to development land (mostly construction sites) and from dipterocarp forest to vacant land use. The prominent change areas are the green valley resort in Mae Rim district, the Mae Kwang dam in Doi Saket district and the housing estate near Hang Dong district, which clearly appears green. Areas of no change appear deep red, for example, hill evergreen forest at Doi Suthep (Suthep Mountain). cropland in Sansai district and soya beans cultivation in Hang Dong district. A false color composite image created by combinations of bands 4 from two dates and its ratio does not give a satisfactory result compared with the first combination, therefore it is not acceptable. 6. Conclusions and Recommendations The application of any change detection technique may be unsuccessful if user do not have enough knowledge about its characteristics in relation to the conditions over the area of study. Generally, the use of more than one technique is preferred by many researchers, because they can compare the results derived, and finally select the best ones for their project. The technique of change detection based on two-date classification has been found to be less restricted in terms of the algorithms used when compared with other change detection techniques. For example, the use of image differencing requires that the geometry of an image must be known as accurately as possible (about a quarter to half a pixel). otherwise substantial errors in change detection will occur. Close phenological correspondence between the images is also required, in order to reduce errors in change detection. However, change detection based on two-data classification requires good results from a classification. The most important factors that should be taken into account when performing change detectin, as recommended by Jensen (1981), have involved the familiarity with the study area, the quality of the data set, and the characteristics of change detection algorithms. It may be concluded that the use of Landsat TM for mapping land use/land cover change in the Chiang Mai area provided a satisfactory result if the appropriate techniques were used in data analysis. However, the research work on land use/land cover should be conducted on a regular interval, so that the information can be updated through time. This research is further investigating the use of GIS data and sequential air photos of the test area as another means of studying land use/land cover change. The large scale of the air photographs allows comparatively accurate interpretation of the relevant land cover boundaries, and these can be recorded in the digital spatial data base through use of a digital Elevation Model (DEM) (Salamanca Software Pty Ltd. 1992). Thus it can be demonstrated that older as well as more modern elements of the national information infrastructure can be used to augment and to test the change detection results. 7. Acknowledgments I wish to acknowledge the Thailand Remote Sensing Center for providing the Lansat data for this research project. Thanks are given to resource persons, from both Thailand and Australia e.g. Prof. Vanpen Surarerks (Dept. of Geography. Chiang Mai University. Dr. David Jupp (CSTRO, Division of Water Resources, Canberra, Australia) for their useful supervisors. A/Prof Jin Peterson and A/Prof. Paul Bishop for their assistance and encouragements throughout my research period. 8. References
|