Digital Classification of
LANDAST TM for Land Cover Mapping of the Pa Wang Phloeng-Khom-Lam Narai
National Forest Reserve, Lop Buri Province, Thailand Kaew Nualchawee and Lilita
Bacareza Introduction Asian Institute of Technology (AIT), Bangkok Thailand The objective of the study was to come up with a land cover map of the Pa Wang Phloeng-Muang Khom-lam Narai National Forest Reserve using various computer-assisted classification of Landsat digital data taken on 18 July 1993 with reference data from 26 January 1994, and to determine practical remote sensing approaches for high classification accuracy result. The demonstration of the stages and final results of the study are presented with focus on the major steps that were employed:
Related Past Studies In a related study of the area done by kasetsart University in 1984, manual analysis of aerial photographs of scale of 1:15,000 (date 18 November 1983) was performed to identify extent of land cover and land use within the forest reserve. Recently, the Prcticum research Team of the Natural resource Program of AIT conducted another study using Landsat Thematic Mapper (TM) digital data to demonstrate the usefulness of satellite remote sensing in yielding information for land cover changes from 1983 to 1994 in the absence of serial photographs (Practicum Research Study, 1994) Information derived from the two studies served as input to the present study. The study area The study area is a part of Khok Charoen District, Lop Buri Province in the upper Central region of Thailand, specifically located in the Yang rak Subdistrict at Latitude 15 degree 15 minutes to 15 degrees 27 minutes North and longituds 100 degrees 52 minutes to 101 degrees East. The average annual rainfall from 1983 to 1994 was 1,054 mm to 1,324 mm, with the rainy season lasting from May to October leaving November to April a distinctly dry season. Average temperature over the past eleven years was 27.8 degrees Celcius. The elevation level in the area varies from 80 m to 560 m above mean sea level, with more than 40% of the land area having a slope gradient of 0-2%, while 4.7% has slope gradient of 40% Vegetation is a mixture of fruit and forest trees. Eucalyptus and Casuarina plantations abound, while the mountains and hills or upland vegetation are dominated by reproduction of salings of Dipterocarpacaea, Myrtaceae and Leguminosae (Practicum research Study, 1994) Methodology Selection of the remotely-sensed Data There was no past experience over the study area as to what season would be best for discrimination of major vegetation cover, etc. The selection of the remotely-sensed data was mainly depended on the availability of the most recent high quality and cloud-free satellite image. In summary, the following images were selected and used for the study:
The Image Analysis System Used The digital analysis was performed at the remote sensing Laboratory (RSL) at AIT, using the ERDAS Image Processing, PC-based System, and CCT was used for data input. The following steps were taken during this study: Geometric Correction of the Image A first order polynomial transformation and resampling with nearest neighbor algorithm was used to spatially geo-reference the Landsat image to UTM map projection system Ground control points identifiable at road intersections in reference to topographic maps (1:50,000) and those obtained from field visits through the Global Positioning System (GPS) were used. A root mean square error of pixel (30m) was accepted for the correction process using a total of 8 Ground Control Points (GCPs). Image Enhancement The contrast stretch image enhancement was applied to the original image to be used as reference for the interpretation of the raw data. For the study, however, the histogram-equalized stretch was used. Identification of the different Resource Classes Based on the field visits to the study area conducted from February to May 1994 and adopting the US Geological Survey classification (Anderson, 1976) land cover classes to be mapped were define, and a classification scheme for the digital analysis was determined. Brief descriptions of the categories are presented as follows.
Determination of Training Approaches for Image Classification In this study, the two basic approaches to training set selection were used, namely uspervised and unsupervised approaches. These two approaches, however, are not always satisfactory for various conditions of environment and variations in cover types, etc. Therefore, additional to the two approaches mentioned, the third approach was also considered, where in both supervised and unsupervised methods were taken into consideration. It involves applying first, a clustering algorithm to the data (the unsupervised way) to define spectral clusters, the results of which can then be manipulated by the analyst for further analysis and training-set determination (the supervised way). Past tests have shown that this "modified (controlled) clustering approach" was judged best because it resulted in savings of man-hours and computer time, as well as having the highest classification accuracy (Fleming, et al., 1975; Rohde, et al., 1977, 1978; Pettinger, 1982) These studies have concluded that the modified clustering approach was best suited when the area are spectrally complex due to variation in vegetation cover and terrain. Results and Discussions The Supervised Classification Approach A total of 10 resource classes were identified based on the supervised classification scheme. Table 2 below show the area and percentage distribution of each resource class.
The Unsupervised Classification Approach In the unsupervised classification approach, spectrally separable classes are first determined and their informational utility is then defined. Clustering algorithms are used to determine the natural spectral groupings present in a data set. The basic premise is that values within a given cover type should come close together in measurement of space, whereas data in different classes should also be well separated (lillesand and Kiefer, 1987) For this study, the sequential method for clustering was employed, and the classification results, based on the final 12 resource classes are shown in Table 3 below. The Table shows the spectral classes represented from the clustering method for unsupervised classification and its corresponding area-wise distribution.
The Modified Clustering Classification Approach This approach, commonly termed hybrid classification, involves elements of both unsupervised and supervised analysis. A hybrid classifier is on which incorporates two or more decision rules (Mather, 1987) such as the independent clustering of the pixels (unsupervised method) which are then subject to analysis for spectral separability and normality (supervised method). Five training areas were statistically clustered into 100 spectrally separate clusters and an evaluation of the statisics of the signatures derived from each cluster was performed. The final Classification using Maximum likelihood classifier shows the results in the Table 4 which presents the area distribution and percentage of various resource classes that have been identified.
Classification Performance for the Three Classification Approaches To evaluate the results of the classifications and to verify the degree to which the land cover maps derived would meet users' needs, classification accuracy assessment was done through machine-assisted procedure. One basic constraint in the study is the lack of in depth ground truth data for the study area and thus accuracy assessment is based mainly on spectral analysis of the digital data. The derivation of accuracy figures are based on the following definitions:
Using the above resoning, the following figures compare the performance of the three approaches:
Conclusions Where areas are characterized by topographic and vegetation complexity as in the case of the Pa wang Phloeng-Muang Khom-Lam Narai National Forest Reserve, an effective method for the classification process suited for the environment results in a significant improvement in identification of various individual cover types. Among the different approaches to classification, the study shows that the modified clustering approach to classification is the best technique to employ since it allows for a high degree of interaction between the analyst and the machine, there by complementing the inadequacies of either party. The difficulty in derivating signatures is effeciently executed through the methodology of hybrid classification, at lesser processing time. The analyst on the one hand, is also given the opportunity to "control" which signatures best qualify to represent a certain resource class considering the fact that he or she holds the reference data from which to base the fine decision for classification. The results of the study also show that the modified clustering approach gives the highest overall classification accuracy of 90.9% for level II resource class. The capability of ERDAS for man-machine interaction also facilities the various steps in digital analysis and the use of Landsat TM data for vegetation mapping has produced reasonableaccuracy results. Its main limitation however, is therefor, is the presence of cloud cover in the image, which constrains the interpretation of some areas. It is therefore recommend that further studies on the area with use of remote sensing data be realized in conjunction with the use of radar digital data whose capabilities allow the user to 'see' through the clouds (aschbacher, 1991) Acknowldgements The authors are grateful for the advice and assistance given by many people involved, especially Mr. Yian Kawan Ang, Manager of the Remote Sensing Laboratory (RSL), Mr. Than Naing, and Mr Quang Instrumentation Engineer and System Analyst, respectively. Also sincere thanks are due to Mr. More Myint, a doctral candidate of the STAR Program of AIT. References
|