Land Cover Change Detection Radio metrically-Corrected Multi-Sensor Data

Land Cover Change Detection Radio metrically-Corrected Multi-Sensor Data

Muhamad Radzli Mispan* And Paul M.Mather
Department Of Geography, University Of Nottingham
University Park NG7 2RD Nottingham
United Kingdom
*present address;Strategic, Environemtn and Natural Resources Research Centre,
MARDI. P.O. Box 12301, 50774, Kuala Lumpur.
E-mail :radzali@mardi.my

Abstract
remote sensing offers practical benefit in the field of land cover change detection. However, cloud cover restricts the use of optical remotely sensed data in tropical regions such as Malayisa. Therefore, data from different sensors such as Landsat TM and SPOT HRV are required in order to ensure as complete a templete a temporal coverage as possible. On the other hand, quantitative change detection study requires data to be corrected and converted to physical values such as radiance or reflectance factor. This study desribes a method of using reflectance factor of multi-sensor data for change detection analysis. The radio metrically- corrected multi-sensor data were transformed into uncorrelated component images in N-dimensional space using the Gramm Schmidt orthogonal transformation technique (GSO). The transformation is based on the four stable components and two change components extracted from the image data. The result of the transformation were then used to identify change in land from the image data. The study indicates that the GSO technique provide reliable information on the nature and type of change that is taking place over a period of time.

Introduction
Remote sensing techniques offer benefits inthis field of land cover change. One of the major advantage of satellite remote sensing system is their synoptic and repetitive coverage capability that can used to identify and monitor changes at regional and global scales. The spatio-temporal patterns of change in surface radiance offer reliable information sources on the state and nature of the surface feature and the process of changes that has taken place over a period of time. However, cloud cover restricts the use of optical remotely sensed data in tropical regions such as Malaysia. Therefore, data from different sensors such as Landsat TM and SPOT HRV are required in order to ensure as complete a temporal coverage as possible.

It is also critical that the type of technique and approach used to detect these changes are compatible with the characteristics of the surface features. Thus, for successful change detection analysis, the sections of an appropriate change detection algorithm is also an important factor. The selection should be based on the capability and flexibility of the algorithms the nature and physical of the are and the nature and type of Change that need to be addressed. This paper discusses methods of change detection utilizing multi-temporal and multi-sensor remote sensing data.

Materials and methods
An attempt is made to employ a Gramm Schmidt Orthogonalisation technique proposed by Collins and Woodcock (1994). In contrast with their approach, this study uses radiometrically corrected data of three homologous bands of green, red and near-infrared channels of landsat TM and SPOT HRV. The radiometric correction efforts are described in Mispan and Mather (1997) and Mispan (1997)

Resources
The study are is located at the central western coast of Peninsular Malaysia (Latitude 3o25'N and longitude 101o 45'E) about 20 km south of Kuala Lumpur. This study used Landsat -5 Thematic mapper ™ and SPOT HRV-2 data acquired on the 6 March 1990 and 26 December 1990 respectively (Table 1). The image processing and analysis were carried out using ERDAS-Imagine software at the Department of Geography, University of Nottinghanm

Table 1: Input reflectance (%) for Gramm-Schmidt orthogonal transformations

features	XS1	SX2	XS3	TM2	TM3	TM4
Wet soil	7.540	9.525	18.120	6.510	7.125	15.345
Dry soil	23.525	43.025	37.215	25.325	43.910	32.750
Vegetation	3.21	1.713	27.748	2.633	1.630	27.959
Water	4.350	2.150	1.850	4.900	2.900	2.250
Bare-vegetation	25.867	42.725	37.661	4.088	2.064	27.039
Vegetation-bare	4.796	2.389	27.759	13.519	18.833	32.741

Gramm Schmidt orthotogonalisation
The Gramm-Schmide orthogonalisation (GSO) technique was used by Kauth and Thomas (1976) to derive the coefficients of the Tasselled Cap Transformation (TCA) for singly-data Landsat MSS data. Jackson (1983) describes how the coefficients for n-space indices can be calculated using the Gramm-Schmidt Orthogonalisation technique with minimum data points. The first step in the construction of transformation vectors is the selection of a point in N dimentsional space that acts as the origin of the new co-ordinate system. To obtain the first index, an equation for a line through soil data points must be derived. A minimum of two soil points is required, with points differing considerably in reflectance preferred. For this study, reflectance values of wet and dry soils available in the image were chosen to drive the soil line. This is the first component derived from GSO technique, which can be said to represent the stable brightness index. The second stable component is orthogonal to the first component. The reflectance value of vegetation cover was used to derive this component. The dark dense vegetation 9DDV) wa chosen to represent the vegetative cover. This component can be said to be stable greenness index.

The third stable component, which represents a stable wetness index, is a component orthogonal to the first and second stable components. The representative reflectance for this component is selected from water bodies. The fourth component is a change from bare to vegetation. A similar situation applies to the fifthe component. However in contrast to the fourth component, this component represents a change component from vegetation to bare.

Calculation of the transformation matrix
A computer program to implement the Gramm-Schmidt Orthogonalisation described by Jackson (1983) was developed. The program requires an input file containing the initial value for the transformation coefficients was produced and shown Table 2. The transformation coefficients were later used to transform input data to new N-dimensional space. The result of the transformation were uncorrelated component images.

Table 2: Transformation coefficients of the Gramm-schmid orghogonal.

band	Gs1	Gs2	Gs3	Gs4	Gs5
XS1	0.26095	-0.18073	-0.24689	0.04577	-0.19943
XS2	0.54687	-0.31396	-0.47537	0.48842	0.06547
XS3	0.31172	-0.55726	-0.51663	-0.56281	-0.00233
TM2	0.30715	-0.14964	0.23175	-0.21913	0.87124
TM3	0.60050	-0.18355	0.53798	-0.31704	-0.44091
TM4	0.28413	0.70862	0.32098	0.54228	0.04975

Result and Discussions
The GSO transformation process produces a file containing five component images from the original six channels input data. This is in accordance with Jackson (1983) and Mather (1987), in that the process should result in N-1 component images, where N is the number of spectral band in the input data. Based on these coefficients, the five GSO component images were produced and these are shown in Figure 1.

Figure 1: Component images derived from Gramm-Schmidt Orthogonalisation techniques.
The transformation coefficient (Table 2) shows that all spectral bands contribute positive values in the first component (gsl), within which each of the red channels of both data sets contribute more than 50% to the component loading. Examining visually of the first visually of the first component image (gsol), three major cover types, namely, water bodies, vegetative cover and non-vegetative covers can easily be discriminated. In the image, water bodies appears dark in tone, vegetative cover in grey tones and non-vegetative in bright tones. In the vegetative area, it also found that there is variation in grey level values, which may due to different proportions of ground cover. However, the type of vegetative cover as well as change in vegetation cannot visually be discriminated in this component.

There is a distinct difference in the contribution of component loading between visible and near-infrared channels observed in the second component (gs2). The visible channels of both data sets contribute low and negative loading whereas the near-infrared channels contribute high and positive loading to the component vector. The component image (gso2) exhibits high visual contrast between two distinct land surface features; vegetative and non-vegetative covers. The non-vegetative cover, which include water bodies, urban area and bare appears dark in the image data but cannot be distinguished visually. The vegetative cover appears grey with various degrees of brightness. From its appearance, this component resembles a vegetation index image and may be used to discriminate non-vegetative against vegetative areas. However, similar to gsol, land cover type and potential change in its cover cannot be seen in this component.

The third component (gs3) exhibits almost similar component loading between the two data sets but in different. The table shows that SPOT HRV data contribute negative loading whereas Landsat TM data contribute positive loading to the component image. As expected, the component image (gs03) shows area of change from vegetation to non-vegetation. The area of change is visually distinguishable, in that it appears dark in the component image. Different degrees of brightness can also be seen in the vegetative area, which represent different type of vegetation. This corresponds to the graph shown in Figure 2, which shows substantial reductions in pixel values in area of change. Examining the area of change in the input data sets (corrected landsat and SPOT data), it appears that most of the changes are from vegetation to non-vegetation.

In the fourth component (gs4), the transformation coefficients in the visible bands of both data sets seem to be in reverse direction as compared to third component (gs3). Again, near-infrared channels of both data sets dominate the component loading are similar to third component. Visual inspection of the component image indicates change in surface feature between two data sets. Changes in cover types from non-vegetation to vegetation cover are shown in dark pixels, whereas non-vegetative cover for both change and unchanged areas appear bright. Similar to the third component (gs03), different variations in grey tone are also observed in vegetative areas. These indicate different type of vegetation cover. This phenomenon can be seen graphically in Figure 2. The fifth component shows the residual information on the GSO transformation. From the image it appears that this component contributes very little information on the study of change.

No	Class-Name	No.	Class_Name
1	Water	6	Stressed Rubber
2	Forest	7	Mixed Horticulture
3	Mature Oil palm	8	Bare soil/urban area
4	Young Oil Palm	9	Vegetation to Bare
5	Mature Rubber	10	Bare to Vegetation

Figure 2: Pixel value of various covers type derived from GSO techniques.
Conclusions
This discussion shows that multi-temporal and multi-sensor data can be made quantitatively comparably by converting the data into a common scale or datum. This is particularly useful in the continuous assessment of land surface feature, in which archived data are to be used. The Gramm-Schmidt Orthogonalisation method appears to be a much more promising technique not only in detecting land cover change but also providing reliable information on the nature and type of change that is taking place over a period of time. Furthermore this technique is independent of the image data, thus permitting the technique to be used with confidence.

Acknowledgements
The authors express their gratitude to the Director General of the Malaysian Agricultural Research and Development Institute (MARDI) for financial support during the study, and to colleagues at the Department of Geography, University of Nottingham, for their support and encouragement. The Malaysian Remote Sensing Centre, kuala Lumpur, Kindly supplied the satellite data used in this work.

Reference

Collins, J.B., and Woodcock C.E., 1994, Change detection using the Gramm-Schmidt Transformation applied to mortality. Remote Sensing of Environment, 50, 267-279.
Jackson, R.D, 1983, Spectral Indices in n-space. Remote Senisng of Environment, 13, 409-421.
Mather, P.M., 1987, Computer Processing of Remotely Sensed Images. (Chichester: John Wiley).
Mispan, M.R. and Mather, P.M., 1997, Multi-Sensor Radiometric Correction: A Case Study from Malaysia, Paper Presented At the 18th Asian conference on Remote Sensing, 20 -25 October, 1997, Kuala Lumpur,
Mispan, M.R., 1997, Multi-sensor remote sensing data for change deection: A case study from Peninsular Malaysia, unpublished Ph.D. thesis, University of Nottingham, 221 pp.