Vegetation Change Detection of Central part of Nepal using Landsat TM Kalpana G. Bastakoti Department of Geography, University of Calgary, kalpanagb@gmail.com Abstract This paper presents a study of detecting vegetation change in the central part of Nepal from 1989 to 2009 using Landsat TM data. The main techniques used were Principal Component Analysis (PCA) and Normalized Difference Vegetation Index (NDVI) differencing. Results of both methods indicate that vegetation has been decreased in urban and other built-up areas whereas it is increased in other parts over the past 20 years. Background and Relevance Timely and accurate change detection of surface features of earth provides the information about the relationships and interactions between human and natural phenomena which helps to make better resource management decision (Lu et al, 2004). In recent decades, due to the increasing awareness of depletion of tropical forests and its impacts particularly on biodiversity and global climate change, various global to local initiatives have been started to halt deforestation and conserve the resources. In Nepal, after the implementation of the Master Plan for Forestry Sector (MPFS) in 1989, several new initiatives have been started in the forestry sector policy. The major initiatives are community forestry and leasehold forestry in the mid-hills and mountains region, collaborative forest management in the Terai areas, and declaration and extension of protected area in biologically significant sites. Therefore, it is important to understand the effect of Nepal's new forest policy and initiatives in the forest cover dynamics. For a successful implementation of a change detection analysis using remotely sensed data, careful considerations of the remote sensor systems, environmental characteristics, and image processing methods are important. A single spectral band cannot reflect all the changed information due to the complex landscapes therefore use of transformed images or vegetation indices can be more effective in extracting the differences of changed features (Lu et al., 2005). Red and Near Infrared (NIR) channels of the sensors on board satellites are particularly well-suited to the study of vegetation (Mather, 2004).In general, change detection techniques can be roughly grouped into two categories. The first category include those detecting binary 'change or no-change' information, such as using image differencing, image rationing, vegetation index differencing, Normalized Difference Vegetation Index (NDVI) and Principal Component Analysis (PCA) etc. The second category include those detecting detailed from-to change trajectory, such as using the post-classification comparison and hybrid change detection methods (Lu et al., 2004). This study follows the first category of techniques and explore the forest cover change dynamics of central part of Nepal for a period of 1989 to 2009 using Landsat TM data.
Methods and Data Two Landsat TM satellite images of path 141 and row 041, acquired in October 31, 1989, and November 7, 2009 were used in this study. These images were geo-referenced in WGS 84, UTM 45 systems. The actual location of the image on the map is displayed in Figure 1. This study considers only the central part of Nepal. The data were analyzed using PCA transformation and NDVI differencing techniques utilizing IDRISI Selva software. Figure 1: Map showing the exact location of satellite image on the map of Nepal (Source: http://glcfapp.glcf.umd.edu:8080/esdi/index.jsp) Principal Component Analysis PCA is a data reduction or data transformation technique that can be used to transform an original image database into a more compact dataset composed of a smaller number of orthogonal (uncorrelated) variables. Previous studies have shown that selecting band each from the visible, near infrared, and mid-infrared spectral regions results in the optimal waveband combination for vegetation discrimination (DeGloria, 1984; Horler & Ahern, 1986; Sader, 1989). Therefore, in this study, bands 3 (visible red), 4 (near infrared), and 5 (mid-infrared) were extracted from each 1989 and 2009 data sets and PCA was performed on both data set. Interpretation of the PCA outputs was done based on algebraic signs of the eigenvector/loadings and the visual assessment of the principal component (PC) images. Normalized Difference Vegetation Index NDVI is the most commonly used technique for land cover change detection particularly for vegetation monitoring which separates the green vegetation from its background soil brightness. It is simply the difference between the near infrared and red bands normalized by the sum of those bands. NDVI index value ranges from -1 to 1 where higher value represents abundant green vegetation and lower value indicates the low or no vegetation.
The mathematical formula for calculating NDVI is NDVI = NIR (TM4) - R (TM3)...(1) NIR (TM4)+ R (TM3) Where, NIR = the spectral reflectance measurements acquired in the near infrared band R = the spectral reflectance measurements acquired in the red band. The difference image of 2009 and 1989 was created by subtracting 1989 NDVI values from 2009. DIF = NDVI [ 09] - NDVI [89]...(3) From the NDVI difference image, vegetation change map was prepared using density slicing which indicats negative change (value < 0), no change (value = 0) and positive change (value > 0). Results Principal Component Analysis The bar graph in Figure 2 shows the variance explained by each principal component. The first four principal components PC1, PC2, PC3 and PC4 comprise a total of 97% of the original variance. Figure 3 shows the loading values of all six PC which reflects the relationship between the principal components and original channels. Figure 2: Variance accounted by each PC Figure 3: Loading values of all PC for all bands Analysis of the loadings (Figure 3) of the transformed data indicates that PC1 represents an average of all the bands of both dates as the loadings of this components are all positive and similar values. PC2 has negative loadings for all bands in 1989 and positive for all bands in 2009 which implies that PC2 is a difference image of these two dates. PC3 is notable for having positive value for visible and negative for near infrared both in 1989 and 2009 suggests inter-image differences among the visible and infrared (IR) bands that could be associated with average of the vegetation of the two dates which
does not provide information about change. Conversely, in PC4, band 3 of 1989 is positive and band 4 is negative which is reverse in 2009, suggesting that this component is a difference image of 1989 and 2009 and indicates change in greenness. Therefore, in this study, PC2 and PC4 contain most of the change information whereas PC2 shows overall change and PC4 mostly shows the changes in vegetation. PC1 PC2 PC3 PC4 PC5 PC6 Figure 4 : Principal component images (enhanced) obtained from bands 3,4,5 of 1989 and 2009 Landsat 5 TM images
In Figure 4 green color represents the vegetation and other colors indicate nonvegetated area. The first four PC components show clear geographical pattern and latter contain mostly noise. PC2 and PC4 which represent most of the change information in which darker green color displays improved vegetative area and lighter color (yellowish) shows area that is converted from vegetation to urban/ barren land. Figure 5, RGB composite image of PC4, PC2 and PC1 display the change and no change area. Blue color represents no change, bright whitish color represents changes of vegetation area to built up/barren land and green color represents improved vegetation. It is clear in the image that major cities like Kathmandu and Hetauda have been expanded over the period. Some area in hid-hills seems improved in greenery. Kathmandu Hetauda Figure 5. PCA composite image RGB PC4,PC2,PC1 Normalized Difference Vegetation Index Figure 6 displays NDVI images of 1989 (a), NDVI images of 2009 (b), NDVI difference image (c) and NDVI difference classified image (d). In the NDVI difference image (c), values that are negative indicate decreased greenness, values close to 0 indicate areas remained relatively unchanged such as water, barren land, ice, snow etc, and positive values represent improved greenness which is clearly seen in Figure 6 d classified map.
From the figure, it is apparent that in major cities like Kathmandu and Hetauda vegetation has negative change but in other areas, there is positive change. a. NDVI 1989 b. NDVI 2009 Kathmandu Hetauda c. NDVI difference d. NDVI difference classified map Figure 5 : NDVI image of 1989 (a), NDVI image of 2009 (b), NDVI difference image (c), and classified NDVI difference map (d) Comparison between PCA and NDVI Comparing the results from PCA and NDVI, both methods are found useful to detect vegetation change. From the visual inspection of the images from PCA and NDVI, change from vegetation to urban/barren land seems almost similar in both the images but the change in vegetation area seems slightly higher in NDVI difference map than in PCA. However, because of the lack of ground data to verify the results, it is not possible to conclude which method performed better over other. The NDVI differencing method is relatively easy to interpret, but it cannot provide complete matrices of change directions (Lu et al., 2004) and the index differencing is also subject to registration error (Gong & Howarth, 1992). When using the NDVI
differencing method, some changes can be caused by other factors than the actual land use land cover change such as plant phenological change, illumination difference, and atmospheric condition change between two dates may affect final change detection results. One important advantage of NDVI is that it combines the best features of difference and ratio indices and minimizes the topographic effects. PCA is data dependent, so it can't be determined beforehand which of the PCs best displays the change and the direction in which the change occurred (Byrne et al, 1980). Therefore, a good knowledge of the history of the study area is necessary to identify the changes in the PCA outputs and separate the various change types. Furthermore, the algebraic signs on the eigenvectors can be interpreted in terms of "greenness" and "brightness" changes, but this is subjective and not based on standard correlation, such as that associated with the NDVI (Hayes & Sader, 2001). Therefore, the interpretation and threshold of PCA change imagery is more complicated than NDVI differencing (Hayes, 2001). Conclusions This study demonstrated the feasibility of using time-series Landsat satellite image and different change-detection approaches to extract vegetation change information. Two change-detection techniques i.e. Principal Component Analysis and Normalized Difference Vegetation Index differencing were used in this study. Both of the methods are found useful to display the general pattern of change. Change from vegetation to urban/barren land seems similar in the images from PCA and NDVI but the changes in greenness seems slightly higher in NDVI difference map than in PCA. Results of both methods indicate that vegetation has been decreased in urban and other built-up areas and it is increased in the major conservation areas and participatory forestry intervention area over the past 20 years. References Byrne, G.F., Crapper, P.F. & Mayo, K.K. (1980). Monitoring land cover change by principal component analysis of multi-temporal Landsat data. Remote Sensing of Environment, 10, 175-184. DeGloria, S.D. (1984). Spectral variability of Landsat-4 Thematic Mapper and Multispectral Scanner data for selected crop and forest cover types. Geoscience and Remote Sensing, IEEE Transactions on, 22(3), 303-311. Eastman, J.R.( 2012). IDRISI Selva. Worcester, MA: Clark University.
Gong, P., & Howarth, P.J. ( 1992). Frequency-based contextual classification and gray-level vector reduction for land-use identification. Photogrammetric Engineering & Remote Sensing, 58,423. Hayes, D., & Sader, S.A. (2001). Change detection techniques for monitoring forest clearing and regrowth in a tropical moist forest. Photogrammetric Engineering & Remote Sensing, 679, 1067 1075. Horler, D.N.H., & Ahern, E.J. (1986). Forestry information content of Thematic Mapper data. International Journal of Remote Sensing, 7, 405-428. Lu, D., Mausel, P., Brondízio, E., & Moran, E. (2004). Change detection techniques. International Journal of Remote Sensing, 25(12), 2365-2401. Lu D., Mausel, P., Batistella, M., & Moran, E. (2005). Land-cover binary change detection methods for use in the moist tropical region of the Amazon: A comparative study. International Journal of Remote Sensing, 26, 101 114. Mather, P.M. (2004). Computer processing of remotely sensed images: An introduction, (3rd edn), Chichester: John Wiley & Sons Ltd. Sader, S.A. (1989). Satellite digital image classification and multispectral characteristics of northern hardwood and boreal forest communities in Maine. Maine Agricultural Experiment Station, Miscellaneous Report 334.