|
Abstract
GreenRevolution Gujarat, India SatelliteImagery
Dams Irrigation Desertification
|
In order to understand the types of land cover changes and environments that cause differences between the vegetation change results of NDVI and SMA, it is helpful to first examine the initial vegetation abundance values generated by the two methods. The scatter plot in Figure 6 compares the initial SMA vegetation abundance results with the NDVI vegetation abundance results for the area shown in the image in Figure 7. Each point on the graph represents a pixel in the scene. If the two methods determine relatively equal amounts of vegetation for a pixel the point should fall along a best-fit line, showing a linear relationship between the two methods. As the graph shows, the majority of the points fall along this line for low vegetation abundance, however, the graph becomes less linear as vegetation values increase. Overall the scatter plot does not show a tight linear relationship, but rather good correlation at medium to low vegetation abundance with divergence at higher values. In order to determine the types of land covers that differ between the two methods, we can color code classes of points in the graph and highlight the corresponding pixels in the scene. If we compare where the pixels are in the image to the original land cover of the MSS color image, we can begin to determine the land covers that may be interpreted differently by NDVI and SMA.
High
Vegetation Abundance and Complex Forest The primary classes of areas that differ between methods correspond to areas with high vegetation abundance (specifically undisturbed, continuous forests, highlighted in Red in Figures 8 & 9), and water (highlighted in blue in Figures 8 & 9). Because this study is concerned with vegetation change, difference in forest cover vegetation values is much more important than difference in water values. The red areas seem to be darker vegetation in the original image (Figure 7). Most likely these darker areas of vegetation are forests of complex structure, and shadowed due to canopy layers, topographical features, or the sun angle. This rounding off of the graph corresponds to what would be expected from the saturation problem discussed earlier, which is associated with NDVI results (see Figure 4) (Baruet & Guyot, 1991; Hatfield et al., 1985). It is a well-known phenomena that as leaf area index and biomass reaches a threshold level, NDVI is a poor indicator due to saturation effects (Bégué, 1993; Chance, 1981; Wanjura & Hatfield, 1987; Weigand et al., 1991). SMA also saturates at high levels of vegetation abundance, however, SMA is able to accurately detect vegetation cover to much higher levels of abundance than NDVI (Smith, 1990 a&b). The threshold for NDVI has been found to lie between a leaf area index of 2-6 depending on the type of vegetation and the experimental conditions used to determine accuracy levels (Hatfield et al., 1985). Overall, SMA has far greater sensitivity to subtle differences in high vegetation abundance than NDVI. This can be illustrated through a coarse but more detailed dissection of the scatter plot of SMA v. NDVI. The color highlights in Figures 10 & 12 correspond to the land covers of the pixels in the images in Figures 11 & 13 (refer to Figure 7 for reference). These colors show small, equal intervals of vegetation abundance according to NDVI and SMA. What is most noticeable from these plots is that SMA has more different classes of vegetation abundance at high values of vegetation than NDVI, indicating a potentially greater ability to detect subtle change in forest. Implications of Soil and Saturation Effects for Change Results While it is interesting to note the differences between the initial vegetation abundance results of NDVI and SMA, the importance is the implications of these discrepancies for vegetation change detection. Based on the nature of the differences in vegetation abundance as calculated by SMA and NDVI, we can speculate on the types of land cover changes that may cause differences between vegetation change calculations. For example, Figure 14 shows the possible implications of saturation and background soil effects for vegetation change analysis. As the diagram indicates, the saturation problem associated with NDVI should cause the values of vegetation change calculated by NDVI to be lower than those calculated by SMA in areas of high vegetation abundance. Subtle changes in healthy forest, such as small degrees of degradation or drought effects, may not be detected by NDVI, while these should be detected by SMA. Additionally, Figure 14 shows the implications of soil background color effects for vegetation change calculations.[10] If highly vegetated areas are deforested, exposing background soil levels, the change detected by NDVI should vary depending on the overall soil brightness. Analysis of Observed Vegetation Change Results The correlation between vegetation change values as calculated by SMA and NDVI is generally quite good (see Figure 15). Overall, the change values correspond linearly, although some pixels do differ. Therefore, it is likely that few if any subtle changes in continuous forest cover were observed or detectable at this resolution and over this time period. In addition, the number of pixels with continuous forest and significantly different types of background soil that experienced the same actual forest change may be few or the change may have been so drastic that the subtle soil effects were not significant. However, discrepancies between the two methods become important when individual areas are examined in detail. The chart below shows examples of aggregated vegetation change data using SMA and NDVI for aggregated areas within a 10 km area of influence around two village centers (see Figure16). Because these two methods yield vegetation abundance measurements on different scales (NDVI between -1 and 1 and SMA between 0 and 1) the values generated by the two methods are not directly comparable. However, the overall percent change in vegetation abundance from year to year for each method and area is shown in the chart below. Examples
of Vegetation Change Values from 1976-1980 over Aggregated Village Areas (10 km
radius)
The overall percent vegetation change calculated by these two methods is different in absolute values, but they do not differ drastically in the magnitude or direction of change. This difference may be significant depending on the use of such data within larger studies. Analysis of Forest Change results While overall vegetation change information is useful, many researchers may need information on the changes in forested areas specifically. Separating out types of vegetation is difficult when using multispectral data because it has so few spectral wavelengths to use to distinguish between types. One common technique for estimating forest change using NDVI is to set a threshold point, where everything above a specified level of vegetation abundance is classified as forest. The absolute area classified as forest can then be tracked over time to calculate spatial changes in forest cover. This methodology, however, has some problems. One obvious problem is that this method would mean that any kind of highly vegetated area, including agricultural plots, would be classified as forest. This may not be a problem if the highly vegetated areas in a scene are generally all forest anyway. For example, researchers have avoided this issue by picking dates during which agricultural fields are fallow or have not yet reached high vegetation abundance. However this seasonal flexibility is not always optimal, especially if agricultural practices or productivity is also an interest for a particular project. For many studies, such as the one presented here, this is a significant obstacle. It is virtually impossible to separate forest from agriculture based on NDVI values alone.[11] An additional problem with using absolute thresholds is that it is extremely difficult to find an accurate threshold without obtaining expensive and time consuming ground truthing data, and the choice in threshold can drastically change forest change results. If a large number of pixels are clustered just above the chosen threshold value (as might be expected in NDVI due to the saturation problems at high vegetation abundance (see Figure 4)), then small changes in NDVI values may cause pixels values to fall below the threshold, and therefore indicate more forest change than actually occurred. For example, in 1972, most of the forested areas had NDVI values greater than 0.5. However, when this same threshold level was applied to 1976 many forested areas were no longer classified as forest. The numbers indicated that 120.99 km2 of forest were lost in this four-year period, a much larger amount than seemed reasonably accurate. In order to avoid this problem, a lower threshold could be chosen; however, this requires that more areas with lower vegetation abundance be part of the forest change analysis. The images in Figures 17, 18, & 19 show the areas that would be classified as forest if a threshold value of 0.35 was chosen. The raw MSS data image can be used for reference when interpreting the land cover highlighted in these figures (see Figure 20). As the images show, many more pixels were classified as forests in 1980 despite using the same value of .35 for each year. The reasons for this become clearer when the histogram for the three years is examined. Figure 21 shows the number of pixels for each value of NDVI for each year for the area shown in Figures 17-19. The red line indicated the threshold level of 0.35, and the green line is fixed as 0.6. As the plots show, the distribution is similar from year to year for most of the histogram below approximately 0.2, however, the overall distribution of the data between 0.2 and 0.7 shifts slightly from year to year. One possibility is that these shifts are caused by small errors in the process of radiometric alignment. However, because the overall distribution of the data is not equally shifted across vegetation abundance levels, alignment problems are less likely than the possibility that actual changes in vegetation cover occurred. Given the changes in the histograms over time, overall vegetation decreased over the first four years and then increased over the second four years. The quantitative results showed a decrease in forest cover of 13.88 km2 from 1972-1976 and a later increase of 84.32 km2 from 1976-1980. While we do not have actual field data to quantitatively verify this result, it seems likely that the increase from 1976-1980 is larger than actual forest increases due to the addition of many small plots in the left of Figure 19 that are most likely not well-structured forest. This methodology is actually calculating the change in medium-high and greater vegetation abundance. Therefore, it is likely that while actual forest coverage may not have changed drastically, background levels of vegetation abundance and secondary growth along the edges did change significantly. Unfortunately, however, separation of these various types of vegetation is not possible through this methodology, and therefore this distinction is impossible to verify using NDVI alone. If spectral mixture modeling is available, other methods for separation of vegetation types exist. One of the most successful was developed by Adams et al. (1993). They separated land covers based on quantitative fractions of endmembers and temporal changes in endmember values. Well-structured forests, for example, were found to have specific ranges in the fractions of shade, vegetation, and soil. Areas with these specific ratios of endmembers could be grouped as forest and tracked over time. Similarly, pastures and secondary growth could be classified based on different ranges of endmembers and the changes in the fractions of endmembers over time. This kind of analysis is important to consider as it shows the advantages of SMA’s additional information, in the form of other endmember fractions, over NDVI vegetation data alone. The additional endmember fraction data that is generated by SMA is not only useful in drawing conclusions on the changes in that land cover itself, but can be used to classify more complex land uses, and track the changes in those as well. To test this method of using SMA fraction ratios to define forest cover, the forested test site used for NDVI forest analysis was used (see Figure 20). Analysis of scatter plots comparing soil, shade, and vegetation fractions was used to develop a range of these endmembers which characterized forests areas as separate from other highly vegetated areas. The final characterization defined forests as having soil values less than 0.19, vegetation values greater than 0.27, and shade values between 0.2 and 0.68. Given this definition, forested areas were calculated to have increased by 5.27 km2 from 1972 to 1976, and then decreased by 0.22 km2 from1976 to 1980. Therefore the overall change in forest from 1972 to 1980 was 5.05 km2. This indicates extremely small amounts of change in the overall forested area, and differs sharply from the forest change calculated by NDVI using the threshold method discussed earlier. The results of the NDVI forest change analysis showed a decrease in forest cover from 1972-1976 and an increase from 1972-1980, exactly opposite results from those presented from the SMA analysis. Since actual forest change for this region is unknown we cannot verify either one, although comparisons of these two methodologies in earlier in this study indicate that NDVI may be less accurate than SMA in these areas. It is possible that the definition of forest chosen for the SMA analysis was not adequate, or it may be that while highly vegetated areas changed significantly over this time period, actual forested areas were generally stable. Given the information available to us, we cannot confidently rule out either possibility. But it is important to note how radically different the results of forest change analysis are using NDVI as opposed to SMA. Uncertainty: The uncertainty involved in this analysis enters in many steps of the process. The resolution is low and therefore features smaller than the pixel size (57m by 57m) are generally not detectable . In addition processes such as radiometric alignment, and registration do alter the data, adding small levels of error. However, generally errors in radiometric alignment and registration should not affect differences seen in NDVI and SMA as they were performed previous to both these methods and are base conditions of the data. |
|
For more information please contact Laurel_Firestone@yahoo.com |