Conclusions on the comparison of vegetation derivation methods

Overall the two methods yielded similar vegetation change vegetation change results. When looked at spatially, the two images show similar areas of high and low vegetation abundance. The overall similarity in the results indicates that for most areas with little continuous forest, NDVI is fairly accurate in its analysis of overall vegetation abundance. However, when compared in scatter plots, the two methods reveal marked differences for areas with very high vegetation abundance. Saturation effects indicate that subtle changes in undisturbed and continuous forests may not be accurately detectable through NDVI. It is important to note that it is generally the value of vegetation abundance that is inaccurate and not the spatial distribution of those pixels containing vegetation. Therefore, inaccuracies in NDVI seem most significant when quantifying vegetation. If the goal is simply to map the spatial extent of vegetation cover, NDVI seems to be sufficiently accurate. Once a value of vegetation abundance within a pixel is attached, however, the effects of soil backgrounds and saturation take effect.

The vegetation change results indicate that given the types of changes seen at 57m resolutions over four year intervals for the land covers associated with the image, NDVI sufficiently detects actual vegetation change. This is especially encouraging to researchers interested in detecting primarily drastic anthropogenic changes in vegetation cover. Overall, analysis of forest change is difficult using NDVI and may not be reliable. While it is difficult to determine the absolute level of accuracy that can be obtained through a cutoff methodology without additional ground truth data, it seems clear that many accuracy obstacles arise when thresholds are used. The results presented above suggest that it may be more accurate to analyze vegetation change as a whole rather than trying to separate out forested areas using a threshold methods. A potential to determine human induced forest thinning exists using SMA endmembers to separate forested areas (as described in Adams et al., (1993)). This analysis is potentially more accurate than any methodology using NDVI values alone, although actual accuracy verification of forest change data was not possible for this study. Most importantly, the forest change results calculated through NDVI and SMA drastically differed, indicating that if forest change is an important component of the study, the methodology used is an important variable to consider.

Back to Index

For more information please contact Laurel_Firestone@yahoo.com