Chapter 6 - Summary & Conclusions
Limitations of the Study
The multispectal mapping of the land associated with digital remote sensing techniques is characterized by, but not restricted to, inherent limitations. No map produced by digital manipulation of multispectral data is ever 100% correct when it is produced by a computer (Robinove, 1981). By nature, the process of classifying such a broad range of the Earth’s features into specific and often simplified land use and land cover classes introduces error by drawing boundaries around geographically located classes that are ‘homogeneous’ or acceptably heterogeneous. However, these limitations can often be overcome by sound statistical analysis to produce acceptably accurate land use and land cover maps as derived from multispectral satellite data.
In retrospect, there are a few, specific limitations in this research which should be addressed as a means for improvement or potential strategies for further study. The first limitation focuses on the ground truth data acquired for accuracy assessment. By utilizing the process of obtaining the ground truth data by extensive GPS field surveys as discussed in Chapter Four, Methodology, bias with respect to proximity to roads is characteristic of the data. It should be noted that this is not critical to the overall accuracy assessment of the land use and land cover map, however, it is important to mention. Obviously, the characteristics of the landscape are not the same where the impacts of roads are not felt. A potential solution for this limitation could be to combine both road and hiking or foot GPS surveys. One additional bias of the data utilized for ground truth is inconsistency with respect to number of pasture sites as compared with the other land class data points. This issue has been previously addressed in Chapter Four, Methodology.
A second limitation of this project reiterates the strict rectification standards required for accurate and quantitative change detection. Unfortunately, research on this subject by Dai and Khorram (1998) was obtained after registration was performed and classifications procedures were well underway. The Root Mean Square (RMS) error of 0.5 pixel was deemed acceptable at the time of registration. Possible implications have been previously addressed in Chapter Four, Methodology. It is imperative to perform extensive research of all facets incorporated in a project of such complexity prior to undertaking such rigorous and technologically involved statistical analysis.
The third and probably most important limitation associated with this research is the lack of a quantifiable assessment for change detection. Unfortunately, the author did not have access to similar GPS ground truth data to coincide with the 1984 Landsat image. Otherwise, a separate classification of the 1984 Landsat image could merit statistical analysis for a quantitative change detection assessment. Nonetheless, a 1999 land use and land cover map for Carroll County, Arkansas characterized by sound, quantifiable accuracy was derived as a result of this research. By considering the shortcomings associated with this methodology, perhaps this study could serve as a basis for further research and analysis of land use and land cover change in Carroll County in the future.