GEOG 323 Reflection 3: Error and Uncertainty
As the name of this blog post suggests, we are currently studying error and uncertainty in the context of spatial analysis. Our discussion of this topic is grounded in Chapter 6 of Longley et al. (2008), which discusses various types of error and uncertainty that can emerge in GIS work. Longley et al. diagram the progression and development of uncertainty in Figure 6.1 (p. 129), modeling how "filters" between "real world," "conception," "measurement & representation," and "analysis" can distort data and conclusions by the time they appear in the final analysis.
Thinking about error in my vector GIS analyses is not something I have often done. It's much easier to believe that geospatial data perfectly represents the truth than to stop and consider how error might have wormed its way into the data and analysis. Moreover, as Longley et al. point out, in a GIS, "results will have been reported to high apparent precision, with more significant digits than are justified by actual accuracy, and lines will have been drawn on maps with widths that reflect relative importance, rather than uncertainty of position" (p. 153). Such results can be inherently misleading, inspiring disregard for uncertainty in addition to unwarranted confidence in the maps and numbers shown.
However, I do have some experience with modeling uncertainty in raster image classification analyses. In Remote Sensing & Land Use (GEOG 351) with Prof. Niwaeli Kimambo last fall, we made confusion matrices to analyze the accuracy of some of our tree-cover classification maps. This was a great experience because it allowed us to critically examine the strengths and weaknesses of our classification methods and to compare how well the classifications worked for different image collections. It was also a good introduction to how statistics relates to error since the value of a classification method is determined largely based on how well it performs relative to its expected accuracy, which is derived from the distribution of sample data in the input raster.
Although the topic of uncertainty is almost by definition an uncomfortable one, it's also one worth embracing by geographers. We wouldn't have so many different spatial data formats, projections, collection methods, and more if each one didn't have its benefits and its drawbacks. As Longley et al.'s Figure 6.1 demonstrates, error is inherently propagated along the way from the real world to the analysis as we interpret what's around us and develop representations for it. In order to make our work as truthful and understandable as possible, and to enhance reproducibility and replicability, it's important to document potential sources and locations of error and uncertainty. If we don't pay attention to uncertainty, we also don't have any incentive to develop new forms of collecting, storing, and analyzing data that increase precision and accuracy.
There are a variety of ways that geographers can make sure uncertainty gets the attention it deserves. If relevant, results should be reported with confidence intervals, and the number of significant figures should be appropriate to the context of the input data and analysis methods. The pros and cons of using different file types, data structures, and data collection methods should be highlighted to encourage using the best possible dataset for the task at hand. Perhaps most importantly, we must remember that maps are not perfect representations of the world, as much as we might like to believe they are. Maps can intentionally lie, mislead, or spread propaganda, but even when none of these are the case, a map is still an "abstraction" of the real world (Longley et al., p. 128). It glosses over certain details while highlighting others, and in doing so it inherently generates uncertainty and the potential for error. According to the Dutch geostatistician Gerard Heuvelink, as paraphrased by Longley et. al, "accuracy [is] the difference between reality and our representation of reality" (p. 128, emphasis in original). We will never achieve perfect accuracy, and that's okay. But paying attention to accuracy, and by the same token, uncertainty, is key to producing and interpreting geographic research as well as possible.
References: Longley, P. A., M. F. Goodchild, D. J. Maguire, and D. W. Rhind. 2008. Geographical information systems and science 2nd ed. Chichester: Wiley.