COS 57-5 - How well do vegetation classification systems capture long-term changes?

Abstract: Developing classification systems from field-collected vegetation data has been a common practice in vegetation science for decades. Such “expert classifications” have served as a baseline to assess shifts in vegetation over time and have aided in the optimization of conservation and management efforts, albeit mostly at small geographic scales. Over the past couple decades, there has been a push towards integrating the data from these studies into a classification that represents a broad geographic area (e.g., United States National Vegetation Classification; USNVC). Given the effort that goes into their development, it is important that these large-scale classification systems are able to account for vegetation dynamics over the long and short term. Our objective is to quantitatively assess the robustness of the USNCV. Specifically, in a supervised classification methodology, we use clustering methods to test how well current USNVC units agree with numerical (i.e., clustering) solutions and how well existing units represent new and increasingly different data points.


For our analysis, we used three vegetation data sets that each span several decades between original sampling and re-sampling: eastern deciduous forest in North Carolina (sampled in 1977 and 2009), Chihuahua Desert Grassland in Texas (sampled in 1981, 1996, 2007 and 2014), and montane and subalpine forests in Colorado (sampled in 1972 and 2012). Cluster and USNVC groupings resulted in similar noise levels (i.e., proportion of plots that do not fit in existing clusters), indicating similar robustness of both approaches when presented with new data. Larger datasets with non-random sampling locations performed better (as little as 12% noise) than small data sets with random sampling locations (up to 76% noise). Among our three examples, the more xeric Big Bend data seems to be naturally slower in vegetation change and thus we expect the classification to last longer as well. Overall, our results indicate that the robustness of current classification systems should not be assessed in a single metric but needs to be carefully evaluated on a case by case basis, depending on the type and quantity of data on which a vegetation unit is based.