The creation of model-based USNVC vegetation maps for the Pacific Northwest national parks involved training data collection at 6000 field plots representing 294 distinct vegetation associations—far too many to map. In such cases, National Park Service methods suggest lumping to higher levels of the NVC hierarchy. Based on this approach, our initial draft maps consisted of 40 map classes formed from vegetation alliances and groups.
We assessed the separability of the resulting map classes for purposes of mapping and field identification, using Random Forests modeling and floristic similarity analyses, respectively. Unfortunately, hierarchical lumping did not produce well-distinguished map classes. On average, the map class an association was assigned to received just 0.5% more model votes than its second best fit. Floristics were also not well-separated: association-level floristics were only 3.6% more similar to their assigned class than to their second best fit.
We used a data-driven approach to refine the classification to reduce confusion. Starting with map classes based on highly distinct associations, we grew them in a stepwise fashion based on each association’s modeling and floristic fit to each of the nascent map classes. The consequences of each assignment were assessed on an individual and a classification-wide basis.
The resulting map classification had 43 vegetated map classes. On average, the map class an association was assigned to received 5.7% more model votes than its second best fit, a more than 11-fold improvement over modeling based on NVC hierarchy. Floristic separability also improved markedly: association-level floristics were 8.2% more similar to their assigned class than to their second best fit.
The revised classification maintained the key ecological communities that local ecologists and NPS staff expect in these maps; draft map class descriptions have been approved by local ecologists and park management staff. Furthermore, the crosswalk is consistent across Parks, and the classification does a good job at representing cross-Park ecological similarities and differences. Two-thirds of the map classes are shared, while one-third are unique to one Park. If a map class was not represented by enough plots to map at a particular Park, it was lumped with the next most similar type, affecting just 0.9% of field plots.
Our results suggest that this approach will provide increases in both achievable map accuracy and consistency of map class field identification, increasing the utility of the maps while maintaining key USNVC concepts.