A Bayesian belief network assessment of vegetation succession and spatial dynamics in response to fire and hydrological conditions in the Okefenokee National Wildlife Refuge, Georgia, USA

Background/Question/Methods

Vegetation within the Okefenokee National Wildlife Refuge (ONWR) is fire dependent. Wildfires are common and maintain the Okefenokee Swamp in a complex of wetland prairies, shrubs, and forest. In the past 12 years, three extensive fires have swept the swamp, burning nearly 80% of the refuge swamp wetlands. Fine resolution vegetation maps of the refuge and post-fire change assessments indicate that recent fires have reduced vegetation structural complexity in the swamp. Acquiring knowledge about post-fire recovery and burn patterns provides a valuable tool for managing the expansive landscape of ONWR, in particular for characterizing available fuels, determining fuel models, and understanding the potential for vegetation recovery as burn frequency and severity respond to climate dynamics. We have developed a spatially explicit Bayesian belief network (BBN) to forecast vegetation succession under different fire and hydrological condition scenarios. Our model is informed with vegetation maps developed from aerial photography and satellite imagery captured in 1977, 1990, 2001, 2008, and 2012, fire severity maps developed in 2002, 2007, and 2011, and historical data on precipitation, water level, and fire frequency. This approach spatially relates ecological factors (e.g., fire and hydrology) with vegetation community composition in a Geographic Information System.

Results/Conclusions

The BBN is populated with nodes derived from spatially explicit data from vegetation and fire severity maps and nodes describing key environmental variables including non-growing season precipitation, growing season water levels, and time since previous fires.  Initial application of the 2001 and 2008 vegetation maps suggests that the model prediction of vegetation composition and distribution is fairly accurate. The 10-fold cross-validation error was 22.6%, and commission errors generally exceeded omission errors, with greatest error for the “Bare Ground” and “Canopy Pines” vegetation classes. These types have the smallest coverage areas (0.3 and 1.9% of the total area, respectively) of the nine modeled vegetation types, and are found primarily in the swamp upland perimeter. Some error within the model would necessarily be derived from the original vegetation classifications, which were produced with overall accuracy (Khat) of 0.84 (2001) and 0.83 (2008). Error may be reduced as we incorporate data from the remaining vegetation maps (1977, 1990, 2012) and evaluate the BBN conditional probability tables in a sensitivity analysis to identify sources of variation for specific vegetation classes. This model is a tool to forecast vegetation succession under anticipated fire scenarios informed by climate projections for the region.