Does density-dependent dispersal explain mountain pine beetle spread? Performance of models across years and regions in BC (Canada)

Background/Question/Methods

The population dynamics of many irruptive forest insects are spatially autocorrelated  and synchronized. For mountain pine beetle (Dendroctonus ponderosae), the proximity of other beetle infestations (i.e. “beetle pressure”) is an important predictor of infestation risk. Here, we ask whether density-dependent formulations of beetle pressure predict new infestations better than alternatives where influence is proportional to or independent of infestation severity. Finding a good predictive model of spatial dependence is of interest for management, and proponents of pattern-oriented modeling have also argued that model comparison can yield insight into pattern generating processes. We vary dispersal kernel (1-40 km average distance, radially symmetric exponential and Gaussian kernels) and the sources of dispersing beetles (source intensity) to generate 296 alternative models. Source intensity scenarios include: all infestations contribute equally; intensity is proportional to severity; and intensity increases with cumulative impact (dispersal rate increases as resources are depleted). We assess model performance across 20 years and 41 forest districts in British Columbia (20 years x 41 districts = 240 cases). For each case, we evaluate the relative ability of alternative models to explain the pattern of new infestations. Model performance measures include AIC, AIC weight, R², and AUC.

Results/Conclusions

We find that the simplest intensity scenario (i.e. all infestations contribute equally) is most often best. The best dispersal kernel varies widely between cases. Models with high average dispersal distance (40 km) tend to perform better in outbreak conditions (infestation start probability >10%), and in relatively flat, pine-dominated districts (through the Chilcotin Plateau and Prince George forest region). Thus, we show that the extent of spatial dependence varies with geography and outbreak conditions, but in most circumstances new infestations arise near to other new (low severity) infestations. We argue that infestation probability is lower near severely infested areas because remaining uninfested areas are relatively poor habitat, not because beetle dispersal pressure is low. This result highlights inherent difficulties in inferring generating processes from observed patterns.