Iterative forecasting improves forecasts of spotted lanternfly invasion

Background/Question/Methods

Iterative ecological forecasting allows for continual model improvement through the incorporation of new data and scientific discoveries. Incorporating new scientific knowledge is especially important for models of invasive species spread when little is known about biological tolerances. Here we use a process-based spatio-temporal spread model called PoPS (Pest or Pathogen Spread) to show the continual improvement of the model through new data incorporation and show that when new scientific discoveries are incorporated into the model we can see larger increases in model skill. We iteratively use annually collected spotted lanternfly (SLF) (Lycorma delicatula) data from 2014 to 2021 from USDA APHIS and state departments of agriculture surveys; 2014 to 2020 data was used for calibration and 2021 for validation. Approximate Bayesian computation is used to update model parameters each year with new data. We have incorporated observations from the field on density dependant dispersal and long-distance dispersal along rail lines into the model. Additionally, a new lab study on SLF survival based on temperature was incorporated into the model to improve our previous estimates of SLF temperature tolerances. We use accuracy, precision, recall, specificity, and the Matthews Correlation Coefficient (MCC) to measure model skill across parameter sets and models.

Results/Conclusions

Our preliminary results show that model skill increases over time by approximately 15% from 2016 to 2020 as new data is incorporated into model parameter estimates. We compared random Cauchy long-distance dispersal vs. rail-directed long-distance dispersal, density-dependent dispersal vs. non-density-dependent dispersal, and updated temperature tolerance vs. non-updated temperature tolerance. Combinations of these model components were tested for a total of eight versions of the model. Both the updated temperature tolerances and rail-directed long-distance dispersal improved model accuracy by approximately 8% and when included together improved model accuracy by approximately 13%. The overall best model based on accuracy, precision, specificity, and MCC was the rail-directed long-distance dispersal with updated temperature tolerances and rail-directed long-distance dispersal with density dependant dispersal and updated temperature tolerances performed best based on recall. While density dependant dispersal didn’t have a significant impact on model skill based on any of our measures, it did improve trust in the model for stakeholders using it to determine surveillance and treatment strategies. Including features that improve decision-makers, trust is an important part of our participatory forecasting cycle that increases the use of forecasts as a decision support tool.