PEST RISK ANALYSES FOR TEMPERATE FRUIT FLIES IN EXPORTED FRUITS
Location: Fruit and Vegetable Insect Research
Project Number: 5352-22430-001-22
Specific Cooperative Agreement
Start Date: Sep 12, 2011
End Date: Dec 31, 2014
Develop ecological niche models using existing and unpublished data on apple maggot and western cherry fruit fly for prediction of potential of the establishment and spread of these pests in tropical countries importing tree fruits from the Pacific Northwest.
Develop databases and predictive maps on the climate of the importing countries and range of suitable hosts for apple maggot and western cherry fruit fly as well as development of databases and predictive maps for areas where these pests are normally found in the U.S. The databases will be used in ecological niche modeling programs to determine: 1) the potential for establishment and spread of these pests in importing countries; and 2) identification of researchable data gaps that may be needed to improve the accuracy of the models and may be used in the development of risk assessments.
Ecological niche models, which are based on the relationships between organisms and features of the environment, are increasingly being used to model and map native and invasive species distributions. Combining statistical algorithms with a geographic information system (GIS), ecological niche models attempt to predict probability of occurrence of a species by using presence-only or presence-absence data in combination with environmental variables. These models are based on Hutchinson’s classical niche concept: the distributions of species are constrained by biotic and abiotic gradients (e.g. elevation, temperature and precipitation) and species interactions (e.g., competition and predation). Although ecological niche models have become popular tools for invasive species ecology, they are also increasingly being used for rare and endangered species, defining conservation priority areas, phylogeographic studies, and the potential impacts of climate change However; their potential use in pest risk assessment is not yet fully explored with the exception of a few studies. These techniques therefore have the potential to advance pest risk assessment.
We will use maximum entropy model or Maxent which is a presence-only method. Our choice of presence-only method is best suited to the proposed research because absence data for the species are not available or are unreliable as they may be in the early stages of invasion. Recent studies on distribution modeling of insect pests in different parts of the world have demonstrated the effectiveness of the Maxent model. Maxent is a general purpose, machine learning, non-parametric, predictive model that uses presence-only data. This method estimates the probability distribution of a species by finding the probability distribution of maximum entropy, which is a probability that is closest to uniform. It automatically includes variables interactions and can handle continuous and categorical predictor variables. It uses a set of features (e.g., linear, quadratic, product, threshold and hinge) which are functions of environmental variables that constrain the geographical distribution of a species. It uses a regularization parameter, which is determined empirically, to control model overfitting. Maxent generates an estimate of probability of presence of the species that varies from 0 (lowest probability) to 1 (highest probability). Maxent has consistently fared well in model comparison studies.
Model evaluation and validation: We will evaluate model performance using a number of threshold-dependent and threshold-independent metrics to allow for better overall evaluation (Lobo et al. 2008). Threshold-dependent metrics will include Cohen’s Kappa, sensitivity, specificity, percent correct classification, odds ratio, and true skill statistic or TSS. The threshold-independent evaluation measure will include Area Under the receiver operating characteristic (ROC) Curve, or AUC. Wherever possible we will validate our models using an independent data set. However, in case of unavailability of independent validation data we will use “split sample approach”, to create a quasi-independent dataset for model validations.