Using biased sampling data to model the distribution of invasive shot-hole borers in California

Abstract

The invasive shot-hole borers (ISHB), Coleoptera: Curculionidae: Scolytinae: Euwallacea spp. and their ambrosial fusaria symbionts were first detected in Southern California in 2003 and have since caused dieback and tree mortality in urban, agricultural, and natural environments. This research assesses the relative predictive ability of distribution models for ISHB. A series of 100-m resolution models were developed using the Maxent modeling tool to test the effects of three main parameters: length of sampling period, spatial extent, and spatial filtering. Five chronologically cumulative sampling periods, simulating stages of invasion, were used to train the models. To evaluate model predictions, we used independent test records from the 2012–2016 sampling period, as well as presence and absence data collected via 2018 field surveys. Testing the models against both datasets demonstrated that as length of sampling period increased, prediction errors decreased, and overall accuracy improved for predictions of ISHB host occurrence. All models achieved high presence-only, area under the receiver operating characteristic curve (AUCPO) values > 0.93 and correctly classified 87.7 ± 18.8% of independent test records, indicating high model performance regardless of the degree of temporal bias. Spatial filtering produced more discriminating results without compromising model sensitivity to test records. Sensitivity was consistently higher for models that used the larger spatial extent (state of California), which suggests that for an emerging species, larger backgrounds may produce less discriminating predictions. The relative contribution of the top environmental predictor variables, including minimum temperature of the coldest month, percent impervious surface, isothermality, and dry-season normalized difference vegetation index, differed according to the modeled spatial extent. Regardless of the parameters used, the study finds that ISHB distribution models reliably identified areas vulnerable to infestation. Such predictions, particularly those early in the study period, may have aided early containment and detection efforts.