Embracing Ensemble Species Distribution Models to Inform At-Risk Species Status Assessments

Abstract

Conservation planning depends on reliable information regarding the geographic distribution of species. However, our knowledge of species' distributions is often incomplete, especially when species are cryptic, difficult to survey, or rare. The use of species distribution models has increased in recent years and proven a valuable tool to evaluate habitat suitability for species. However, practitioners have yet to fully adopt the potential of species distribution models to inform conservation efforts for information-limited species. Here, we describe a species distribution modeling approach for at-risk species that could better inform U.S. Fish and Wildlife Service’s species status assessments and help facilitate conservation decisions. We applied four modeling techniques (generalized additive, maximum entropy, generalized boosted, and weighted ensemble) to occurrence data for four at-risk species proposed for listing under the U.S. Endangered Species Act (Papaipema eryngii, Macbridea caroliniana, Scutellaria ocmulgee and Balduina atropurpurea) in the Southeastern U.S. The use of ensemble models reduced uncertainty caused by differences among modeling techniques, with a consequent improvement of predictive accuracy of fitted models. Incorporating an ensemble modeling approach into species status assessments and similar frameworks is likely to benefit survey efforts, inform recovery activities, and provide more robust status assessments for at-risk species. We emphasize that co-producing species distribution models in close collaboration with species experts has the potential to provide better calibration data and model refinements, which could ultimately improve reliance and use of model outputs.

Predicting suitable habitats of Melia azedarach L. in China using data mining

“…Additionally, we observed that the prediction accuracy of the XGBoost was very close to that of RF as the XGBoost has good generalization performance 35 . Although, previous studies have shown that MaxEnt, SVM, and GBM models performed well in simulating species suitability distribution 36 , 37 , our results have shown that the prediction accuracy of these models was intermediate relative to the performance of the seven tested models. These phenomena may indicate that species characteristics and sample size also have influence on the accuracy of species distribution models 38 .…”

Ecological niche modeling predicting the potential distribution of African horse sickness virus from 2020 to 2060

“…Besides, it had 0.99 with TSS and 0.95 in Kappa accuracy evaluation, indicating correctly classifying suitable and unsuitable territories of the world for AHSv. The ensemble model's accuracy metrics were better than individual models indicating the ensemble approach was the best choice than using individual models alone 25 – 27 . Bioclimatic variables like Solar radiation, mean maximum temperature, mean precipitation of the year, altitude and precipitation seasonality contributed 36.83%, 17.1%, 14.34%, 7.61%, and 6.4%, respectively.…”

Geographic Distribution of the Puerto Rican Harlequin Butterfly (Atlantea tulita): an Ensemble Modeling Approach

“…An AUC value of 0.5 is considered poor performance, while 1 would be excellent. Given the relative low number of presence records, we used a jackknife approach (Pearson et al 2007; Ramirez-Reyes et al 2021) where models are constructed using all presence and pseudoabsence records minus one observation ( n - 1). We used the omitted observation's environmental data to solve the n - 1 model and obtain a predicted value.…”