Ensemble approach for potential habitat mapping of invasive Prosopis spp. in Turkana, Kenya

Abstract

Aim Prosopis spp. are an invasive alien plant species native to the Americas and well adapted to thrive in arid environments. In Kenya, several remote‐sensing studies conclude that the genus is well established throughout the country and is rapidly invading new areas. This research aims to model the potential habitat of Prosopis spp. by using an ensemble model consisting of four species distribution models. Furthermore, environmental and expert knowledge‐based variables are assessed.LocationTurkana County, Kenya.MethodsWe collected and assessed a large number of environmental and expert knowledge‐based variables through variable correlation, collinearity, and bias tests. The variables were used for an ensemble model consisting of four species distribution models: (a) logistic regression, (b) maximum entropy, (c) random forest, and (d) Bayesian networks. The models were evaluated through a block cross‐validation providing statistical measures.ResultsThe best predictors for Prosopis spp. habitat are distance from water and built‐up areas, soil type, elevation, lithology, and temperature seasonality. All species distribution models achieved high accuracies while the ensemble model achieved the highest scores. Highly and moderately suitable Prosopis spp. habitat covers 6% and 9% of the study area, respectively.Main conclusionsBoth ensemble and individual models predict a high risk of continued invasion, confirming local observations and conceptions. Findings are valuable to stakeholders for managing invaded area, protecting areas at risk, and to raise awareness.

Integration of remote sensing and bioclimatic data for prediction of invasive species distribution in data-poor regions: a review on challenges and opportunities

“…Sample size and design in presence and absence data is also another challenge in SDM (Fithian et al 2014;Rocchini et al 2015a, b). Most researchers used the only presence or pseudo absence data for modeling prediction of species distribution as absence data for invasive species are difficult to obtain (Ng et al 2018;Václavík and Meentemeyer 2012). In addition, employing a reasonable sample size using appropriate techniques can enhance the prediction of species distribution (Xie et al 2008).…”

“…Some other researchers first compare different models and then use the best performing model for prediction (e.g. Engler et al 2013;Ng et al 2018;Shiferaw et al 2019a). Other papers suggest that ensemble models are preferable over single models (e.g.…”

“…Other papers suggest that ensemble models are preferable over single models (e.g. ; Früh et al 2018;Ng et al 2018)). As indicated in Fig.…”

“…5, GLM and MaxEnt are largely used in bioclimatic and remote sensing based research. This might be attributed to the challenge of collecting absence data for invasive species (Ng et al 2018;Václavík and Meentemeyer 2012). MaxEnt is popular in SDM, through evaluating the quality of SDM is difficult (Filho et al 2010;West et al 2014).…”

“…Hence, as eradication of invasive species is difficult and costly, early detection and prevention, using an integrated data source, need to be prioritized for its management (Paz-Kagan et al 2019;West et al 2014;Zimmermann et al 2007). Although mapping current distribution and modeling of suitable habitat for invasive species are of particular interest, published reports are limited (Fischer et al 2013;Ng et al 2018).…”