Modelling species habitat suitability from presence-only data using kernel density estimation

• Published in: Ecological indicators Vol. 93; pp. 387 - 396
• Main Authors: Zhang, Guiming, Zhu, A-Xing, Windels, Steve K., Qin, Cheng-Zhi
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.10.2018
• Subjects: aerial surveys; case studies; Ecological monitoring; ecosystems; environmental factors; environmental indicators; Habitat suitability modelling and mapping; habitats; Kernel density estimation; Minnesota; model validation; monitoring; national parks; Odocoileus virginianus; prediction; Presence-only data; probability; probability distribution; Resource availability; Minnesota; Habitat suitability modelling and mapping; Resource availability; Ecological monitoring; Presence-only data; Kernel density estimation
• ISSN: 1470-160X; 1872-7034
• DOI: 10.1016/j.ecolind.2018.04.002

•A novel approach for modelling habitat suitability from species presence-only data.•Suitability is computed based on species presence and background distributions.•Case study shows the approach performs well in terms of training and test accuracy.•Modelled suitability-environment responses are amenable to ecological interpretation.•The approach is useful for ecological monitoring at ecosystem or species level. We present a novel approach for modelling and mapping habitat suitability from species presence-only data that is useful for ecosystem and species monitoring. The approach models the relationship between species habitat suitability and environment conditions using probability distributions of species presence over environmental factors. Resource availability is an important issue for modelling habitat suitability from presence-only data, but it is in lack of consideration in many existing methods. Our approach accounts for resource availability by computing habitat suitability based on the ratio of species presence probability over environmental factors to background probability of environmental factors in the study area. A case study of modelling and mapping habitat suitability of the white-tailed deer (Odocoileus virginianus) using presence locations recorded in aerial surveys at Voyageurs National Park, Minnesota, USA was conducted to demonstrate the approach. Performance of the approach was evaluated through randomly splitting the presence locations into training data to build the model and test data to evaluate prediction accuracy of the model (repeated 100 times). Results show that the approach fit training data well (average training area under the curve AUC = 0.792, standard deviation SD = 0.029) and achieved better-than-random prediction accuracy (average test AUC = 0.664, SD = 0.025) that is comparable to the state-of-the-art MAXENT method (average training AUC = 0.784, SD = 0.021; average test AUC = 0.673, SD = 0.027). In addition, the suitability-environment responses modelled using our approach are more amenable to ecological interpretation compared to MAXENT. Compared to modelling habitat suitability purely based on species presence probability distribution (average training AUC = 0.743, SD = 0.030; average test AUC = 0.645, SD = 0.023), incorporating background distribution to account for resource availability effectively improved model performance. The proposed approach offers a flexible framework for modelling and mapping species habitat suitability from species presence-only data. The modelled species-environment responses and mapped species habitat suitability can be very useful for ecological monitoring at ecosystem or species level.