Earth observation based indication for avian species distribution models using the spectral trait concept and machine learning in an urban setting. (April 2020)

• Record Type: Journal Article
• Authors: Wellmann, Thilo; Lausch, Angela; Scheuer, Sebastian; Haase, Dagmar
• Is Part Of: Ecological indicators. Volume 111(2020)
• Publication Date: 2020-04
• Subjects: Remote sensing -- Spectral traits -- Species distribution model -- Random forest -- Urban birds -- Machine learning; Environmental monitoring -- Periodicals; Environmental management -- Periodicals; Environmental impact analysis -- Periodicals; Environmental risk assessment -- Periodicals; Sustainable development -- Periodicals; 333.71405
• DOI: 10.1016/j.ecolind.2019.106029
• Languages: English
• ISSNs: 1470-160X

Abstract

Graphical abstract: Highlights: New methodology to create species distribution models (SDM) for bird species. Random forest is the most suitable machine learning technology for the SDMs. Texture metrics are the most important indicator describing bird-breeding ranges. High to medium accuracies (91–59%) for 44 bird species in an urban setting. Repeatable and cost effective methods for deriving high-resolution SDMs. Abstract: Birds respond strongly to vegetation structure and composition, yet typical species distribution models (SDMs) that incorporate Earth observation (EO) data use discrete land-use/cover data to model habitat suitability. Since this neglects factors of internal spatial composition and heterogeneity of EO data, we suggest a novel scheme deriving continuous indicators of vegetation heterogeneity from high-resolution EO data. The deployed concepts encompass vegetation fractions for determining vegetation density and spectral traits for the quantification of vegetation heterogeneity. Both indicators are derived from RapidEye data, thus featuring a continuous spatial resolution of 6.5 m. Using these indicators as predictors, we model breeding bird habitats using a random forest (RF) classifier for the city of Leipzig, Germany using a single EO image. SDMs are trained for the breeding sites of 44 urban bird species, featuring medium to very high accuracies (59–90%). Analysing similarities between the models regarding variable importance of single predictors allows species groups to be determined based on their preferences and dependencies regarding the amount of vegetation and its spatial and structural heterogeneity. When combining the SDMs, models of urban bird species richness can be derived. The combination of high-resolution EO data paired with the RF machine learning technique creates very detailed insights into the ecology of the urban avifauna, opening up opportunities of optimising greenspace management schemes or urban development in densifying cities concerning overall bird species richness or single species under threat of local extinction.