Modelling urban bird breeding ranges with remotely sensed heterogeneity in plant traits using a random forest

Birds strongly respond to vegetation structure and composition, yet typical habitat models based on earth observation (EO) data use pre-classified data such as land use state classes for the habitat modelling. Since this neglects factors of internal spatial composition of the land use classes, we propose a new scheme of deriving multiple continuous indicators of urban vegetation heterogeneity using high-resolution earth observation datasets. The deployed concepts encompass spectral trait variations for the quantification of vegetation heterogeneity as well as subpixel vegetation fractions for the determination of the density of vegetation. Both indicators are derived from RapidEye data, thus featuring a continuous resolution of 5 meters. Using these indicators of plant heterogeneity and quantity as predictors, we can model the breeding bird habitats with a random forest machine learning classifier for our case study Leipzig while exclusively using one input dataset. Separate models are trained for the breeding ranges of 60 urban bird species (including 10 on the German red list), featuring medium to high accuracies (54–87%). Analysing similarities between models regarding variable importance of the single predictors allows species groups to be discriminated based on their preferences and dependencies regarding the amount of vegetation on the one hand, and its structure and heterogeneity on the other. The combination of continuous high-resolution EO data paired with a machine learning technique creates novel and very detailed insights into the ecology of the urban avifauna opening up possibilities of analysing and optimising different greenspace management schemes or future urban developments concerning overall bird species diversity or single species under threat of local extinction.