Application of UAV remote sensing and machine learning to model and map land use in urban gardens

Abstract

Abstract Urban gardens are an integral part of urban agricultural systems, contributing to ecosystem services, biodiversity and human wellbeing. These systems occur at fine scales, can be highly complex and therefore offer the opportunity to test mechanisms of ecological patterns and processes. The capacity to confidently characterize urban gardens and their land uses is still lacking, while it could provide the basis for assessing ecosystem service provision. Land classifications from remote sensing platforms are common at the landscape scale, but imagery often lacks the resolution required to map differences in land use of fine-scale systems such as urban gardens. Here, we present a workflow to model and map land use in urban gardens using imagery from an unoccupied aerial vehicle (UAV) and machine learning. Due to high resolutions (<5 cm) from image acquisition at low altitudes, UAV remote sensing is better suited to characterize urban land use. We mapped six common land uses in 10 urban community gardens, exhibiting distinct spatial arrangements. Our models had good predictive performance, reaching 80% overall prediction accuracy in independent validation and up to 95% when assessing model performance per cover class. Extracting spatial metrics from these land use classifications, we found that at the garden and plot scale, plant species richness can be estimated by the total area and patchiness of crops. Land use classifications like these can offer an accessible tool to assess complex urban habitats and justify the importance of urban agriculture as a service-providing system, contributing to the sustainability and livability of cities.