Image background assessment as a novel technique for insect microhabitat identification

Habitat fragmentation under increased urbanisation, industrial agriculture and land clearing, are changing the way insects occupy habitat. Some species are highly adaptable and may occupy urbanised areas, utilising anthropogenic microhabitat-scale features. Other species are dependent on natural elements of their habitats, having to locate small regions of natural microhabitat within increasingly hostile landscapes. Consequently, humans are encountering insects in new settings. Identifying and analysing insects’ use of natural and anthropogenic microhabitats is therefore important to assess their responses to a changing environment, for instance to improve pollination or manage invasive pests. But such studies are labour-intensive. Traditional studies of insect microhabitat use can now be supplemented by machine learning-based insect image analysis. Typically, research has focused on automatic insect classification, but valuable data appearing in image backgrounds has been ignored. In this research, we analysed the backgrounds of insect images available in the Atlas of Living Australia database to determine the microhabitats in which they were commonly photographed. We analysed the image backgrounds of three globally distributed insect species that are common across Australia: Drone flies (Eristalis tenax), European honey bees (Apis mellifera), and European wasps (Vespula germanica). Image backgrounds were classified broadly as either natural or anthropogenic using computer vision and machine learning tools benchmarked against a manual classification algorithm. Our automated image background classification achieved 97.4% accuracy when compared against manual classification. Mis-classifications were scarce, usually less than 1%, and primarily for backgrounds of wood and soil or bare ground. Our results indicate that drone flies and European honey bees were predominantly photographed against natural backgrounds (flies manual classifier 95±3%, automated classifier 94%, bees 89±2%,87%), implying frequent observations by humans in natural microhabitat. European wasps were less frequently photographed against natural backgrounds (70±6%,63%). Within this data set, observations of the wasps in anthropogenic microhabitats were more common than for flies and bees. Our results are aligned with the expectation that the wasps are relatively well-suited to urban environments, and that European honey bees and drone flies utilise natural features of their environment. In general, although biases in data collected without formal protocols limits their application, our new automated approach for image background analysis can provide valuable data about insects’ interactions with humans, our artefacts, and natural features of their environments.