Automated mapping and identification of shrub individuals in South Africa's Fynbos biome using drone imagery and deep learning

Abstract

In order to understand population and community dynamics, many ecological studies require comprehensive knowledge of the spatial distribution of individual organisms, but obtaining this data is a time and labor-intensive process. In this study we develop a workflow to automatically determine the species of shrubs of the Proteaceae family in South Africa's Fynbos region from drone-based photogrammetric data. We applied deep learning to segment five species of shrub individuals from the background based on spectral and height information. The spectral-height model achieved an average prediction accuracy of 74.4%, compared to 61.6% when using spectral information alone. Despite the challenge in distinguishing sprawling shrubs from the background, which may be overcome with additional training data, the presented workflow holds promise for the efficient mapping of shrub communities.