Deep-learning based population monitoring of the endangered plant species Gladiolus illyricus: lessons learned for implementation of a technology-based biodiversity monitoring approach

New technologies offer promising possibilities in biodiversity monitoring to increase standardization of sampling methods and improve cost efficiency. Among the former, uncrewed aerial systems (UAS) are widely used today to produce orthomosaics of a particular area. At the same time, computer-intensive methods for automated object detection within images have increased accordingly. While they are widely used in science, applied nature conservation makes little use of these methods. The current study aimed to test the applicability of UAS in combination with a deep-learning based object detection workflow in Schütt-Graschelitzen, a small-scale Natura 2000 protected area near Villach, Austria. For this purpose, we trained a YOLO_v8 algorithm with flowers of Gladiolus illyricus from an orthomosaic. The orthomosaic was split into about 1000 equally sized tiles with 80 tiles used for training and 20 tiles used for validation. For ground truthing, the individual inflorescences were counted manually. Our main findings indicated moderate model performance with the training and validation dataset and also with new data. Moderate – rather than strong – performance is likely a result of too little training data. While object detection worked considerably well, background revealed too high variability, making reliable classifications challenging. Comparing the different work steps (without UAS mission) suggests that creating a representative training dataset is the most time-intensive part of the workflow. For small areas and a single survey, this is likely not efficient compared to traditional field sampling methods. However, its efficiency increases with each resurvey event, as pretrained deep-learning models developed during prior monitoring cycles can be reused. This can reduce the amount of training data required in a subsequent survey. Additionally, UAS- and deep-learning based monitoring can help at sites with high sensitivity to trampling and favors large study areas, as its efficiency increases with the sample size area.