Transportation networks such as highways and railroad lines are important corridors for the spread of invasive species. In this project, an automated detection and mapping of two invasive alien plant species was developed and applied along highways in the Swiss Central Plateau, based on artificial intelligence (AI, Deep Learning).
The project was carried out in cooperation with the Computer Vision Lab at ETH Zurich (Dr. Radu Timofte) and demonstrates the possibilities of artificial intelligence in automated species detection and mapping. It documents the current distribution of invasive alien plant species and provides detailed information for practical use (green maintenance, management of invasive species) as well as for current research in the field of biological invasions.
The narrow-leaved ragwort (Senecio inaequidens) from South Africa and the tree of heaven (Ailanthus altissima) from China served as test species. While the tree of heaven can lead to increased maintenance costs due to its rapid growth and can affect road safety, the narrow-leaved ragwort is particularly problematic for agriculture if the species would spread from streets to crops, meadows and pastures. Ingredients of the ragwort can poison the liver of livestock and harm humans. Both species are on the «Black List» of invasive neophytes in Switzerland. The tree of heaven is listed in the European Union on the «List of invasive alien species of Union concern» (Union list) and the narrow-leaved ragwort is one of the species that is prohibited in Switzerland under the «Release Ordinance».
For the mapping, side and median strips were filmed with two cameras at an average speed of 82 km/h from a car. The mapped highway sections with a length of c. 690 km correspond to about half of the entire Swiss highway network. They were recorded three times during the vegetation period (May, August, October). To train and test the neural networks, experts marked the presences and absences of the two invasive species in images of the films - in the case of the narrow-leaved ragwort, additionally classified into flowering and non-flowering occurrences. The labeling of the species was done per image in 15 tiles of the same size. Different neural networks were tested, of which ResNet152 proved to be the most suitable for species identification. The detected occurrences of the tree of heaven agreed in 98.5% of the cases with the information of the experts, whereby 88.6% of the labeled image tiles were found. In the case of the narrow-leaved ragwort, 97.7% were correctly detected and 87.5% were found.
For mapping, images were extracted from the films every five meters and the two invasive species were automatically detected. The mapping is thus based on around 1.35 million georeferenced images with over 20 million automatically processed image tiles. This provided detailed distribution maps with good reproducibility.
The two invasive alien species are currently growing more frequently on the central strips than on the side strips of highways in the Swiss Central Plateau, but otherwise show different distributions. The tree of heaven is found especially near urban areas and is still missing on larger sections of the highways. In contrast, the narrow-leaved ragwort already grows on almost the entire highway network. While the spread of the tree of heaven is obviously dependent on the seed dispersal from ornamental and street trees in the settlement area, the ragwort spreads by itself along the highways. Analyses on the level of the territorial units responsible for maintenance also show an influence of different green maintenance on the occurrences of the two invasive species.
The developed deep learning based mapping approach has proven to be reliable and efficient. It produces well reproducible, spatially high-resolution distribution maps that cannot be provides with reasonable effort using classical expert mapping. The approach is therefore also suitable for monitoring invasive alien plant species on larger road networks. This also applies to other species which can be well recognized in the images. It can be assumed that with the current development in the fields of artificial intelligence and camera technology, corresponding applications of automated species detection and mapping will rapidly gain in importance for research and practice.
2018 - 2020