Benchmarking Invasive Alien Species Image Recognition Models for a Citizen Science Based Spatial Distribution Monitoring

Abstract. Recent developments in image recognition technology including artificial intelligence and machine learning led to an intensified research in computer vision models. This progress also allows advances for the collection of spatio-temporal data on Invasive Alien Species (IAS), in order to understand their geographical distribution and impact on the biodiversity loss. Citizen Science (CS) approaches already show successful solutions how the public can be involved in collecting spatio-temporal data on IAS, e.g. by using mobile applications for monitoring. Our work analyzes recently developed image-based species recognition models suitable for the monitoring of IAS in CS applications. We demonstrate how computer vision models can be benchmarked for such a use case and how their accuracy can be evaluated by testing them with IAS of European Union concern. We found out that available models have different strengths. Depending on which criteria (e.g. high species coverage, costs, maintenance, high accuracies) are considered as most important, it needs to be decided individually which model fits best. Using only one model alone may not necessarily be the best solution, thus combining multiple models or developing a new custom model can be desirable. Generally, cooperation with the model providers can be advantageous.