Investigating key drivers influencing AI-based detection and identification of plants

Abstract

In recent years, AI-driven platforms have transformed citizen science by collecting and generating valuable records of living organisms for monitoring biological data. Many applications utilize visual similarity and geospatial information to identify species based on photographs. This study investigates how location impacts plant identifications made by iNaturalist, particularly in detecting invasive alien plants (IAP) that are not established in an area. We also compare the accuracy of iNaturalist and PlantNet while exploring potential biases. To assess iNaturalist’s taxonomic accuracy under varying location parameters, specimens of plants that are native and naturalized in Ontario, termed “established plants” for the purpose of this study, were collected and photographed (n = 61) and photographs of plants from Canada’s regulated pest list, which are either not present or have a very limited distribution in Canada, termed “outsider plants” for the purpose of this study were exported from iNaturalist and GBIF (n = 402). We used photographs of the established plants to compare taxonomic accuracy between applications, considering factors such as plant families, distribution status, and visible parts. A scoring system was established, and a cumulative linked mixed model was applied to analyze taxonomic accuracy. Our findings reveal that restricting location significantly hinders iNaturalist’s ability to identify IAP, highlighting the potential for missed detections. While sample size limitations prevented a robust comparison between applications, we also found significantly lower identification accuracy for species in the Poaceae family and for photographs featuring only leaves. Ultimately, recognizing the influence of location is essential for effectively monitoring IAP and leveraging iNaturalist as a tool for early detection.