Mapping the distribution of an invasive tree using a crowdsourced geospatial data platform and deep learning

Abstract

Keeping maps of invasive plants (IP) updated is challenging because of the fast expansion they can undergone in the short term. Traditional mapping methods (i.e., systematic sampling) are resource-intensive and often unaffordable at large scales. Citizen science tools (e.g., iNaturalist) are emerging as complementary sources of information on IP occurrences, with these data centralized by the Global Biodiversity Information Facility (GBIF). However, citizen science records are often biased towards species with ornamental value and highly populated areas. Mapillary (www.mapillary.com) is a crowdsourced geospatial data platform that enables users to collect and upload georeferenced street-level images using smartphones while traveling by car, train, or bicycle. Images are captured at regular intervals, providing continuous visual coverage of routes, and are subsequently stored and made accessible through geographic coordinates. Mapillary images can be used as a complementary tool to identify and map conspicuous plant species, particularly those expanding through roadsides. Ailanthus altissima is an invasive tree, native to SE China, introduced in Europe in the 18 th century as ornamental and shade tree. It has secondarily expanded along roads and railways thanks to the profuse production of winged seeds with high germination capacity. Nowadays it is considered invasive in most European countries. We explore the utility of Mapillary –as source of images– combined with deep learning techniques for the automatic detection of A. altissima , to generate an updated map of occurrences in roadside environments of central Spain. An R script was developed to download georeferenced Mapillary images within defined coordinates. A total of 390 images containing A. altissima in different phenological stages were selected and annotated with bounding boxes in LabelImg. We used a patch-based strategy focused on visually distinctive parts of the image rather than whole individuals, resulting in 969 annotated instances (Fig. 1). The annotated data were used to train a deep learning object detection model that we developed in Python based on a convolutional neural network architecture (YOLOv8, Ultralytics), configured as a single-class detector to identify the presence of A. altissima . Given the fragmented nature of the annotations, model outputs were post-processed by spatially aggregating nearby detections into coherent zones of presence. Model performance was evaluated using a zone-based recall metric, achieving a detection rate of ~80% of annotated presence areas. The trained model will be applied to additional Mapillary images systematically sampled across the study area to generate a spatially explicit occurrence map of A. altissima in central Spain. Finally, the resulting distribution will be compared with GBIF records to assess consistency and complementarity between both data sources. This approach highlights the potential of combining street-level imagery and deep learning to support large-scale, cost-effective monitoring of invasive plant species in human-modified landscapes.