Searcharxiv⌕ Search

arXiv subjects

David M. Richardson

Publications and source records attributed to David M. Richardson.

1 recordsLinked to original sources

Automated detection of wild and cultivated plants in citizen science records

Citizen science plant occurrences may help the early detection of naturalisations of alien species, yet they are affected by misclassifications of cultivated plants as wild occurrences. To differentiate cultivated from wild observations, we proposed a decision framework for the manual annotation of records. We annotated this cultivation status for 1,491 iNaturalist records in Cape Town, comprising 12 invasive alien plant species. To automate the classification, we trained algorithms on annotated records based on 39 predictors from split into 4 sets: (i) basic record and observer features, (ii) visual features extracted from images, (iii) geographic features, and (iv) iNaturalist observer statistics. 11 Classification algorithms were trained on 8 combinations of predictor sets, to assess predictors importance, yielding 88 models evaluated with crossvalidation. 25.2% of all annotated records were cultivated, and 21.2% among Research Grade records, while 24.8% of annotated records were doubtful. Automated classification yielded an accuracy in distinguishing cultivated from wild records reaching 87%, improving over the iNaturalist cultivation tag (74%), and the accuracy was robust when predicting on new species (85%). Model performance was overall highest when combining iNaturalist-specific, geographic and visual predictors, but visual predictors alone performed better for certain species. Our results highlight the potential to automatically distinguish cultivated and wild-growing plant records to support invasion risk assessments and sanitise ecological analyses currently contaminated by cultivated records. Our framework is transferable to other crowdsourcing platforms, even though the manual annotation capacity can limit transferability.

q-bio.QM↗