ArTreeficial in Advancing Novel Datasets and Developing Deep Learning Models for Spotted Lanternfly Identification

Lycorma delicatula, the spotted lanternfly (SLF), is an invasive species that has been present in the United States of America for about a decade. Its prolific feeding significantly damages a large variety of plants. Although artificial intelligence (AI) powered technologies are labor-saving, there is a significant dearth of data and technology directly addressing this urgent ecological issue-which is ongoing with previous work in ArTreeficial-and robust models are needed to handle less than photographically-ideal data. This paper develops a novel SLF original image database of unprecedented size to date: 500 original images, augmented to 3000 images total. On a YOLOv8 model, models trained on original and augmented images increased experimental accuracy to 90% (compared to 87<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup>), reducing false positive identifications by three times (9.8% versus 30.9%), compared to the model trained only on original images. Our results suggest that image augmentation not only results in a more robust model under low-fidelity environments, but also demonstrates emergent properties which partial augmentation models fail to provide.