PS 11-134 - Using Machine-Learning Models to identify crown-class condition of EAB-infested forests.

Abstract: A classical biocontrol program was developed and implemented by the U.S. Department of Agriculture against the invasive beetle Agrilus planipennisi (emerald ash borer, EAB) after its first detection in the U.S. in 2002. The EAB biocontrol program has released several introduced parasitoid wasps as biocontrol agents across 32 EAB-infested states. While recent field studies have observed the successful establishment of these agents and significant EAB suppression in release sites, the contribution of this EAB biocontrol program to ash recovery and regrowth is currently not known. This study will establish whether machine learning (ML) models and drone photography can serve as a new fieldwork methodology to identify crown class conditions and forest canopy cover in EAB-infested forests and whether the crown class conditions of EAB-infested forests are recovering over time given biocontrol deployment.



In the summer of 2022, 50 x 2 meter transect surveys were conducted at field sites with confirmed EAB sightings across the New England region. Each transect survey inspected ash trees > 5 cm in diameter for visible signs of EAB infestations, and each tree was classified on a qualitative 1-3 scale of crown class condition, a popular indicator of forest canopy health. Simultaneously, a DJI Air2S commercial drone was flown at 50 meters above the forest canopy level and collected aerial imagery of the forest canopy cover for use in the machine learning models.



A suite of Convolutional Neural Network (CNN) image-classification models coded in Jupyter Notebook were trained on these ~4,000 aerial tree images to test their ability to predict forest crown class conditions. To evaluate model performance, a variety of metrics were used, including confusion matrices, t-SNE plots, and F1 scores. Preliminary confusion matrices for a MobileNet model indicate 60% and 55% accuracy for predicting class 1 and 2 respectively, but a 39% accuracy for class 3.



This study suggests that ML models can be used to detect EAB damage in native ash forests, and with increased data and high-resolution imagery, ML can better assess both the impact that the EAB has on native ash trees, as well as the impact biocontrol agents have on both the EAB and native ash forests after deployment. After a sufficient benchmark model is developed, its predictions will then be compared to both the field transect surveys collected in 2022 and new field sites located in New England.