Testing and training data for machine learning models to detect, classify and count blackbirds damaging agriculture using drone-based imagery

We used drones to capture images of mixed-species blackbird (Icteridae) flocks damaging sunflower (Helianthus annuus) in North Dakota. Images included several blackbirds (Icteridae) that breed in North Dakota and are considered agricultural pests, including red-winged blackbirds (RWBL) (Agelaius phoeniceus), common grackles (Quiscalus quiscula), yellow-headed blackbirds (Xanthocephalus xanthocephalus), brown-headed cowbirds (Molothrus ater), and European starlings (Sturnidae: Sturnus vulgaris). This study was implemented between September 2021 and October 2022 in multiple counties in North Dakota, USA, where blackbird damage to sunflowers is prevalent. We simultaneously hazed and captured video and photographs of flocks with the drone, thus images consist of airborne flocks with sky, green vegetation, or tan vegetation backgrounds. Images were used to train and test two models: 1) a ResNet-18 convolutional neural network (CNN) model to detect flocks of varying size and distance from the camera and 2) Faster Region-based Convolutional Neural Network (Faster-RCNN) models to detect individual blackbirds, classify individual blackbirds by species and for RWBL sex and age class, and count blackbirds. The Faster-RCNN model required individual birds in the images to be manually annotated by trained biologists for model training. This data publication contains the data and R code used to analyze these data, as well as the 400 images used in the models to detect blackbird flocks, the 131 images used in the models to detect and classify individual blackbirds, and the 131 image annotation files.
We designed this study to assess efficacy of drone-based aerial imagery combined with deep learning algorithms to accurately detect mixed-species blackbird flocks, as well as detect, classify, and count individual birds on varying backgrounds.
For more information about this study and these data, see Duttenhefner et al. (2025).