Machine learning to detect, classify, and count blackbirds damaging agriculture using drone-based imagery: Supporting AI-driven automation for deployment of damage management tools

As human populations expand and land use changes, human-wildlife conflicts are increasing, requiring cost-effective management tools that balance human well-being and wildlife conservation. Hazing devices are often used to mitigate conflicts between wildlife and agriculture with drones serving as both frightening devices and monitoring systems. Artificial intelligence (AI) enables automated detection, identification, and counting of wildlife, allowing for field-based Internet of Things (IoT) systems to selectively deploy tools when target species reach a critical threshold. We acquired drone imagery of mixed-species blackbird flocks – dominated by red-winged blackbirds (RWBL) Agelaius phoeniceus – damaging sunflowers ( Helianthus annuus ) in North Dakota (September–October 2021–2022). We trained a ResNet-18 Convolutional Neural Network (CNN) model to A) detect flocks (accuracy = 95.0 %); and Faster Region-based Convolutional Neural Network (Faster-RCNN) models to B) detect individual blackbirds (accuracy = 65.7 %, precision = 97.6 %), C) classify individual blackbirds by species and for RWBL sex and age class, and D) count blackbirds (% difference: birds = 37.5 %; male RWBL = 39.6 %; female RWBL = 43.1 %). The model correctly classified RWBL to species (87.6 %), sex, and age class (adult males = 89.8 %; hatch-year males = 27.6 %; females = 80.0 %). The RWBL were misclassified as other blackbirds that inflict crop damage, not non-target species. Variability in background (e.g., sky, green vegetation, tan vegetation) and complexity (e.g., contrast, texture), along with bird camouflage, required background removal to enhance performance when moving drones captured still images of moving targets. Camera orientation (e.g., depth perception, target overlap) and image quality (e.g., blurred, shadowed objects) affected detection, classification, and counts. Automated deployment reduces labor and wildlife habituation, increasing longevity and efficacy of management tools. • Increasing human-wildlife conflict highlights the need for effective management tools. • Our model accurately detected blackbird flocks damaging crops with >95 % average accuracy. • Faster-RCNN effectively detected and classified blackbirds by species, age, and sex. • Background elimination drastically increased automated detection performance. • Findings lay groundwork for autonomous bird monitoring and crop protection systems.