Dataset: Detecting leaf-cutting ants through computer vision in vineyards
Description
Between 20% and 40% of global agricultural production is lost annually due to pests, generating significant economic impacts—estimated at 70 billion USD from invasive insects alone—according to the Food and Agriculture Organization of the United Nations (FAO). In addition, the expected rise in average global temperature is likely to increase the destructive potential of leaf-cutting ants [1], reinforcing the need for sustainable pest management strategies in modern agriculture. In this context, the present dataset aims to support the identification and geolocation of leaf-cutting ants and their foraging trails. The provided images and ground truth data were generated through the use of agricultural machinery equipped with cameras, combined with image preprocessing techniques to detect ants and their trails in real-world conditions. Data collection was conducted in a vineyard located in Uruguay, providing a valuable real-world case study to advance the Technology Readiness Level (TRL) of AI-based solutions aimed at supporting Smart Agriculture practices.
Files
Steps to reproduce
The provided dataset comprises a total of 3033 annotated images (2,912 images in the training set, 85 in the validation set, and 36 in the test set). Annotations were made on Roboflow.com in YOLOv8 format and include four classes: ant trails, A. lundi nests, A. heyeri nests, and vineyard damage. The related files are organized according to the YOLOv8 format compatible with the Ultralytics framework. In summary, the database is organized as follows: images folder: Contains the input image files (e.g., .jpg, .png), divided into training, validation, and optionally test sets. labels folder: Contains the corresponding annotation files in .txt format, with the same filenames as the images. data.yaml: Configuration file that defines the dataset parameters such as the path to the dataset root directory and a list of class names indexed by label IDs. You can use the image-based dataset to train a Yolo v8 model or another version of Yolo using, for example, the train.py command in a suitable environment, and use the generated weights to perform inferences. The present dataset is intended to support the detection of leaf-cutting ants in real-world scenarios. It includes annotations for ant trails, Atta lundi nests, Atta heyeri nests, and vineyard damage. The authors invite researchers from other regions and institutions to contribute to the expansion of the dataset, particularly by incorporating additional scenarios, species, or characteristics that were not covered within the scope of this study.