The AgroInsect dataset comprises 1,510 images and 1,654 annotations, capturing a representative diversity of agriculturally relevant insect species under real-world field conditions. The distribution of samples across classes reflects natural occurrence patterns observed in field surveys and the intrinsic variability of the public data sources used in its construction, including iNaturalist and the Global Biodiversity Information Facility (GBIF). In several cases, a single image contains multiple insects, resulting in a higher number of annotations than images, with each detected specimen treated as an independent instance.The dataset is organized according to the widely adopted YOLO format, in which image files and their corresponding annotation files are stored separately. Annotations were generated using the Label Studio platform and follow the YOLO standard, where each annotation line consists of a class identifier and the normalized coordinates of the associated bounding box. This structure allows the dataset to be directly used for training and evaluating deep learning–based object detection models without additional preprocessing.AgroInsect was developed to support the training, evaluation, and benchmarking of machine learning and deep learning approaches for insect classification and agricultural pest detection. The images represent realistic field scenarios, characterized by variations in illumination, partial occlusions, different insect developmental stages, and complex backgrounds. These characteristics make the dataset particularly suitable for assessing the robustness and generalization capability of computer vision models.The practical applicability of AgroInsect has been demonstrated in studies employing state-of-the-art object detection architectures, including models from the YOLO family, optimized variants for embedded and edge computing environments using TensorFlow Lite (TFLite), as well as frameworks based on Detectron2.The dataset includes four insect classes of high agricultural relevance—Diabrotica speciosa, Dalbulus maidis, Diceraeus spp., and Spodoptera frugiperda. These species were selected due to their significant economic impact on soybean and maize production systems in the state of Mato Grosso and other major agricultural regions of Brazil. In this context, AgroInsect constitutes a strategic resource for applied research in digital agriculture, fostering the development of artificial intelligence–based solutions for pest monitoring and supporting sustainable agricultural practices.

The AgroInsect dataset comprises 1,510 images and 1,654 annotations, capturing a representative diversity of agriculturally relevant insect species under real-world field conditions. The distribution of samples across classes reflects natural occurrence patterns observed in field surveys and the intrinsic variability of the public data sources used in its construction, including iNaturalist and the Global Biodiversity Information Facility (GBIF). In several cases, a single image contains multiple insects, resulting in a higher number of annotations than images, with each detected specimen treated as an independent instance.The dataset is organized according to the widely adopted YOLO format, in which image files and their corresponding annotation files are stored separately. Annotations were generated using the Label Studio platform and follow the YOLO standard, where each annotation line consists of a class identifier and the normalized coordinates of the associated bounding box. This structure allows the dataset to be directly used for training and evaluating deep learning–based object detection models without additional preprocessing.AgroInsect was developed to support the training, evaluation, and benchmarking of machine learning and deep learning approaches for insect classification and agricultural pest detection. The images represent realistic field scenarios, characterized by variations in illumination, partial occlusions, different insect developmental stages, and complex backgrounds. These characteristics make the dataset particularly suitable for assessing the robustness and generalization capability of computer vision models.The practical applicability of AgroInsect has been demonstrated in studies employing state-of-the-art object detection architectures, including models from the YOLO family, optimized variants for embedded and edge computing environments using TensorFlow Lite (TFLite), as well as frameworks based on Detectron2.The dataset includes four insect classes of high agricultural relevance—Diabrotica speciosa, Dalbulus maidis, Diceraeus spp., and Spodoptera frugiperda. These species were selected due to their significant economic impact on soybean and maize production systems in the state of Mato Grosso and other major agricultural regions of Brazil. In this context, AgroInsect constitutes a strategic resource for applied research in digital agriculture, fostering the development of artificial intelligence–based solutions for pest monitoring and supporting sustainable agricultural practices.