Early detection of caterpillars using artificial intelligence

Entomofauna are essential for agricultural productivity and ecological balance. Understanding interactions among pollinators and pests is key to sustainable agriculture. Caterpillars, as pests, damage crops by feeding on tissues and spreading diseases, and reducing the yield and its quality. Therefore, early detection of caterpillars is essential for agriculture, it can be achieved using artificial intelligence embedded into agrobots or unmanned aerial vehicles following the concept of precision farming. This study introduces the dataset Caterpillar640, which contains 1,300 images of caterpillars, which were captured in natural conditions in Latvia. The images are annotated and prepared for artificial intelligence training by using YOLO architecture. The collected dataset is publicly available and distributed under an open license, providing valuable resources for research in precision farming and biodiversity monitoring. The artificial intelligence was developed using three convolution neural network architectures: YOLOv8, YOLOv9 and YOLOv10. The medium-sized models (YOLOv8m, YOLOv9m, and YOLOv10m) were selected for the comparison experiment. The comparison showed that YOLOv8m achieved the best results 0.887 mAP@0.50, YOLOv9m and YOLOv10m achieved 0.873 mAP@0.50 and 0.859 mAP@0.50, respectively. The experiment was conducted by repeating the training of each neural network model five times. The trained models can be embedded into IoT sensors, agrobots and drones for autonomous monitoring of fields, orchards and gardens offering an efficient solution for pest and biodiversity monitoring. Additionally, the dataset Caterpillar640 can be applied by researchers to train general convolutional neural networks for agricultural tasks combining it with other agricultural datasets that can impact on the accuracy of pre-trained models.