Vineyard Image Dataset for Disease Detection, Cluster and Canopy Estimation

Dataset The dataset consists of vineyard images collected using the Viewer device during field trials of the EU Project OpenAgri, Sustainable Innovation Pilot 1. The color images were captured in a vineyard of the Savatiano variety located in Spata, Athens, in July 2025. Images were captured sequentially along vineyard rows and are organized into three subsets depending on the functionality used: yield estimation, disease detection, and canopy characterization. Each subset contains two subfolders: one containing the images and another containing the annotations in YOLO format. The annotation folders follow the standard YOLO dataset structure. Each image has a corresponding .txt file with the same filename containing the object annotations. Each line in the annotation file represents one labeled object and includes the class ID along with normalized spatial information relative to the image dimensions, provided either as bounding box parameters (center coordinates and width/height) for detection tasks or as segmentation polygons for canopy-related analysis such as vigor estimation. Additional configuration files such as obj.names, obj.data, and train.txt define the dataset classes, dataset paths, and the list of training images used for YOLO model training. If an image listed in the training dataset does not have a corresponding annotation entry, this indicates that no target object is present in that image. The yield estimation subset contains annotated images of Savatiano grape clusters at BBCH 75–79 stage, enabling cluster detection and counting for accurate yield prediction. The disease detection subset includes annotations of artificial targets representing downy mildew (Peronospora) symptoms placed on grapevine leaves for controlled detection experiments. The vigor detection subset includes annotations of vine structures such as trunk and foliage, enabling segmentation-based analysis of canopy geometry and spatial distribution. These annotations support the generation of vigor maps and variable-rate application (VRA) strategies by capturing variations in canopy development along vineyard rows. Models Separate YOLO-based models were trained for each of the three functionalities tested during the first season of the OpenAgri project: vigor estimation, disease detection, and yield estimation. For each functionality, two model variants are provided, corresponding to two deployment targets: a cloud version, intended for server-side inference with higher accuracy, and an edge version, optimized for on-device deployment on the Viewer hardware with reduced computational requirements. The primary difference between the two variants is model accuracy, with the cloud version achieving higher detection and segmentation performance at the cost of greater computational demand.Each functionality folder contains two subfolders — one for the edge version and one for the cloud version — each structured identically. The trained model is provided in two formats: a PyTorch checkpoint (model.pt, ~153 MB) for cloud deployment or further fine-tuning, and an ONNX export (model.onnx, ~49 MB, IR version 8, opset 17, PyTorch 2.7.0) for edge deployment across a wide range of inference runtimes. Both variants accept RGB images of 640×640 pixels as input.Each folder also includes a labels.txt file listing the class names recognized by the model, one per line. The class definitions differ per functionality: the yield estimation model uses the class grape_savatiano_cluster; the disease detection model uses grape_savatiano_downy_mildew_artificial; and the vigor estimation model uses two classes, trunk and foliage.Finally, a results.csv file records the per-epoch training and validation metrics logged during model training. Each row corresponds to one epoch and includes training losses (box, segmentation, classification, DFL), validation metrics for both the detection head (precision, recall, mAP@50, mAP@50–95) and the segmentation mask head, validation losses, and learning rate values.All models were trained for 400 epochs. Training and validation losses converged steadily across all runs. As an indicative example, the disease detection cloud model reached a peak bounding-box mAP@50 of 0.136 (epoch 349) and a peak segmentation mAP@50 of 0.135 (epoch 352), with final-epoch precision of approximately 0.70 for both heads.