Field-Based Image Dataset for Visible Red Palm Weevil Detection in Date-Palm Environments

This dataset supports research on the automated detection of visible red palm weevils (Rhynchophorus ferrugineus) in real-world agricultural environments using machine vision and object-detection models.

The dataset was developed for the study titled “Field-Based Machine Vision for Visible Red Palm Weevil Detection Using YOLOv12 and RF-DETR.” It contains field photographs captured at a commercial date-palm farm in Al-Kharga Oasis, Egypt, during repeated visits under natural and uncontrolled outdoor lighting conditions. Images were acquired using multiple smartphone cameras, including an iPhone 15, Samsung Galaxy S21, and vivo V2109, to introduce variation in image resolution, viewpoint, color rendering, illumination, background complexity, object size, and image quality.

The original study dataset comprised 1,034 images:

974 field-acquired images from the commercial date-palm farm.

60 licensed stock images used to increase background and viewpoint diversity.

Due to third-party licensing restrictions, the licensed stock images are not included in the publicly distributed dataset. The Harvard Dataverse deposit therefore contains the field-acquired images, corresponding annotations, and any distributable derived metadata or dataset documentation.

Each visible red palm weevil was annotated using a tightly fitted bounding box under a single object class, RPW. Partially occluded insects were annotated according to their estimated full extent, while overlapping insects were labeled separately. The dataset also includes background images containing no visible red palm weevils, allowing object-detection models to learn from negative examples.

The complete raw dataset used in the study included:

126 images with no visible red palm weevil.
809 images containing one visible weevil.
93 images containing between two and six visible weevils.
6 images containing more than six visible weevils.
For model development, images were standardized to 640 × 640 pixels and processed using auto-orientation and adaptive contrast-limited histogram equalization. Derived augmented versions were generated using transformations such as horizontal flipping, rotation, zooming, cropping, shearing, hue and saturation adjustment, exposure variation, blur, noise injection, and grayscale conversion.

The dataset is intended for research in agricultural computer vision, small-object detection, pest monitoring, precision agriculture, and the comparative evaluation of one-stage and transformer-based object detectors. It is specifically designed for detecting externally visible adult red palm weevils and cannot be used to identify larvae or infestations hidden inside palm trunks.

Users should note that most field images were collected from a single farm and that the dataset contains a concentration of centrally positioned and single-instance objects. Cross-farm, cross-season, and cross-region validation is therefore recommended before operational deployment.