Harmonized Tea Weed–Disease–Pest Object Detection Dataset (v1.0)

**DOI:** 10.5281/zenodo.23911768 **Dataset release:** Version 1.0**Repository:** Zenodo**Dataset type:** Annotated images for object detection**Annotation format:** YOLO bounding boxes**Classes:** Weed (0), Disease (1), Pest (2)**Associated study:** *Calibrated knowledge distillation for embedded multi-stressor detection in tea plantations* — Sikudhan Lucas Mpuhus and Zhao Jin. --- ## Overview This record deposits a harmonized three-class object-detection benchmark for detecting weeds, visible disease symptoms, and pests in tea imagery. It brings together three publicly available datasets under a single ontology, label format, and partition scheme, and distributes two fixed configurations: the original harmonized dataset (9,760 images; 12,849 annotated instances) and a moderately augmented configuration (9,975 images; 13,585 annotated instances) in which 215 generated training images were added to the training partition only. Validation and test partitions are identical between the two configurations. The deposit contains the complete image and label files for both configurations, the corresponding `data.yaml` files, a full `README.md`, and an `augmentation_protocol.md` documenting the augmentation procedure and its reproducibility limits. ## What this deposit contributes, and what it does not The authors of this record **did not collect the original photographs**. The original images remain the work and property of the creators of the three source datasets listed below. This deposit is a **derivative dataset**: the authors' contribution consists of source-label harmonization, annotation conversion to YOLO format, filename disambiguation, organization of the source-defined partitions, integrity assessment, and training-only minority-class augmentation. No new rights are claimed over the underlying images. The DOI assigned to this record identifies the harmonization and integration work and the released derivative files; it does not transfer ownership of the photographs. ## Source datasets and attribution | Source | Repository and attribution | Images | Link ||---|---|---:|---|| TEA WEED Dataset | Roboflow Universe; *teaweed* | 87 | https://universe.roboflow.com/teaweed/tea-weed-29rfc || Tea Pest Dataset | Roboflow Universe; *wenyu-wang* (hosted under *mia-kg8ss*) | 82 | https://universe.roboflow.com/mia-kg8ss/tea-pest-ecn4j || Tea-Leaf Disease and Pest Detection Dataset (v1.0) | Figshare; *S. C. Xiang* | 9,591 | https://doi.org/10.6084/m9.figshare.32253357 | All three source datasets are reported in the associated study as licensed under **Creative Commons Attribution 4.0 International (CC BY 4.0)**. Users of this deposit must preserve the attribution above and comply with the upstream licenses when redistributing the images or derivatives. ## Class mapping | Class ID | Class | Meaning ||---:|---|---|| 0 | Weed | Weed objects in tea-growing imagery || 1 | Disease | Visible disease symptoms or affected leaf regions labeled as disease targets || 2 | Pest | Pest organisms annotated as objects | Labels for Red Spider- and Tea Mosquito bug-infested leaves were assigned to Disease because their bounding boxes identify affected leaf regions rather than individual organisms. **Disease is an object-detection category for visible affected regions, not a confirmed pathogen-level diagnosis.** ## Configurations and partition statistics | Configuration | Train images | Val images | Test images | Total images | Total instances ||---|---:|---:|---:|---:|---:|| 01_Original_Harmonized | 7,392 | 808 | 1,560 | 9,760 | 12,849 || 02_Moderately_Augmented | 7,607 | 808 | 1,560 | 9,975 | 13,585 | The augmented configuration retains every original training image and adds 215 generated training images containing 736 additional annotations (370 Weed, 366 Pest). Disease instances are unchanged. The two configurations share the same underlying source photographs and **must not be added together** when reporting the number of independent images. ## Intended use and known limitations The benchmark is intended for controlled comparisons of detection architectures, class-imbalance mitigation, knowledge distillation, and resource-constrained agricultural perception. It is **not** a field validation of an autonomous UAV–UGV system. Users should note three properties of the benchmark: - **Class imbalance:** Disease accounts for approximately 96.4% of original instances; the fixed test partition contains only 18 Weed and 40 Pest instances. Class-wise metrics should accompany aggregate scores.- **Source–class association:** Each class derives predominantly from one source dataset, so class-wise differences may reflect source-specific imaging conditions as well as detection difficulty.- **Near-duplicate caution:** Integrity checks found no byte-identical images across partitions, but perceptual-hash screening flagged cross-partition similarity candidates, and the Disease source contains visually related image families that cannot be guaranteed to lie within a single partition. Results should be read as performance on these fixed partitions, not as independent-field generalization estimates. ## License This deposit is released under **Creative Commons Attribution 4.0 International (CC BY 4.0)**, consistent with the source datasets. The original images remain © their respective creators. This harmonized derivative compilation is © Sikudhan Lucas Mpuhus and Zhao Jin, released under the same license with the attribution above. ## Citation Cite this record using the Zenodo-generated citation (DOI 10.5281/zenodo.23911768). When reusing the images, also acknowledge the three original data providers listed in the attribution table above. ## Contents of this deposit - `README.md` — full dataset documentation (sections 1–13)- `augmentation_protocol.md` — training-only augmentation procedure and reproducibility limits- `01_Original_Harmonized/` — images, YOLO labels, and `data.yaml` for the original harmonized configuration- `02_Moderately_Augmented/` — images, YOLO labels, and `data.yaml` for the moderately augmented configuration