UAV RGB Image Dataset for Object Detection of Ganoderma-Affected Oil Palm Trees

Description

This dataset provides UAV RGB image tiles and crown-level object annotations for the detection of Ganoderma-affected oil palm trees. Version 2 is a methodologically reconstructed release designed to support geographically independent object-detection benchmarking. The canonical release contains 1,352 RGB image tiles (640 × 640 pixels) derived from an August 2023 UAV orthomosaic at an approximate ground sampling distance of 0.0408 m/pixel. The dataset contains 9,400 accepted oil-palm crown annotations in COCO format, comprising 722 field-derived Ganoderma-positive annotations and 8,678 survey-negative-normal annotations. Disease status was derived from field survey records linked to palm identities; RGB-image reviewers were used to validate crown geometry rather than visually diagnose Ganoderma infection. To reduce spatial leakage, the train, validation, and test partitions were defined using geographically disjoint plantation blocks before image extraction. The final release contains 1,128 training tiles, 106 validation tiles, and 118 test tiles, with no block or anchor-palm overlap between partitions. Cross-split file-integrity checks also identified no identical image files across partitions. Bounding boxes were produced using a human-calibrated semi-automatic crown annotation workflow followed by risk-based human quality control. Of 9,439 candidate object instances, 9,400 were accepted and 39 were excluded during quality control. No image augmentation is included in the canonical public release; augmentation, when used for model development, should be applied only to the training partition after the provided geographic split has been preserved. The term “survey-negative-normal” indicates a counted palm without a matched Ganoderma-positive field record under the dataset's survey-linkage protocol and should not be interpreted as laboratory-confirmed absence of Ganoderma infection. Independently field-verified annotations for other foliar stresses are not included. The release represents one plantation environment and one selected acquisition month; these limitations should be considered when evaluating model generalization. The repository includes RGB image tiles, COCO annotations for the complete dataset and each predefined split, public object- and tile-level metadata, a dataset card, a data dictionary, and SHA-256 checksums for release-integrity verification. Operational block identifiers, persistent internal palm identifiers, and exact UTM coordinates have been excluded from the public release.

Steps to reproduce

The dataset is provided as a canonical, geographically partitioned object-detection benchmark. Users should preserve the predefined train, validation, and test partitions supplied with the release. These partitions were defined using geographically disjoint plantation blocks before RGB tile extraction to reduce spatial leakage. RGB images are provided as 640 × 640 pixel tiles. Object annotations are supplied in COCO format for the complete dataset and separately for the predefined training, validation, and test partitions. The two annotation categories are ganoderma_affected_palm and survey_negative_normal_palm. Ganoderma-positive status originates from field survey records linked to palm identities; RGB-image review was used to validate crown geometry rather than to visually diagnose disease status. For model development, use instances_train.json with the training images. Hyperparameter selection and model selection should use only the validation partition. The test partition should remain untouched until final evaluation. If image augmentation is used, it should be applied only to the training partition after the predefined geographic split has been preserved. No augmented images are included in the canonical release. For reproducibility and integrity checking, users should verify repository files against SHA256SUMS.csv and consult README.md, DATASET_CARD.md, DATA_DICTIONARY.md, and metadata/dataset_summary.json before analysis. The supplied public tile and object metadata describe the released benchmark without exposing operational block identifiers, internal palm identifiers, or exact UTM coordinates. Recommended object-detection reporting includes precision, recall, F1-score, AP50, and AP50–95 on the predefined test partition. Results should explicitly state that the benchmark represents one plantation environment and one selected UAV acquisition month. The survey_negative_normal_palm category should not be interpreted as laboratory-confirmed absence of Ganoderma infection.

Institutions

• Bina Nusantara University

  DKI Jakarta, West Jakarta

Categories

Licence