DECAFIA: A Field-Collected Coffee Leaf Disease Detection Dataset from Santander, Colombia (CoffeeLeaf-CO)

DECAFIA is a multi-source annotated dataset of 2,786 coffee leaf images comprising 15,181 annotated instances across four classes, designed for object detection of coffee leaf diseases and pest damage under Colombian Andean agronomic conditions.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━CLASSES━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ID 0 — roya      | 5,031 instances | Hemileia vastatrix (coffee rust)ID 1 — coco      | 5,394 instances | Curculionidae weevil defoliationID 2 — hojas     | 2,371 instances | Whole-leaf outline (geometric reference)ID 3 — minador   | 2,385 instances | Leucoptera coffeella (leaf miner)Total            | 15,181 instances across 2,786 images

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━DATASET SPLIT━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━Train : 1,934 images (69.4%)Val   :   549 images (19.7%)Test  :   286 images (10.3%)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━IMAGE SOURCES — FULL DECLARATION━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Source 1 — Original field collection (Colombia)Images with prefixes: COCO_*, ROYA_*, SANAS_SOCORRO_*, SANAS_NUEVAS_*Location: Coffea arabica plantation, El Socorro, Santander, Colombia(6°28′N, 73°19′W; ~1,500 m a.s.l.)Captured: 2022–2023, consumer Android smartphones, natural field lighting (no controlled setup). All annotations are original to this work.

Source 2 — RoCoLe (Ecuador) — re-annotated under CC BY 4.0Images with prefix: SANAS_ROCOLE_*Original dataset: Parraga-Alava, J. et al. "RoCoLe: A robusta coffee leaf images dataset." Data in Brief, 2019.DOI: https://doi.org/10.17632/c5yvn32dzgThese healthy leaf images were redistributed and independently re-annotated by the DECAFIA authors under CC BY 4.0 terms. Original RoCoLe disease labels were not used.

Source 3 — Silva et al. 2020 (Brazil) — re-annotated under CC BY 4.0Images with prefix: MINEIRO_*Original dataset: Silva, L.B. et al. "Rust and leaf miner in coffee crop (Coffea arabica)." Mendeley Data, 2020.DOI: https://doi.org/10.17632/vfxf4trtcg.5These leaf miner images were redistributed and independently re-annotated by the DECAFIA authors under CC BY 4.0 terms.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ANNOTATION━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━Tool: VGG Image Annotator (VIA) v2.0 + Segment Anything Model (SAM)Format: YOLO bounding box (.txt per image, normalized coordinates)Inter-annotator agreement: Cohen's κ = 0.91 (near-perfect)Annotators: 2 independent + 1 adjudicator for disagreements

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━FILE FORMAT━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━Each image has a corresponding .txt label file (same filename).One annotation per line: class_id x_center y_center width heightAll coordinates normalized to [0, 1].Images without disease annotations are healthy leaves (hojas only).

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━BASELINE BENCHMARK━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━Architecture: YOLOv8m (fine-tuned from MS COCO pre-training)Test set (286 images): mAP50 = 92.4% | mAP50-95 = 76.6%Per-class AP50: roya 88.2% | coco 94.1% | minador 94.9%Model weights: https://huggingface.co/estebanr25/decafiaSource code: https://github.com/estebanr25/decafia

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━AUTHORS━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━Luis Esteban Rosas RuizAndrey Fernando Salom MedinaDirector: Jaime Guillermo Barrero PerezUniversidad Industrial de Santander, Bucaramanga, Colombia

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━LICENSE━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━Creative Commons Attribution 4.0 International (CC BY 4.0)Images from Source 2 and Source 3 are redistributed under their original CC BY 4.0 licenses with proper attribution.