Tapziy/oil-palm-detection-data

SawitGuard-GNN — Dataset Card

Three sub-datasets used by the SawitGuard-GNN oil palm risk-ranking pipeline (Datathon 2026, RISTEK Fasilkom UI). Code and full documentation: SawitGuard-GNN. Companion weights: oil-palm-detection-weights.

⚠ Read "Known data-quality caveat" below before using layer1_uav_crowns for anything. It is not three independent datasets — it is one dataset, tile-duplicated ~30×.

---

Summary

Sub-dataset	Modality	Unit	Label	License
layer1_uav_crowns	UAV RGB, nadir ortomosaic tiles	5,077 unique tree crowns	Healthy / Unhealthy (generic canopy health)	CC BY 4.0
layer2_eg9pp_panel	Tabular field census, 25 years	1,200 palms × 45 censuses	Field-verified Ganoderma/BSR symptom + death events	CC BY-SA 4.0
peru_palm_anomaly	UAV RGB, nadir	424 images (86 validation)	PalmAnom / PalmSan crown anomaly classes	CC BY 4.0

Only layer2_eg9pp_panel is field-verified for Basal Stem Rot (BSR) / Ganoderma boninense. The two UAV datasets use generic Roboflow-crowdsourced crown-health labels — not a BSR diagnosis. This distinction is load-bearing throughout the project; do not blur it.

---

layer1_uav_crowns — UAV crown inventory (Layer 1)

• Files: ds_B/ (raw, 2,303 JPG tiles + 3 COCO _annotations.coco.json), frozen/layer1_crowns.csv (5,077 rows, one per unique tree), frozen/layer1_tiles_disjoint.csv (43 rows, the non-overlapping tile subset)
• Source: Roboflow health-detection/oil-palm-health-detection, export v2, 2024-04-21. License: CC BY 4.0, "Provided by a Roboflow user."
• Modality: UAV RGB nadir, ortomosaic, GSD ≈ 8.7 cm/px
• Label: Healthy / Unhealthy — generic canopy health, not BSR, no field verification
• Unit of analysis: 5,077 unique trees (deduplicated from 151,060 raw annotation boxes — see caveat below)
• Positive class: 66 unique Unhealthy trees (1.30%) — 17/31/18 per orthomosaic
• Spatial units: 3 orthomosaics — 1,379 / 1,849 / 1,849 trees
• Split: split_fold = ortho → leave-one-orthomosaic-out, 3 folds. Random splits are invalid on this dataset (see caveat).
• Forced limit: n=66 unique positives ⇒ wide confidence bands; class-weighting is indistinguishable from noise at this n. Maximum honest claim = single-site generic canopy-health demonstrator, not a BSR detector.

⚠ Known data-quality caveat (read before using)

Roboflow tiles were exported at random offsets, not on a grid, so tiles overlap heavily and one physical tree appears in a median of 32 overlapping tiles (range 1–77). Each orthomosaic is only ~5,000×5,000px (fits ~25 non-overlapping 1,024px tiles) but the export contains 737–799 tiles per site — only 13/14/16 tiles per orthomosaic are actually non-overlapping (layer1_tiles_disjoint.csv). Training or evaluating on the raw 151,060 annotation rows means training on 5,077 physical trees replicated 29.8× — this is documented in full, with verification checks (0 label conflicts between duplicate boxes, 0 "phantom neighbors" under 0.5× planting distance, 5,048/5,077 canonical views ≥60px from any tile edge) in audit/AUDIT_REPORT.md in this repo. frozen/layer1_crowns.csv is the already-deduplicated, ready-to-use table — one row per unique tree, using its single most-central ("canonical") tile view. Never treat 151,060 as a sample size.

---

layer2_eg9pp_panel — 25-year field epidemic panel (Layer 2)

• Files: Eg9PP_Phenotypes.csv (raw source), frozen/layer2_nodes.csv (1,200 rows), frozen/layer2_panel.csv (54,000 rows, tree × census long format), frozen/layer2_edges.csv (3,354 rows)
• Source: Tisné, S. et al. (2017), G3: Genes, Genomes, Genetics 7(6):1683–1692, doi:10.1534/g3.117.041764 — SOCFINDO estate, Medan, Indonesia. License: CC BY-SA 4.0, copyright PalmElit & CIRAD. See Eg9PP_LICENSE.md in this folder for full terms; attribution to the source paper is required in any derived work.
• Label: field-verified Ganoderma/BSR — first-symptom date and death date per palm, with censoring indicators. This is the only field-verified BSR label in the whole project.
• Unit of analysis: 1,200 palms, 14 families, 80 plots, 2 parcels
• Time range: 45 census dates, years 0.5–25.5 (54,000-row panel)
• Events: symptomatic 702 (58.5%), dead 366 (30.5%); 498 palms never symptomatic during observation; earliest censored observation at t=6.0y
• Contact graph: root-contact edges at r = 1.5× planting distance → 3,354 edges, mean degree 5.59
• Split: fold = parcel → leave-one-parcel-out, 2 folds. Verified safe: 0/3,354 edges cross a parcel boundary, and all 14 families appear in both parcels (fold split is not confounded with genotype).
• Forecasting task: predict whether an asymptomatic (A) palm becomes symptomatic/dead within h census steps. Positive rates: h=1 1.58%, h=2 3.03%, h=3 4.45%, h=4 5.65%.
• Forced limit: no imagery at all, and the latent "exposed" (E) disease compartment is never observed — only first-symptom and death times. A full SEIR head is not identifiable from this data; the project uses a reduced SI(D) head instead, and results from the two are not comparable (see model card).
• Geometry correction: raw X_POSITION/Y_POSITION are not to scale. xm = X × cos(30°) makes the six nearest neighbors land at exactly distance 1.000 (equilateral triangular planting). Without this correction the contact graph is wrong. frozen/layer2_nodes.csv already has this applied.
• Censoring: palms that leave observation get status C and are excluded from the risk set from that point on — never treated as healthy through the end of the study.

---

peru_palm_anomaly — Third, independent evidence line

A deliberately separate UAV dataset from a different estate and country, used only to test whether the same crown-detection approach transfers across sites — never merged with layer1_uav_crowns.

• Files: images/{train,valid,test}/ (424 JPG, 800×600, + _annotations.csv per split, TF Object Detection format), images/README.roboflow.txt, images/README.dataset.txt
• Source: Roboflow proyecto-palmera-aceitera/oil-palm-tree-detection-4, export v15, 2024-06-12, underlying data from Mendeley Data doi:10.17632/nh7d23dgnw.1. License: CC BY 4.0.
• Modality: UAV RGB, oil palm plantation, Peru
• Label: two crown-anomaly classes, PalmAnom and PalmSan — again not BSR
• Unit: 424 images; 86 validation images, 109 ground-truth boxes
• Forced limit: only 1 fold / 1 seed was run (not 3-fold like layer1_uav_crowns) ⇒ no mean±std, not directly comparable to the ds_B numbers. Maximum honest claim = qualitative support that crown detection transfers cross-site. Also: the detector over-predicts by 36% (1.72 boxes/image vs. 1.267 ground-truth; 28/86 images exceed GT count), so a high mAP50 does not mean tree counting is accurate — which matters because counting is exactly what graph construction needs.

---

What is NOT joined, and why

layer1_uav_crowns and layer2_eg9pp_panel are never merged: different estates, different eras (the Eg9PP plots were removed in 2012; the UAV imagery is post-2013), no join key, and neither source is georeferenced. What is tested instead is interface compatibility — do the two layers' output graphs have the same shape? Mean degree at r=1.5× planting distance, inner trees only: Eg9PP (planting positions) 5.74, Layer 1 predicted (YOLOv12n) 5.54±0.12, Layer 1 ground-truth boxes 5.62±0.05. Predicted vs. Eg9PP differ by 3.5% — both are degree-6 triangular lattices, consistent in shape though not identical. Full derivation in audit/AUDIT_REPORT.md and the code repository's docs/RESULTS.md.

Citation

If you use layer2_eg9pp_panel, please cite:

Tisné, S., Pomiès, V., Riou, V., Syaputra, I., Sudarsono, Cros, D., Yangera, A., Nodichao, L., Cochard, B., & Denis, M. (2017). Identification of Ganoderma disease resistance loci using natural field infection of an oil palm multiparental population. G3: Genes, Genomes, Genetics, 7(6), 1683–1692. https://doi.org/10.1534/g3.117.041764

For the UAV sub-datasets, please attribute the respective Roboflow sources linked above (CC BY 4.0).