CRB Damage Monitor — Sulav Paudel, Ph.D.

Check coconut palm photos for CRB damage. Add up to 10 items, then press Start check:

Photos from your computer or phone (JPEG, PNG), or pasted with Ctrl+V
PDFs (reports, papers, slides exported as PDF; up to 15 MB): the photos inside are checked one by one
Links: an iNaturalist observation (all its photos), a GBIF record, an image, a web page (its main photos), or a public GitHub folder (all photos and PDFs in it)

Each photo gets one of three answers: Likely CRB damage Possible damage — please check No clear damage. The model judges the whole photo, so it works best on clear views of palm crowns. Photos are deleted as soon as the check is done.

One photo is not enough

A single "likely" can be a look-alike (wind, other insects, disease), and a single "no clear damage" can miss light damage. Send photos of different palms from the same area, and look at the share that come out Likely CRB damage. A quarter or more is a strong sign of CRB damage in that area.

Most palms look damaged: check 20–30 photos. If a quarter or more (5 of 20, 8 of 30) are "likely", CRB damage is very probable.
About half the palms look damaged: check 40–60 photos, with the same "quarter or more" rule.
Only a few palms damaged (new or light infestation): photos cannot confirm or rule out CRB, however many you check. Treat any "likely" result as a reason to inspect the site.

Based on our tests of the model on places it had never seen (Oct 2026); numbers will be updated as the model improves. Photos are screening only: CRB is confirmed by finding beetles, larvae or breeding sites.

Drop photos or PDFs here or

JPEG, PNG or PDF. You can also paste a copied image (Ctrl+V). For a PowerPoint deck, use File → Export → PDF first.

To check

E-mail me the results (optional)

Your e-mail address is used only to send these results and is deleted when the check finishes.

How a photo is checked

01 · Collect the photos Uploaded photos are used as they are. From a PDF, the photos embedded in it are taken out (logos and small pictures are skipped). An iNaturalist observation or GBIF record gives all of its photos; a web page gives its main photos.
02 · Palm check A fast image–text model (CLIP) first checks that the photo shows a coconut palm. Photos of people, other plants, insects or documents are reported as No coconut palm found and are not scored, because the damage model was trained only on palm photos.
03 · Damage model Each photo is scored by an image model (DINOv2-base, fine-tuned) trained on 1,479 labelled photos of CRB-damaged and undamaged coconut palms. Three copies of the model, trained separately, are averaged, and each photo is also scored as its mirror image.
04 · Three answers Likely CRB damage (score 0.88 or more), Possible damage — please check (0.44–0.88) and No clear damage (below 0.44). In tests on places the model had never seen, 5% of undamaged and 60% of damaged photos scored "likely", and 80% of undamaged and 16% of damaged photos scored "no clear damage".
05 · Human review Results are screening only. Frond damage from wind, other insects, disease or pruning can look similar, and light damage or distant palms can be missed. Found a photo worth a closer look? Email Sulav.

Training data

The model was trained on 1,479 photos (740 with CRB damage, 739 without), each labelled by eye, from field surveys and published field photos across the CRB range: Vanuatu (Efate and Santo surveys, about 570 photos), Fiji, Papua New Guinea, Solomon Islands, Samoa, Cook Islands, Tahiti, Guam, Hawaiʻi, Timor, Australia (Cairns) and India (Kerala). Photos the model got wrong in testing were re-checked by eye without seeing their original label; labels were corrected where needed, and photos that could not be judged at their resolution were left out.

How it was tested

New places: the places were split into five groups. Each group was left out in turn, a model was trained on the rest and tested on the group it had never seen. Across all photos it ranked a damaged photo above an undamaged one 90% of the time (AUC 0.90).
A new survey year: 705 damaged photos from the Efate (Vanuatu) roadside survey of November 2025, never used for training: 88% came out as likely CRB damage, 8% as possible and 4% as no clear damage. Lightly damaged palms were the hardest: 59% likely, 23% missed.

Limitations

The model judges the whole photo. With several palms in view, a small damaged palm can be missed, and it cannot say which palm is damaged.
Frond damage from wind, other insects, disease or pruning can look like CRB damage. In testing, most false alarms were palms with other kinds of damage.
Light damage, distant palms, blurry or low-resolution photos, and screenshots with text are less reliable.
It was trained on Pacific, Asian and Australian palms; photos from new regions (e.g. the Americas) may be scored less accurately until local photos are added.
The answer levels are screening aids, not a diagnosis: please confirm by eye.

Next steps

Palm by palm: find each palm crown in the photo and score it separately, so results point to the damaged palm.
Damage grades: estimate how severe the damage is on the 1–5 survey scale, not just whether it is there.
V-shaped cuts: check that the model is using the characteristic V-cuts in the fronds, and train on them directly.
More places: add labelled photos from new regions, especially more undamaged palms from recent surveys.

Credits & citation

Damage model, training data and site: Sulav Paudel.

The first version of this monitor used Aubrey Moore's (University of Guam) SAM3 + elliptic Fourier V-cut detector (aubreymoore/CRB-2026-05-13), which shaped this work.

Paudel, S., & Moore, A. (2026). CRB Damage Monitor [Web application]. https://s-paudel.github.io/crb-monitor.html

Moore, A. (2026). CRB-2026-05-13 [Computer software]. GitHub. https://github.com/aubreymoore/CRB-2026-05-13

Image sources

Photos are shown directly from iNaturalist and GBIF. They are credited to their photographers under the licence shown on each photo, and are not copied or stored here. DINOv2 is by Meta AI.