Oil Palm Counter

Count oil palms in a drone image, right in your browser. Your image is never uploaded: the model runs on your own computer.

For good results

A straight-down (nadir) drone RGB image, ideally about 2 cm per pixel (a typical drone flight at 60–80 m). JPG or PNG.
Works best on young to mid-age plantations like the one it was trained on. On close-up photos of mature palms with overlapping crowns it finds only about half of them.
The first count downloads the model (45 MB); after that it loads from your browser.

Drop a drone image here, or click to choose one No image chosen

Confidence threshold: 0.25 0.25 is the setting used for the production counts. Move it after a count to see the result change.

Image resolution, cm per pixel (optional) If you know it, the image is resized to the 2.13 cm/px the model was trained on, and the area and palms per hectare are shown.

About the model

A YOLOv8s detector trained on 4,041 hand-labelled palm crowns from drone orthomosaics of a commercial plantation in West Africa, and used there to count 718,751 palms over 12,166.6 ha. On 79 held-out test tiles:

0.845precision: share of detections that are real palms

0.774recall: share of real palms found

+0.2%palm count vs. hand count, tile interiors (519 vs 518)

Large images are cut into overlapping 1280 px tiles exactly as the full pipeline does, and palms seen by two tiles are counted once. Full evaluation and limitations: model card.

Whole orthomosaics

This page works on ordinary images, so positions are in pixels. For GeoTIFF orthomosaics with map coordinates and a GIS-ready palm inventory, use the free Python pipeline or the model in QGIS (how).

Collaborate

Next for this work: validating on more plantations and regions, handling mature overlapping canopy, an open benchmark of labelled palm crowns, and missing-palm maps from repeat flights (roadmap).

I welcome collaboration with research institutes, universities, NGOs, government agencies, companies and startups. The most useful contributions are imagery and field counts, GPU compute or cloud credits, and joint research and publications. In return I can offer palm inventories, a model tuned to your imagery, pilot projects and training. Contact o.oyin1115@gmail.com.