Project aim

Count cocoa plants automatically from drone imagery, across both regular commercial plantings and irregular smallholder farms, and say how confident each count is.

  • YOLOv8Detector
  • 2Layout types covered
  • EnsembleUncertainty scoring
  • Held-outTest set
Illustrative map for Cocoa Plant Detection & Counting Illustrative graphic

Overview

YOLOv8 detects individual cocoa plants in UAV imagery. It is combined with multi-temporal analysis, and ensemble uncertainty scoring flags detections the model is unsure about. The approach is evaluated on a held-out test set covering smallholder and commercial layouts.

The challenge

Commercial farms are planted in rows, while smallholder plots are irregular and mixed with shade trees, so a detector tuned to one fails on the other. Counts are also used for planning, so uncertain detections need to be visible rather than hidden.

Data & inputs

  • UAV orthomosaics
  • Annotated plant locations

Approach

  1. 01

    UAV imagery

    Drone flights capture high-resolution imagery of the farms.

  2. 02

    Detection

    YOLOv8 detects individual cocoa plants.

  3. 03

    Multi-temporal analysis

    Detection is combined with multi-temporal analysis of the imagery.

  4. 04

    Ensemble uncertainty

    An ensemble scores each detection's uncertainty, so low-confidence counts are flagged.

Results

The detector counts cocoa plants across smallholder and commercial layouts and reports uncertainty with each count.

Validation

Evaluated on a held-out test set.