🌴 Palm Oil Fruit Bunch Detection & Counting — YOLOv8

Real-time detection and automated counting of palm oil fruit bunches (Tandan Buah Segar) using YOLOv8 — an industry collaboration project with PT Perkebunan Nusantara (PTPN), developed at Binus University.

---

🎯 Project Overview

This project automates the quality grading and counting of palm oil fruit bunches (Tandan Buah Segar/TBS) using computer vision. The model detects and classifies each bunch into three categories, enabling faster and more accurate quality assessment in palm oil plantations.

Industry Partner: PT Perkebunan Nusantara (PTPN)
Institution: Binus University

---

🍇 Detection Classes

Class	Label	Color	Description
0	Matang 🟢	Dark Green	Ripe — ready for harvest
1	Mentah 🔵	Blue	Unripe — not ready
2	Cacat 🔴	Red	Defective — damaged or abnormal

---

🏗️ Pipeline

Aerial / Field Image
        │
        ▼
  Roboflow Dataset
  (palm-oil-nk2ks / cluster-palm-oil v12)
        │
        ▼
  YOLOv8 Training
  ├─ Model   : yolov8n (nano)
  ├─ Epochs  : 200
  ├─ ImgSz   : 800×800
  ├─ Batch   : Auto (-1)
  └─ Workers : 16
        │
        ▼
  Object Detection & Counting
  ├─ Custom colored bounding boxes per class
  ├─ Numbered labels per detected object
  ├─ Per-class count: Matang / Mentah / Cacat
  └─ Confidence threshold tuning (0.03 – 0.6)
        │
        ▼
  Export to ONNX
  (for deployment / edge device)
        │
        ▼
  TensorBoard Monitoring

---

✨ Key Features

• 3-class palm oil quality detection — Ripe, Unripe, and Defective bunches in a single pass
• Automated object counting — each detected bunch is labeled and numbered on the output image
• Color-coded bounding boxes — green for Matang, red for Cacat, blue for Mentah for instant visual grading
• Confidence threshold tuning — configurable from 0.03 to 0.6 for different field conditions
• ONNX export — simplified model (opset 13) for edge deployment on plantation hardware
• TensorBoard integration — real-time training monitoring
• Google Drive integration — persistent model storage across Colab sessions

---

🛠️ Tech Stack

Category	Tools
Model	YOLOv8 (Ultralytics)
Dataset Management	Roboflow API
Computer Vision	OpenCV (cv2)
Export Format	ONNX (opset 13, imgsz 416)
Training Monitor	TensorBoard
Environment	Google Colab + Google Drive
Language	Python 3.10+

---

📁 Project Structure

YOLO8/
├── ultralytics/                  # YOLOv8 cloned repository
│   ├── runs/
│   │   └── detect/
│   │       └── train/
│   │           └── weights/
│   │               └── best.pt   # Best trained model weights
│   ├── Data_17 05 2025/          # Field test images
│   │   ├── Percobaan1.jpg
│   │   ├── Percobaan3.jpg
│   │   └── Percobaan8.jpg
│   └── target_prediction/        # Prediction targets
├── cluster-palm-oil-12/          # Roboflow dataset (YOLOv8 OBB format)
├── chip.yaml                     # Dataset config
└── YOLO8_COUNT_OBJECT.ipynb      # Main notebook

---

🚀 Getting Started

1. Setup Environment

# Mount Google Drive
from google.colab import drive
drive.mount('/content/drive')
%cd /content/drive/MyDrive/YOLO8

# Install dependencies
!pip install -U roboflow
!git clone https://github.com/ultralytics/ultralytics
%pip install -qe ultralytics

2. Download Dataset via Roboflow

from roboflow import Roboflow

rf = Roboflow(api_key="YOUR_API_KEY")
project = rf.workspace("palm-oil-nk2ks").project("cluster-palm-oil")
dataset = project.version(12).download("yolov8-obb")

3. Train the Model

yolo task=detect mode=train \
     model=yolov8n.pt \
     data=/content/drive/MyDrive/YOLO8/chip.yaml \
     epochs=200 \
     workers=16 \
     batch=-1 \
     imgsz=800 \
     save=True \
     cache=ram \
     val=True \
     plots=True

4. Run Detection & Counting

from ultralytics import YOLO
import cv2

model = YOLO('runs/detect/train/weights/best.pt')
class_names = {0: 'Cacat', 1: 'Matang', 2: 'Mentah'}

results = model(source='your_image.jpg', conf=0.5)

# Count per class
Total, Matang, Mentah, Cacat = 0, 0, 0, 0
for result in results:
    for class_index in result.boxes.cls.tolist():
        Total += 1
        if class_index == 1: Matang += 1
        elif class_index == 2: Mentah += 1
        elif class_index == 0: Cacat += 1

print(f"Total: {Total} | Matang: {Matang} | Mentah: {Mentah} | Cacat: {Cacat}")

5. Export to ONNX

yolo task=detect mode=export \
     model=runs/detect/train/weights/best.pt \
     format=onnx simplify=True opset=13 imgsz=416

---

⚙️ Training Configuration

Parameter	Value	Notes
Base Model	yolov8n.pt	Nano — fast inference
Epochs	200	Transfer learning
Image Size	800×800	Optimal for cluster detail
Batch Size	-1 (auto)	Adapts to GPU VRAM
Workers	16	Faster data loading
Cache	RAM	Faster epoch cycling
Confidence	0.03 – 0.6	Tunable per condition

---

🌐 Deployment

The model is exported to ONNX format (opset 13, 416×416) for potential deployment on:

• Edge devices at plantation sites
• Mobile inspection tools
• Integrated quality grading systems

---

🤝 Collaboration

This project is a university-industry collaboration between:

University	Binus University
Industry Partner	PT Perkebunan Nusantara (PTPN)
Use Case	Automated TBS quality grading at harvest points

---

📄 License

This project was developed for academic and industrial research purposes in collaboration with PTPN. Dataset and model weights are proprietary.