SambaGuard Pest & Disease Detector
Overview
SambaGuard Pest & Disease Detector is a YOLOv8s-based object detection model trained to detect Fall Armyworm (FAW) pest infestations and Maize Streak Disease in smallholder maize farms. It is part of the SambaGuard AI system, developed at Dedan Kimathi University of Technology, Kenya, targeting edge deployment on resource-constrained devices such as the Raspberry Pi 5.
This model is the strongest performing version across all SambaGuard training experiments, achieving mAP50 of 0.459 across 5 classes — surpassing all previous versions including the FAW-only v2 model (mAP50 = 0.347).
This repository contains the trained model weights, evaluation metrics, training visualizations, and full experiment outputs.
Detected Classes
| Class ID | Class Name |
|---|---|
| 0 | Fall Armyworm Egg |
| 1 | Fall Armyworm Frass |
| 2 | Fall Armyworm Larva |
| 3 | Fall Armyworm Larval Damage |
| 4 | Maize Streak Disease |
Model Architecture
| Property | Value |
|---|---|
| Architecture | YOLOv8s |
| Framework | Ultralytics / PyTorch |
| Input size | 640 x 640 pixels |
| Parameters | 11.1M |
| GFLOPs | 28.4 |
| Export formats | ONNX, TFLite |
| Number of classes | 5 |
Dataset
The model was trained on a subset of the KaraAgro AI Maize dataset, accessed via Dataset Ninja in Supervisely format. The original dataset was collected by KaraAgro AI from maize farms across Ghana and is available on Harvard Dataverse (DOI: 10.7910/DVN/CXUMDS, License: CC0 1.0).
Dataset preparation steps:
- Retained only the 5 most meaningful classes: FAW egg, frass, larva, larval damage, and maize streak disease
- Dropped healthy maize, healthy images, and none-healthy classes
- No manual augmentation applied — YOLOv8 built-in augmentation only
- No oversampling applied
- Annotations converted from Supervisely JSON format to YOLOv8 normalized bounding box format
- Out-of-bounds coordinates clipped to the valid range of 0.0 to 1.0
| Split | Images | Labels |
|---|---|---|
| Train | 5,084 | 5,084 |
| Val | 1,715 | 1,715 |
| Test | 664 | 664 |
Training Configuration
| Parameter | Value |
|---|---|
| Base model | yolov8s.pt (COCO pretrained) |
| Epochs | 50 |
| Batch size | 16 |
| Image size | 640 |
| Seed | 42 |
| Workers | 2 |
| Augmentation | YOLOv8 built-in only |
| Oversampling | None |
| Hardware | NVIDIA Tesla T4 GPU |
Validation Results
Best checkpoint obtained at epoch 49.
Overall
| Metric | Value |
|---|---|
| Precision | 0.500 |
| Recall | 0.466 |
| mAP50 | 0.459 |
| mAP50-95 | 0.220 |
Per-Class Performance
| Class | Precision | Recall | mAP50 | mAP50-95 |
|---|---|---|---|---|
| Fall Armyworm Egg | — | 0.135 | 0.090 | 0.030 |
| Fall Armyworm Frass | 0.404 | 0.249 | 0.228 | 0.074 |
| Fall Armyworm Larva | 0.796 | 0.875 | 0.893 | 0.352 |
| Fall Armyworm Larval Damage | 0.464 | 0.328 | 0.316 | 0.110 |
| Maize Streak Disease | 0.645 | 0.743 | 0.767 | 0.535 |
Fall Armyworm Larva achieves the strongest detection performance (mAP50 = 0.893), and Maize Streak Disease is detected with high reliability (mAP50 = 0.767) due to its visually distinct symptoms. Egg detection remains the weakest class due to the small size and low image count of the class in the training data.
Comparison with Previous Versions
| Version | Dataset | Classes | Epochs | mAP50 | mAP50-95 |
|---|---|---|---|---|---|
| v1 | KaraAgro FAW-only | 4 | 50 | 0.368 | — |
| v2 | KaraAgro FAW-only (cleaned) | 4 | 100 (best 55) | 0.347 | 0.144 |
| v3 | Roboflow 6-class | 6 | 100 (best 49) | 0.161 | 0.122 |
| v4 | Roboflow 6-class + oversampling | 6 | 100 (best 56) | 0.166 | 0.123 |
| This model | KaraAgro 5-class | 5 | 50 | 0.459 | 0.220 |
Key finding: returning to the original KaraAgro dataset with minimal preprocessing — no augmentation, no oversampling, just class filtering — produced the best results across all experiments.
Repository Structure
disease-detector/
├── weights/
│ ├── best.pt
│ ├── last.pt
│ ├── best.onnx
│ └── best.tflite
├── metrics/
│ └── results.csv
├── plots/
│ ├── results.png
│ ├── confusion_matrix.png
│ ├── confusion_matrix_normalized.png
│ ├── PR_curve.png
│ ├── P_curve.png
│ ├── R_curve.png
│ └── F1_curve.png
├── args.yaml
└── README.md
Usage
Install the required library:
pip install ultralytics
Run inference on an image:
from ultralytics import YOLO
model = YOLO("weights/best.pt")
results = model.predict(
source="image.jpg",
imgsz=640,
conf=0.25
)
results[0].show()
Known Limitations
- Egg detection performance is low (mAP50 = 0.090) due to the small visual size of eggs and limited training examples for this class
- Frass detection is moderate — frass can resemble soil particles and other debris in field conditions
- Model has not yet been validated in real field deployment conditions
Future Work
- Edge export optimization for Raspberry Pi 5 deployment
- Integration with Swahili-language LLM advisory layer for farmer SMS guidance
- Field validation with smallholder farmers in Kenya
- Larger annotated dataset for egg and frass classes
Related Resources
- GitHub Repository: SambaGuard
- FAW-only model (v2): ndunge23/SambaGuard-v2
- Dataset: KaraAgro AI Maize on Dataset Ninja
- Original dataset: KaraAgro AI Maize on Harvard Dataverse
Citation
@misc{sambaguard2026,
author = {Annastacia Ndunge},
title = {SambaGuard Pest and Disease Detector: YOLOv8s for Fall Armyworm and Maize Streak Disease Detection},
year = {2026},
howpublished = {\url{https://huggingface.co/ndunge23/disease-detector}},
note = {Part of the SambaGuard AI project}
}
License
This project is licensed under the Apache 2.0 License.
Author
Annastacia Ndunge Electrical and Electronic Engineering, Dedan Kimathi University of Technology, Kenya
GitHub: aneneahs-kanaks Hugging Face: ndunge23 LinkedIn: Annastacia Ndunge
- Downloads last month
- 46