SambaGuard AI is a YOLOv8-based object detection model for the early detection of Fall Armyworm (FAW) infestations in smallholder maize fields. The model was developed as part of a research project at Dedan Kimathi University of Technology, Kenya, targeting edge deployment on resource-constrained devices such as the Raspberry Pi 4.
The system is designed to address a documented gap in FAW management: smallholder farmers cannot scout frequently enough to catch FAW in its narrow early intervention window, and existing tools do not provide severity-graded, localized guidance. SambaGuard AI aims to close this gap through automated, camera-based detection at the field level.
This repository contains the trained model weights, evaluation metrics, training visualizations, and full experiment outputs for version 2.
Detected Classes
The model detects four Fall Armyworm-related classes:
Class ID
Class Name
0
Fall Armyworm Egg
1
Fall Armyworm Frass
2
Fall Armyworm Larva
3
Fall Armyworm Larval Damage
Model Architecture
Base architecture: YOLOv8s (Ultralytics)
Framework: PyTorch
Input resolution: 640 x 640 pixels
Parameters: 11.1M
GFLOPs: 28.4
Dataset
The model was trained on a cleaned and validated subset of the KaraAgro AI Maize dataset, accessed via Dataset Ninja (https://datasetninja.com/kara-agro-ai-maize) in Supervisely format. The original dataset was published by KaraAgro AI and is available on Harvard Dataverse (DOI: 10.7910/DVN/CXUMDS, License: CC0 1.0).
Dataset preparation steps for v2:
Removal of non-FAW classes (healthy maize, maize streak disease) with 10% retained as background negative samples
Detection and removal of duplicate file extensions
Repair of invalid YOLO annotations with out-of-bounds bounding box coordinates
Per-image bounding box coordinate clipping to the valid range of 0.0 to 1.0
Image-label pair verification across all splits
Stratified train/val/test split to preserve class proportions
Oversampling of the egg class in the training split only to address class imbalance
Final dataset verification confirmed no missing images, no missing labels, no invalid annotations, and no unmatched image-label pairs across all splits.
Split
Images
Labels
Train
5,709
5,709
Val
1,320
1,320
Test
664
664
Training Configuration
Parameter
Value
Base model
yolov8s.pt (COCO pretrained)
Epochs
100
Early stopping patience
20
Batch size
16
Image size
640
Optimizer
Auto (AdamW)
Learning rate schedule
Cosine annealing
Mixed precision (AMP)
Enabled
Mosaic augmentation
Enabled
HSV augmentation
Enabled
Horizontal flip
Enabled
Hardware
NVIDIA Tesla T4 GPU
Validation Results
Best checkpoint obtained at epoch 55.
Overall performance:
Metric
Value
Precision
0.479
Recall
0.376
mAP50
0.347
mAP50-95
0.137
Per-class performance:
Class
Precision
Recall
mAP50
mAP50-95
Fall Armyworm Egg
0.336
0.250
0.198
0.085
Fall Armyworm Frass
0.374
0.233
0.196
0.066
Fall Armyworm Larva
0.782
0.716
0.726
0.303
Fall Armyworm Larval Damage
0.425
0.304
0.267
0.093
The larva class achieves the strongest detection performance (mAP50 = 0.726), which is consistent with its larger visual signature and stronger representation in the training data. Egg and frass detection remain areas for improvement in subsequent versions.
Comparison with v1 Baseline
Metric
v1 Baseline
v2 Clean Dataset
Change
Precision
0.486
0.479
-0.007
Recall
0.401
0.376
-0.025
mAP50
0.368
0.347
-0.021
mAP50-95
0.144
0.137
-0.007
Larva mAP50
0.768
0.726
-0.042
Epochs trained
50
100 (best at 55)
—
Overall metrics show a slight decrease from v1. This is being investigated and attributed to differences in the effective training set composition after dataset cleaning — specifically the removal of out-of-bounds annotations that were previously counted as valid training signal. Further analysis is ongoing. See experiment notes below.
This version introduced substantial dataset cleaning improvements over v1, including annotation repair, duplicate removal, bounding box validation, and egg class oversampling. Despite these improvements, overall mAP50 is marginally lower than v1 (0.347 vs 0.368).
This result is consistent with the hypothesis that dataset cleaning removed a number of noisy annotations that, while technically invalid, previously provided approximate training signal. The v2 dataset is cleaner and more reliable, but smaller in effective annotated objects. Subsequent experiments (v3, v4) will isolate the effect of augmentation and image resolution changes to determine the best path forward.
Planned Experiments
Version
Key Change
Status
v1
Baseline — 50 epochs, uncleaned dataset
Complete
v2
Clean dataset — 100 epochs
Complete (this model)
v3
Stronger augmentation for egg and frass classes
Planned
v4
Image size increased to 960
Planned
Future Work
Improving egg and frass detection through targeted augmentation
Edge export to TFLite and ONNX for Raspberry Pi 4 deployment
Real-time inference benchmarking on edge hardware
Integration with a Swahili-language LLM advisory layer for farmer-facing SMS guidance
Field validation with smallholder farmers in Kenya
