Emmanuel-Acquah/g-yolov8-msv-ghana

# G-YOLOv8: Ghost-Engineered YOLOv8 for Real-Time Maize Streak Virus Detection in Ghanaian Smallholder Farms ## Project Summary This repository accompanies the research proposal *"A Lightweight YOLOv8-GhostNet Architecture for Real-Time Maize Streak Virus Detection in Ghanaian Smallholder Farms"* (Emmanuel Acquah, KNUST, 2026). G-YOLOv8 replaces standard backbone convolutions in YOLOv8 with GhostNet Ghost modules ([S]-REPLACE engineering operation, Han et al., 2020), targeting real-time, on-device Maize Streak Virus (MSV) detection on resource-constrained mobile hardware common in Ghanaian smallholder farming contexts. A post-hoc SHAP explainability layer (Lundberg & Lee, 2017) audits the model's spatial reasoning. ## Planned Repository Structure ``` g-yolov8-msv-ghana/ ├── README.md ├── requirements.txt ├── data/ │ ├── public/ # Filtered/relabeled subset of the Roboflow Maize-diseases corpus │ └── field/ # Primary field-validation set (post-ethics clearance) ├── notebooks/ │ ├── 01_baseline_yolov8.ipynb # Phase 1: locked baseline (never modified after first run) │ ├── 02_gyolov8_engineering.ipynb # Phase 2: G-YOLOv8 construction │ └── 03_evaluation_shap.ipynb # Phase 3: comparison, Wilcoxon test, SHAP attribution ├── models/ │ └── checkpoints/ # Mirrored to Google Drive / Kaggle Dataset output (see Section 4D) └── results/ └── comparison_tables/ ``` ## Dataset - **Public baseline:** Maize-diseases (Roboflow Universe), EnochGrahamWS, CC BY 4.0. universe.roboflow.com - **Primary field set:** Collected post-CHRPE ethics clearance (KNUST), minimum 150 images across ≥3 farm plots. ## Status This repository is created ahead of data collection and training, per the project's Phase 0 (reproducible environment setup). Code will be committed incrementally as each phase (baseline → engineering → evaluation) is completed, per the timeline in Section 10 of the proposal. ## Cit …