📦 UAV Crop Classification using Patch-Based Random Forest

A machine learning pipeline for classifying crop types from UAV multispectral imagery using patch-level analysis and Random Forest classifier.

🚀 Features Implemented

• ✅ Patch extraction from orthomosaic TIFFs using polygon-labeled shapefile

• ✅ Supports multi-size patch testing: 128x128, 256x256, and non-square 242x272

• ✅ Filters mixed-label patches to ensure clean supervision

• ✅ Feature extraction per patch:

  • Mean / Std / Max of 10 spectral bands
  • NDVI & GNDVI vegetation indices

• ✅ Class balancing with SMOTE (only when safe)

• ✅ RandomForestClassifier with class_weight="balanced"

• ✅ Train-test split with stratification (80/20)

• ✅ Accuracy reporting + classification_report

• ✅ Automatic confusion matrix PNG export per patch size

• ✅ Accuracy comparison bar chart: patch_size_accuracy_bar.png

---

🧾 Required Inputs

• TIFF images: UAV-captured multispectral reflectance orthophotos

• Shapefile (.shp, .shx, .dbf, etc.): contains crop-type polygons with field crop

• Folder structure:

  Project/
  ├── ms data/
  │   ├── UAV3-MS/          # Contains .tif images
  │   └── metadata/
  │       └── md_FieldSHP/  # Contains shapefile components
  

---

📂 Output

After running the script, you will get:

• confusion_matrix_<WxH>.png for each patch size
• patch_predictions_<WxH>.csv with per-patch prediction results
• patch_size_accuracy_bar.png comparing different patch sizes

---

🧠 How to Run

Install dependencies:

pip install -r requirements.txt

Run the script:

python patch_rf_classifier_multi.py

---

🔬 Next Suggestions (Optional Enhancements)

• Export each patch as .png for manual review
• Add GLCM-based texture features (contrast, homogeneity)
• Add LightGBM or CNN/Vision Transformer comparison
• Add time-aware modeling if image timestamps are available

---

📌 Authors / Group

• ✍️ Code: Marius (with ChatGPT assist)
• 📷 Data: UAV team
• 🧠 Model logic: ML research group

---

📄 License

MIT or Creative Commons (customizable)

Ready to push to GitHub and share with your group! 🚀