🕐 Status: Paper submitted to the 10th International Conference on Computer Vision and Image Processing (CVIP 2025), IIT Ropar, Punjab, India — currently under review.
Authors: Karan Sharma, Rajiv Ranjan, Dinesh Kumar, Shashank Tamaskar Center for Sustainable & Precision Agriculture, Plaksha University, Mohali, India
The CVIP 2025 conference poster is available in this repository:
Your browser does not support embedded PDFs. Download the poster here.
Sugarcane germination in India is still assessed through manual inspections — slow, error-prone, and impractical at scale. CGMap is a geospatially aware, deep-learning-based pipeline that automatically detects early-stage sugarcane saplings from UAV imagery and identifies missing-plant regions ("gaps").
A lightweight YOLOv8 object detection model provides plant locations, which are converted into a georeferenced point cloud for spatial analysis. A Minimum Spanning Tree (MST) based orientation normalization technique enables reliable row and column extraction even in irregular Indian farm layouts. Gap detection is performed using expected intra-row spacing criteria, and all outputs are exported in GIS-ready WKT format for agronomic decision-making.
CGMap enables timely transplantation interventions, supports resource-efficient field management, and provides a scalable, interpretable, and generalizable solution for precision agriculture.
The pipeline is an end-to-end geospatial computer-vision system:
UAV Imagery
↓
Orthomosaic Generation (Asmoli farms, Uttar Pradesh)
↓
Farm Boundary Extraction (from GIS polygon coordinates)
↓
YOLOv8 Sapling Detection (tiled 300×300px chips)
↓
Centroid → Georeferenced Point Cloud
↓
MST-Based Field Orientation Normalization
↓
Row / Column Structural Extraction (linear fitting + perpendicular thresholding)
↓
Gap Detection (inter-plant distance vs. expected spacing)
↓
GIS Export (WKT: plant points, row lines, detected gaps)
| Step | Description |
|---|---|
| Data Collection | UAV surveys over sugarcane farms in the Asmoli region, UP |
| Orthomosaic Stitching | High-resolution aerial maps with accurate field surface geometry |
| Object Detection | YOLOv8 trained on annotated sapling imagery; confidence threshold: 0.1 |
| Point Cloud Construction | Each detection centroid projected to global pixel coordinates |
| MST Orientation | Fully connected graph → MST → dominant field orientation vector |
| Row Extraction | Point cloud rotated to aligned frame; rows/cols via linear regression |
| Gap Detection | Inter-plant distance anomalies flagged as single or consecutive gaps |
| GIS Export | WKT-format CSVs: plant points, structural lines, gap annotations |
CGmap/
├── main.py # Full pipeline entry point
├── best.pt # Trained YOLOv8 model weights
├── detected_plant_points.csv # Output: detected plant coordinates
├── orthomosiacs/ # Input orthomosaic .tif files
│ └── model_farm_2.tif
├── data_asmoli_model/ # 📦 Publicly released training dataset (see below)
└── 851_KaranSharma.pdf # CVIP 2025 conference poster
The training data in data_asmoli_model/ is a publicly released dataset made available for the research community.
It was manually collected via UAV over sugarcane farms in the Asmoli region, Uttar Pradesh, India, and hand-annotated by the authors for early-stage sapling detection. This is, to our knowledge, one of the few openly available annotated datasets for sugarcane seedling detection in Indian farm conditions.
📌 Dataset DOI / citation coming soon — will be updated upon publication.
You are free to use this dataset for research purposes. If you do, please credit this repository and the authors until the formal citation is available.
pip install ultralytics rasterio opencv-python numpy pandas matplotlib networkx scikit-learn scipy shapelyEdit the top of main.py to point to your data:
input_file = 'orthomosiacs/model_farm_2.tif' # Input orthomosaic
model_path = 'best.pt' # YOLOv8 weights
chip_width = 300 # Tile width (px)
chip_height = 300 # Tile height (px)
conf_threshold = 0.1 # Detection confidence
MST_THRESHOLD = 27 # Max MST edge length for row groupingpython main.pyOutputs:
detected_plant_points.csv— all detected plant centroidsmodel_farm_1_points.csv— georeferenced WKT plant pointsmodel_farm_5_lines_wkt.csv— structural row lines in WKT- Gap annotations ready for QGIS / agronomy dashboard import
- Test Site: Sugarcane farms, Asmoli region, Uttar Pradesh, India
- Detection Backbone: YOLOv8 (model-agnostic — compatible with Faster R-CNN, EfficientDet, etc.)
- Runtime: Complete detection + analysis in tens of milliseconds (GPU) / sub-second (mid-range CPU)
- Edge Deployment: Practical runtimes on Raspberry Pi 4 and Pi 5 — suitable for field-deployed devices
- Spatial analysis scales with number of detected plants, not image resolution → efficient at any scale
All outputs are exported as WKT (Well-Known Text) — directly importable into:
- QGIS ✅
- ArcGIS ✅
- Any agronomy decision-support system with GIS capability ✅
Gap types detected:
| Symbol | Meaning |
|---|---|
| 🟢 Detected plants | Healthy plant present |
| 🔴 Single missing plant | Isolated gap |
-- Multiple missing plants |
Consecutive bald patch |
- Expand dataset across multiple seasons and regions
- Integrate multispectral cues (NDVI, NIR) for improved gap detection
- Add temporal growth pattern analysis
- Ground-truth validation against manual emergence surveys
- Generalize framework to other row crops with similar planting geometries
This research was supported by CNH Industrial under Project Pahal, a CSR initiative. Special thanks to Keshika Gajbhiye for manuscript proofreading assistance.
Karan Sharma M.Eng (Software Systems & Robotics), University of Technology Sydney Former Research Fellow, Center for Sustainable & Precision Agriculture, Plaksha University
- GitHub: @karanS08
- Contact: 0802karanS@gmail.com
🕐 Citation will be updated upon publication. For now, if you use this work or dataset please reference this repository.
Made with 🌾 at Plaksha University | Submitted to CVIP 2025, IIT Ropar