This project detects wells and ponds (dry and wet) within a specified geographic region using satellite imagery tiles and a YOLO-based object detection model. It processes images, extracts polygons of detected ponds, filters based on entropy (for wet ponds), converts pixel coordinates to geospatial coordinates, and generates a combined shapefile of all detected ponds.

• Overview
• Dependencies
• Folder Structure
• Workflow
• Usage Instructions
• Output
• Notes

• Input: GeoJSON file of the target region (Masalia_mws.geojson)
• Process: Download satellite image tiles for the region using tile coordinates. Detect ponds (Dry/Wet) using a pre-trained YOLO model. Compute entropy to filter out noisy wet pond detections. Store polygon masks and geospatial data into a CSV file. Convert polygons to GeoJSON and combine overlapping geometries. Export the final combined geometry as a shapefile and ZIP it.

Ensure the following Python packages are installed: pip install numpy pandas geopandas opencv-python pillow requests matplotlib shapely ultralytics scikit-image

• YOLOv11 model via Ultralytics
• Shapely, GeoPandas for geospatial operations
• OpenCV, scikit-image for image processing
• requests for downloading map tiles

• Data/Zoom17/Masalia/ # Image tiles (downloaded here)
• Shapefiles/Masalia_mws.geojson # Input GeoJSON boundary
• CSV_Output/Masalia_Ponds.csv # Output CSV with pond data
• Shapefile_Output/ # Output shapefile ZIPs
• Ponds_best.pt # Pre-trained YOLO model

1. Download Map Tiles

• Converts bounding box of GeoJSON to tile numbers.
• Downloads tiles from Google Maps (satellite imagery) at zoom level 17.

2. YOLO Inference

• Runs object detection using Ponds_best.pt.
• Filters detections using confidence thresholds.
• For wet ponds, calculates entropy to ensure detection validity.

3. Polygon Conversion

• Extracts polygon mask coordinates.
• Converts pixel coordinates to latitude/longitude using tile geo-boundaries.

4. CSV Generation

• Stores polygon data, class labels, entropy, tile locations into a CSV.

5. GeoJSON and Shapefile Creation

• Converts polygons to GeoJSON features.
• Merges overlapping geometries using buffer-union-buffer method.
• Exports as shapefile and compresses it into a ZIP file.

1. Set Paths and Parameters

• Modify the following in the script as needed: model_path, csv_file, image_dir, zoom_level, scale, entropy_threshold

2. Run the Script

• Execute the full script in your Python environment.

3. Output Files

• Processed CSV at: CSV_Output/Masalia_Ponds.csv
• Final shapefile ZIP at: Shapefile_Output/Masalia_Ponds_COMBINED_GEOMETRY.zip

• CSV File: Contains wells and pond classification (Dry/Wet), polygon coordinates, entropy values, tile locations.
• Shapefile (Zipped): Combined geometries of all detected ponds and wells in EPSG:4326.

• Entropy is only computed for wet ponds to filter out noisy detections.
• Tile scale of 1 results in 256x256 pixel images.
• Merging buffer distance can be adjusted (0.0005) for sensitivity.

Similar to ponds and wells, this project includes a module for detecting plantations using zoom level 17 satellite tiles and a YOLOv11-based model. The pipeline involves:

• Tile download from Google Maps at zoom 17

• Object detection of plantation areas using a trained model (Plantations_best.pt)

• Extraction of polygonal masks for detected plantations

• Geospatial conversion and union of nearby polygons to capture continuous plantation patches

• Final output includes CSV with detected polygons and a merged shapefile of plantation zones

To improve the model’s precision and reduce false positives, we used a CutMix-style augmentation strategy. Here's how it works:

• Negative tiles (tiles without plantations) are taken as background

• One valid annotated polygon from real plantation data is randomly selected and pasted onto a random location within a negative tile using its exact mask

• Corresponding labels are accurately transformed to account for new positions

• This process is repeated to create multiple synthetic samples

• The final result is a curated dataset of single-plantation CutMix tiles, which enhances the model's ability to distinguish real plantations from noisy patterns in natural landscapes

This augmentation strengthens the model's discrimination capability and lowers the occurrence of false positive detections in non-agricultural regions.