AI GeoImage Coregistration

A Python package for automated georeferencing of historical satellite imagery (e.g., Hexagon KH-9) using Deep Learning.

This tool uses state-of-the-art feature matching algorithms (LoFTR and SuperPoint+SuperGlue) to find corresponding points between a raw, unreferenced image and a modern georeferenced map. It then automatically warps the raw image using Polynomial or Thin Plate Spline (TPS) transformations.

Features

• Automated GCP Generation: No manual clicking required. Uses AI to find thousands of tie points.
• Robust to Temporal Changes: Handles season changes, urban development, and appearance differences between 1970s and 2020s imagery.
• Large Image Support: Automatically tiles gigapixel images to process them piece by piece.
• Smart Filtering: Uses RANSAC and Spatial Grid Binning to remove outliers and ensure geometric consistency.
• Warping: Outputs both Polynomial (Order 2/3) and TPS (Thin Plate Spline) GeoTIFFs.

---

Prerequisites (Important)

Before installing this package, GDAL must be installed on your system.

Ubuntu / Debian

sudo apt update
sudo apt install libgdal-dev gdal-bin

Conda (Recommended for Windows/Linux/Mac)

conda install -c conda-forge gdal

---

Installation

You can install the package directly from GitHub:

pip install git+https://github.com/sergeikharchenko/ai_geoimage_coreg.git

Optional: SuperGlue Setup

By default, the tool uses LoFTR, which downloads its own weights automatically. If you want to use the SuperGlue algorithm (high precision for detailed structures), you must clone the official repository due to license restrictions:

git clone https://github.com/magicleap/SuperGluePretrainedNetwork.git
export PYTHONPATH=$PYTHONPATH:$(pwd)/SuperGluePretrainedNetwork

---

Data Preparation (Reference Image)

To get the best results, you need a good Reference Image.

1. Format: It must be a GeoTIFF (.tif) with a valid coordinate system (CRS).
2. Resolution: Try to match the resolution of your input image. For Hexagon (~1-2m/px), standard satellite imagery works well.
3. Source: The easiest way to get a reference is using QGIS:
   • Open QGIS.
   • Add a basemap (e.g., ESRI Satellite, Google Satellite, or Bing Aerial).
   • Zoom to the area covering your historical image.
   • Go to Project -> Import/Export -> Export Map to Image.
   • Crucial: Check "Append georeference information (embedded)" or "Save World File".
   • Increase the resolution (DPI) until the "Output width/height" in pixels is sufficient (e.g., 5000-10000 px).
   • Save as .tif.

---

Usage

Run the tool from the command line:

ai_geoimage_coreg --input /path/to/raw_hexagon.tif --ref /path/to/reference_map.tif --output result_name

Arguments

Argument	Default	Description
--input	Required	Path to the raw, unreferenced historical image (TIFF).
--ref	Required	Path to the georeferenced modern image (TIFF).
--algo	loftr	Matching algorithm: loftr (robust) or superglue (precise).
--output	georef	Prefix for output files.
--tile-size	800	Processing tile size in pixels. Reduce to 600 if running out of GPU memory.
--step-size	400	Sliding window step. Lower value = more points but slower processing.
--margin	400	Search margin on reference image. Increase if initial coarse alignment is poor.
--conf	0.75	Neural network confidence threshold (0.0 - 1.0).
--grid	100	Spatial binning size (px). Keeps only the best point per grid cell to avoid clustering.
--ransac	500.0	RANSAC threshold (meters/pixels). Increase for mountainous terrain or high distortion.

Examples

1. Standard run (recommended for most cases):

ai_geoimage_coreg --input KH9_strip.tif --ref modern_map.tif

2. High Precision (slower): Use a smaller step to find more points and a finer grid for better distribution. Good for detailed urban areas.

ai_geoimage_coreg --input KH9.tif --ref modern.tif --step-size 200 --grid 50

3. Difficult Terrain (Mountains/Distortion): If valid points are being deleted because the image is very warped (panoramic distortion), increase the RANSAC threshold.

ai_geoimage_coreg --input KH9.tif --ref modern.tif --ransac 1000.0

---

Algorithms

• LoFTR (Detector-free): Recommended default. It works exceptionally well on images with significant differences (e.g., 50-year gap). It produces dense matches even in low-texture areas.
• SuperPoint + SuperGlue: Excellent for urban areas with distinct corners and buildings. Requires separate installation (see above).

License

This code is provided as-is.

• LoFTR is Apache 2.0 licensed.
• SuperGlue has a non-commercial license (CC-BY-NC-4.0).

---

Author: Sergei Kharchenko