Automatic sub-pixel co-registration of Landsat-8 OLI and Sentinel-2A MSI images using phase correlation and machine...

Area of Science:

• Earth Observation
• Remote Sensing
• Geomatics Engineering

Background:

• Accurate co-registration of satellite imagery is crucial for multi-sensor data fusion and time-series analysis.
• Misregistration between different satellite sensors (Landsat-8/OLI, Sentinel-2A/MSI) and multi-temporal images from the same sensor can limit the precision of derived geospatial products.

Purpose of the Study:

• To investigate and quantify misregistration errors between Landsat-8/OLI and Sentinel-2A/MSI images at 30 m resolution.
• To assess co-registration accuracy for multi-temporal Sentinel-2A images at 10 m resolution, considering same and adjacent orbit acquisitions.
• To evaluate the effectiveness of phase correlation and various transformation functions for precise image alignment.

Main Methods:

• Employed phase correlation for robust control point identification across images with temporal gaps exceeding 100 days.
• Analyzed co-registration of 45 Landsat-8 to Sentinel-2A pairs and 37 Sentinel-2A to Sentinel-2A pairs.
• Compared performance of non-linear Random Forest regression and 1st order polynomial (affine) transformation functions for mapping misregistration.

Main Results:

• Observed misregistration up to 1.6 pixels (30 m) between Landsat-8 and Sentinel-2A, and up to 1.2 and 2.8 pixels (10 m) for same and adjacent Sentinel-2A orbits, respectively.
• Random Forest regression achieved superior accuracy, with average Root Mean Square Errors (RMSE) of 0.07±0.02 pixels (30 m) and 0.09±0.05/0.15±0.06 pixels (10 m).
• Affine transformation resulted in slightly higher RMSEs: 0.08±0.02 pixels (30 m) and 0.12±0.06/0.20±0.09 pixels (10 m).

Conclusions:

• Phase correlation is a reliable method for co-registering satellite images, even with significant temporal separation.
• Random Forest regression significantly improves co-registration accuracy compared to simpler affine transformations for both inter- and intra-sensor comparisons.
• Accurate co-registration is achievable, essential for leveraging the full potential of multi-sensor and multi-temporal satellite data in Earth observation applications.