Patching - jejjohnson/ml4eo GitHub Wiki

Idea I: Leave Native CRS

We leave the native coordinate system as-is. We can hope that the ML method will be able to account for the weird inconsistencies. A similar code base is done with the

• Geo FNO
• Spherical CNNs
• Efficient Spherical CNNs
• Efficient Generalized Spherical CNNs
• Fast Spherical Convolutions

---

Idea II: Reproject to approximate UTM CRS

We choose a small enough patch size, we estimate the utm coordinates based on the center pixel, then we reproject the data based on that coordinate system.

xds_utm = xds.rio.reproject(xds.rio.estimate_utm_crs())
xds_utm.rio.crs

---

Idea III: Choose a general coordinate system.

Here, we choose a generalized coordinate system like UTM or Major-TOM. Then we create patches based on those grid points.