A model meets a different city.

My role

I worked with a teammate on the inference pipeline and geospatial deep-learning toolchain.

Where it stands

LUMS course project. Uses pretrained SpaceNet weights; no from-scratch training or quantitative IoU/F1 evaluation is claimed.

What changed.

Building-footprint masks from local satellite imagery, with qualitative inspection of where the model transferred and where it struggled.

Make inference work end to end

I helped assemble the Solaris/GDAL environment, reconstruct pretrained weight files and connect resizing, data preparation and inference. The model required a consistent image size, so preprocessing was part of the result.

Keep the evidence at its actual scale

The output was assessed qualitatively. Dense residential layouts were the motivating case; close-up imagery exposed weaknesses. Training on local data and measuring overlap against labeled footprints remained future work.

The stack

• PyTorch
• Solaris
• GDAL
• Docker
• VGG16 U-Net
• SpaceNet-7