Computer vision

Satellite land use and change

Land use from all 13 Sentinel-2 bands with a band-ablation study, and building change detection with a Siamese U-Net.

Before and after aerial tiles with predicted building changes

The problem

Sentinel-2 records 13 spectral bands, but most vision models only see RGB. I wanted to know how much the other bands add once a CNN already has spatial context, and whether they help more when labels are scarce. The second half asks a related question for change detection: can raw pixel differences find new buildings, or does it take a learned model?

What I built

  1. Rebuilt the first layer of an ImageNet ResNet-18 for any number of bands: the RGB bands keep their filters and every other band starts from their mean.
  2. Ran a band ablation on one fixed split of 27,000 EuroSAT patches: 9 band sets, a random forest that sees only per-band statistics, and 1 %/10 %/100 % label budgets.
  3. Trained a Siamese U-Net with a shared ResNet-18 encoder on LEVIR-CD. The decoder gets the absolute feature difference of the two dates, so unchanged content cancels.
  4. Compared it with Change Vector Analysis, once with Otsu per tile and once with a threshold tuned on validation, and exported both models to ONNX and CoreML.

Results

EuroSAT test accuracy and macro F1 on 5,400 patches, one seed per configuration.
ModelBandsTest accuracyMacro F1
Random forestRGB0.7970.789
Random forestall 130.9100.905
ResNet-18RGB0.9830.982
ResNet-1810 non-visible bands only0.9830.982
ResNet-18all 130.9870.987

What the results show

  • Spectra and spatial context largely stand in for each other. Without spatial features the extra bands are worth 11 points; with a CNN they add 0.4 points (91 vs. 68 errors, McNemar p = 0.02), mostly in vegetation classes.
  • I expected extra bands to help more with few labels, and they didn't: RGB and all 13 both score 0.896 with 1 % of the labels.
  • For change detection the learned model reaches F1 0.890, while CVA stays at 0.117 even with a tuned threshold. Pixel differences can't tell buildings from seasons, shadows or new roads.
  • Quantising raw 13-band input with one 8-bit scale zeroes the cirrus band and drops accuracy to 0.866, so the input scaling stays in float.

Limitations

  • EuroSAT's random split lets neighbouring patches of one field land in train and test, which flatters every model. A geographic split would measure the leak.
  • One seed per configuration, and validation accuracy was still rising at epoch 8.
  • The LEVIR-CD mirror stores JPEGs, so the change results aren't a like-for-like comparison with papers.

Full results, commands and figures are in the repository README.