Delhi Airshed Land-Use AI Audit

Air-quality work across the Delhi airshed needs to know what is actually on the ground, built-up against cropland against vegetation, over thousands of square kilometres that nobody is going to label by hand.

• Filtered the tile set to the Delhi-NCR boundary with GeoPandas, testing each tile's coordinates as a point against the region polygon.
• Built labels from the ESA WorldCover 2021 raster: reprojected each tile's coordinates into the raster CRS with Rasterio, read a 128x128 window centred on that pixel, and took the modal class.
• Mapped WorldCover codes into four working classes, built-up, cropland, water and vegetation, with everything else grouped as other.
• Split 60/40 into train and test, then fine-tuned a torchvision ResNet18 with cross-entropy and Adam at lr 1e-3, batch size 32, for 5 epochs.
• Evaluated with accuracy and weighted F1, and saved the confusion matrix, class distribution and grid overlay as figures.

ResNet18 (torchvision)

Fine-tuned to classify each tile into the four land-use classes.

ESA WorldCover 2021

Label source: the modal class inside each 128x128 raster window.

Sentinel-2 RGB tiles named by latitude and longitude, filtered to the Delhi-NCR region and labelled against the ESA WorldCover 2021 raster.

Trains and evaluates end to end, with the confusion matrix, class distribution and grid overlay committed alongside the code.

• Accuracy 81.13% on the 40% held-out split
• Weighted F1 0.76
• 5 epochs, batch size 32, Adam at lr 1e-3

• Labelled each tile by the modal WorldCover class across the whole 128x128 window rather than by its centre pixel, which is more stable on mixed-use tiles.