Athithiyanmr/chennai_urban_climate
Deep learning pipeline (UNet) for built-up area extraction from Sentinel-2 imagery — applied to Chennai for urban heat island and climate risk analysis
What's novel
Deep learning pipeline (UNet) for built-up area extraction from Sentinel-2 imagery — applied to Chennai for urban heat island and climate risk analysis
Code Analysis
13 files read · 4 roundsAn end-to-end pipeline that downloads Sentinel-2 imagery from Planetary Computer, builds an 11-band spectral stack with derived indices, generates balanced training patches from Google Open Buildings labels, trains a UNet-Transformer for binary built-up area segmentation, and evaluates with IoU/F1 m
Strengths
The pipeline is genuinely end-to-end and reproducible, with thoughtful details like balanced patch sampling, nodata-aware normalization, sliding-window inference with overlap averaging, auto-threshold search, and rich evaluation outputs (confusion rasters, plots, JSON metrics). Error handling is consistently good with meaningful error messages, file existence checks, and retry logic for downloads.
Weaknesses
Zero test coverage for any component, and the README makes several inaccurate claims (BCE+Dice vs actual Focal+Dice, 256px patches vs 128px default, 10 channels vs 11, augmentation that doesn't exist in code). The model architecture is a standard textbook UNet with a single transformer block bolted on — no novel contribution to the segmentation problem.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
35
Contributors
1
Files
22
Active weeks
5
Repository
Language
Python
Stars
1
Forks
0
License
—