← Back to Portfolio
OSCD Sentinel-2 Satellite Change Detection
A side project I pursued in depth to learn remote sensing from first principles: change detection on Sentinel-2 bitemporal imagery, pairing classical baselines with a compact Siamese U-Net. Every threshold is chosen on validation data alone, and every claim is labeled VERIFIED or UNVERIFIED, including four early claims that were later corrected on the record.
- Dataset Provenance & Licensing: Sourced OSCD and LEVIR-CD directly from their original providers and tracked each dataset's licence explicitly. LEVIR-CD is characterized but kept out of training to respect its stricter terms.
- Classical Detection Baselines, Built From Scratch: Built absolute-difference, Change Vector Analysis and PCA-K-means detectors, plus spectral-index variants. They set a classical floor of F1 0.237 before any learned model enters the comparison.
- Learned Architecture on Laptop-Scale Hardware: A ~269k-parameter Siamese U-Net (shared encoder, absolute-difference fusion, weighted BCE plus soft Dice loss) reaches mean F1 0.378 across 4 seeds on a single machine with no GPU cluster. That is about 1.6x the classical floor.
- Threshold-Honest Evaluation Protocol: Thresholds are chosen on the validation split only, and whole-image metrics are computed after stitching tiles back together. Results are reported per seed as a range (1.43x to 1.71x) rather than one number.
- Radiometric Nuisance Characterization: Found that brightness drift between image dates (roughly -654 to +609 DN) is as large as the real change signal. A label-free correction for it lifted classical mean AUC from 0.660 to 0.830 across six detectors.
- Statistical Power Analysis: Measured spatial autocorrelation of 24x to 270x across tiles, which puts the effective sample size closer to 1,192 than the raw tile count. Most computer-vision projects skip this check.
- Label-Free Quantification Program: Named the open problem as calibrated, label-free change-rate prediction, then tested strategies from the quantification literature against it, including Median Sweep.
- Adversarial Self-Verification: Four of the project's own claims were later overturned and corrected on the record. The project is backed by 135 tests, including 5 deliberately failing tripwires that catch regressions.