Research

TerraShift studies whether transfer learning on Sentinel-2 imagery can classify land cover well enough to support change detection, and whether multispectral input helps.

Research questions

Main: How effectively can transfer learning using Sentinel-2 satellite imagery classify land-cover types, and can these classifications identify potential land-cover changes?
Secondary: Does multispectral Sentinel-2 imagery improve land-cover classification performance compared with RGB imagery?

Dataset

EuroSAT: 27,000 labeled Sentinel-2 patches across 10 land-cover classes, available in both RGB (3-channel) and multispectral (13-band) form. EuroSAT is already organized into labeled classes and based on real Sentinel-2 data, which makes it a controlled starting point before moving toward temporal change-detection analysis.

Beyond classification: change detection

EuroSAT is a single-snapshot classification dataset, not a change-detection dataset. The planned change-detection stage classifies year-1 and year-2 imagery of the same area separately, then compares the results — while accounting for cloud cover, seasonal differences, geographic alignment, and model uncertainty. A classification difference is a candidate transition to investigate, never an automatic “confirmed” environmental change.