Sentinel-2 Earth Observation · Transfer Learning · ResNet50
Classify the land.
See the shift.
TerraShift trains ResNet50 classifiers on Sentinel-2 imagery to identify 10 land-cover types, compares RGB against 13-band multispectral input, and uses Grad-CAM to examine what the model is actually looking at — a live interface onto real research, not a static writeup.
Transparent machine learning research connecting Sentinel-2 satellite pixels to empirical land-cover evidence.
Main question
How effectively can transfer learning using Sentinel-2 satellite imagery classify land-cover types, and can these classifications identify potential land-cover changes?
Secondary question
Does multispectral Sentinel-2 imagery improve land-cover classification performance compared with RGB imagery?
01 / DATASET
Dataset Library
Browse the 10 EuroSAT classes and run real inference on any image.
Open →
02 / EMPIRICAL RESULTS
Results
Metrics, confusion matrices, and RGB vs. multispectral comparisons.
Open →