The first question in almost every environmental ML project is which satellite. For free, global, multispectral imagery the answer is usually Sentinel-2, Landsat, or both.
At a glance
| Sentinel-2 | Landsat 8 / 9 | |
|---|---|---|
| Operator | ESA (Copernicus) | NASA / USGS |
| Best resolution (multispectral) | 10 m | 30 m (15 m panchromatic) |
| Revisit | ~5 days with two satellites | ~16 days each, ~8 days combined |
| Archive starts | 2015 | 1972 (programme), 2013 (Landsat 8) |
| Bands | 13 incl. red-edge | 11 incl. thermal |
| Thermal | No | Yes |
| Cost | Free | Free |
Choose Sentinel-2 when
- You need detail: small farms, field boundaries, urban features. 10 m vs 30 m is a nine-fold difference in pixel area.
- You need frequent images: crop monitoring through a season, cloudy regions where more passes means more clear views.
- You want red-edge bands, which are sensitive to vegetation health and useful for crops and forests.
For EUDR-style plot checks, Sentinel-2's resolution matters because many smallholder plots are only a few hectares. See Detecting Deforestation with Satellite Imagery.
Choose Landsat when
- You need history: deforestation since 2000, urban growth since the 1980s, glacier retreat.
- You need thermal data: urban heat islands, land surface temperature, drought stress.
- You're building on long-term products derived from Landsat, like global forest change datasets.
Use both when
- You want dense time series. Combining both gives more observations.
- NASA's Harmonized Landsat Sentinel-2 (HLS) product aligns them to a common 30 m grid and corrects differences, saving you a lot of preprocessing.
Practical tips
- Cloud masking matters more than satellite choice in the tropics and during monsoons.
- Use surface reflectance (Level-2) products, not top-of-atmosphere, for most ML work.
- Mind band names: Sentinel-2 B8 (NIR) is not Landsat B8 (panchromatic). Map bands explicitly.
- Clouds and floods: neither optical sensor sees through clouds. For floods, use Sentinel-1 radar.