Remote Sensing Basics

Before going into the details of BigEarthNet, the following section will give a short introduction to common terms of the remote sensing domain.

Spatial resolution

One important concept when working with remote sensing data is the spatial resolution of an image. The spatial resolution describes how large a pixel is. A spatial resolution of 10m indicates that each pixel covers a region with a width and height of 10m. In other words, objects smaller than 10m² cannot be represented, as they are too small.

Spectral reflectance

When light hits a surface, the surface partially absorbs and reflects the incoming electromagnetic spectrum. The reflected electromagnetic spectrum is what an observer (our eyes or a satellite) perceives and processes. We perceive a green leaf as green, as the green spectrum is the largest reflected spectrum.

Band

A band represents a segment of the electromagnetic spectrum. For example, the green band is often defined as the spectrum with a wavelength from 500 to 565 nanometers.

Tiles or Granules

When talking about satellite images, a tile (also granule) often refers to predefined regions. The Sentinel-2 mission has defined 100km x 100km tiles that divides the earth into 60 unique zones.

Patches

Patches is used as a term to specify a valid portion of a tile. “Patch” is not a standardized term, but in the BigEarthNet documentation, it refers to a 1200m x 1200m area of a Sentinel tile.

It is necessary to split a tile into multiple patches further because tiles are too large to be effectively used for machine learning. For comparison, a single Sentinel-2 tile is ~800MB large, while the size of an individual BigEarthNet patch is around ~200kB.

Note

The spatial resolution can differ between bands from the same satellite.