Processing levels

In depth

Agencies describe the chain in levels, which Figure 1 summarises. Level 0 holds the raw measurements as downlinked. Level 1 converts them to physical units, such as radiance or top-of-atmosphere reflectance for optical sensors, and attaches position. Level 2 derives a geophysical property of the surface, such as surface reflectance. Level 3 maps or composites that property on a regular grid, and level 4 is the output of models or analysis. The numbering is a convention and sub-levels vary, so the product specification is the authority.

Two steps do most of the work between levels 1 and 2 for an optical image. The first is orthorectification, which removes the displacement caused by relief. A point at height $h$ viewed at an angle $\theta_v$ from the vertical is displaced on the image, relative to its true map position, by about

where $\delta$ is the displacement of the point on the image from its true map position, measured on the ground, in $\mathrm{m}$; $h$ is the height of the point above the reference surface of the elevation model, in $\mathrm{m}$; and $\theta_v$ is the viewing angle from the vertical at that point.

so a 50 m building viewed 20 degrees off the vertical is displaced by about 18 m, and the correction needs an elevation model [1]. The Sentinel-2 level 1C product is orthorectified and reports top-of-atmosphere reflectance [2]. The second step is atmospheric correction. It inverts equation 3 on the optical imaging page for the surface reflectance, using estimates of the path radiance, transmittance and ground irradiance from radiative transfer models and from the aerosol and water vapour content of the scene.

where $\rho$ is the surface reflectance, without unit; $L$ is the spectral radiance at the sensor and $L_p$ the path radiance, both in $\mathrm{W\,m^{-2}\,sr^{-1}\,\mu m^{-1}}$; $\tau$ is the transmittance of the atmosphere from the ground to the sensor, a fraction without unit; $E_g$ is the spectral irradiance reaching the ground, in $\mathrm{W\,m^{-2}\,\mu m^{-1}}$; and $\pi$, in $\mathrm{sr}$, relates the radiance of a surface that scatters equally in all directions to the irradiance it reflects.

The widely used 6S code is one such radiative transfer model [3], and the processor that produces Sentinel-2 level 2A surface reflectance, with a scene classification that flags cloud and shadow, is Sen2Cor [4]. For radar the sequence is similar. Single look complex products keep the phase needed for interferometry, detected products are projected to the ground and averaged to reduce speckle, and terrain-corrected products use an elevation model to correct the geometry and the brightness [5, 6].

Analysis-ready data takes the sequence a step further. The idea is to process imagery to a minimum common set of requirements, with consistent calibration, geolocation and masks, so that users can work with it directly and compare scenes without doing the corrections themselves [7].