Surface reflectance is the fraction of incoming light that a surface reflects, measured at the ground rather than at the satellite. It is a property of the material itself. Radiance at the sensor is not, because the atmosphere sits in between and changes the signal in both directions.
L1 and L2
Most Earth observation archives distribute data at more than one processing level.
Level 1 products carry top of atmosphere radiance, or top of atmosphere reflectance derived from it. This is the calibrated, geolocated measurement of energy arriving at the instrument. It includes everything the atmosphere did to the light: scattering that added a haze signal, absorption by water vapour, ozone, carbon dioxide and other gases that removed energy at specific wavelengths, and aerosol effects that vary by scene and by day.
Level 2 products carry surface reflectance, sometimes called bottom of atmosphere reflectance. This is the estimate of what the ground would have reflected with the atmosphere removed.
Why the correction matters
For any spectral analysis, this step is not optional. Atmospheric water vapour absorbs strongly around 1.4 and 1.9 microns, and it varies from scene to scene. Carbon dioxide and methane leave their own structure in the shortwave infrared. Left uncorrected, that structure sits on top of the surface signal and will be misread as material chemistry, because it appears exactly as absorption features in the same wavelength regions where mineral features live.
Two spectra of the same outcrop acquired on different days will differ at Level 1 and agree at Level 2, assuming the correction worked. Comparison across time or across scenes only means something after correction.
How it is done
Atmospheric correction models radiative transfer through the atmosphere, estimating the state of the atmosphere from the scene itself and from ancillary data, then inverting the model to recover ground reflectance. Water vapour is often retrieved per pixel from the depth of its own absorption bands. Aerosol optical depth is harder and is usually the largest remaining source of error.
The correction is an estimate, not a measurement, and it fails in predictable places: thin cirrus, deep shadow, very dark surfaces where the atmospheric contribution dominates the total signal, and steep terrain where the illumination geometry departs from the model assumption. Those pixels should be masked rather than analysed. A surface reflectance product without an accompanying quality mask is only half a product.