I have a rather fundamental question about the use of optical data to derive spectral indices. Is it better to use TOA or BOA reflectance values when deriving an index such as NDWI, or NDVI, etc.
From my point of view, the result of a spectral index is unitless, not a physical parameter, thus the relation is kept whether I use TOA or BOA reflectance. Is this assumption correct or am I missing something?
To put it in other words, should BOA be used instead of TOA for spectral index derivation
You are missing something.
The TOA represents radiance rather than surface reflectance values.
Radiance is often effected by atmospheric conditions that get corrected for when processed into BOA such as Sentinel-2 L2A products.
The actual correction factor will vary across the spectral bands.
So the band-pair ratio combinations like NDVI (B8-B4)/(B8+B4) that you are computing from the TOA products are likely to be significantly different from those of the processed BOA products.
From what I’ve seen, if you have enough BOA data then it’s best to use BOA because it removes the effects of the atmosphere so you get the true surface reflectance values.
TOA can be “good enough” depending on your area. It depends if your area has a lot of atmospheric haze or not and how much error from atmospheric effects you’re willing to accept in your data.
As others have already mentioned, BOA removes atmospheric influences. For time-critical use-cases (fires, flooding) TOA results are, however, available more quickly.
There may be areas/times where BOA and TOA give similar results, so there can be use-cases where TOA is good enough, but in general you need to compare BOA and TOA to determine when TOA could be useful. This can be part of a “postmortem” investigation after TOA were used for time-critical incidents to fine-tune your understanding of when TOA results are adequate.
I would quickly add that most of the signal measured at TOA, specially in the NIR bands above water surfaces, is almost aerosol (Mie) and molecular (Rayleigh) Scattering, since water (e.g, above 700 nm without any relevant suspended matter) is pratically black (null water leaving radiance), so any reflectance measured in these bands are not from the water-surface but rather by the atmosphere. This same principle, should apply to any surface measurement, since radiance / reflectance along the light spectrum and path will have relavant inter-band differences (relative radiance variations) and magnitudes, depending on type of backscattered surface and atmospheric conditions at the time of data acquisition. Therefore, the band ratio (by itself) isn’t enough to account for (eliminate) the differently affected band radiances attributed to the atmospheric composition above the targeted surface.