SAVI — Soil-Adjusted Vegetation Index

In 1988 Alfredo Huete published a paper in Remote Sensing of Environment (volume 25, page 295) that began with a measurement rather than a theory. He put cotton and range grass over soils of markedly different brightness and looked at where the NDVI isolines fell in the red–near-infrared plane. The theory said they should converge on the origin. The data said they converged somewhere below it — and that discrepancy meant NDVI over partial canopy was reading the soil, not the plants.

His fix is almost embarrassingly simple: add a constant $L$ to the denominator to move the convergence point to where the data actually put it, and multiply the top by $(1+L)$ to rescale the result back into roughly NDVI's range so the two can be read against each other. $\mathrm{SAVI} = (\rho_{NIR}-\rho_{red})(1+L)/(\rho_{NIR}+\rho_{red}+L)$. One constant, one rescaling. That simplicity is most of why it was adopted so widely.

The design detail worth noticing is that at $L = 0$, SAVI is exactly NDVI. Both new terms vanish and the expression collapses to the older index, identically, not approximately. Huete built it that way deliberately: the correction is an extension rather than a replacement, so a SAVI value and an NDVI value from the same pixel can be compared, and the difference between them is a direct read on how much soil background NDVI was counting as canopy.

Choosing $L$ is where the honesty is required. Huete's recommendation was 1.0 for very low vegetation cover, 0.5 for intermediate, and 0.25 for high cover. In practice 0.5 has become the universal default, usually applied without anyone checking whether the cover is intermediate. The awkwardness is circular and unavoidable: the correct $L$ depends on how much canopy there is, which is the thing you were trying to measure. Two later indices exist entirely to escape that circle — OSAVI fixes $L$ at 0.16, which its authors found optimal across a wide range of agricultural cover, and MSAVI solves for the soil adjustment from the image itself using the soil line. Both are worth knowing about, and both concede the same point: a constant nobody can know in advance is a weakness, however useful the correction is.

The practical guidance is straightforward. Reach for SAVI when a substantial part of each pixel is bare ground — early season before canopy closure, row crops between the rows, thin or failed stands, rangeland, orchards and vineyards with bare inter-rows, arid and semi-arid sites, and anything recently planted. Those are precisely the conditions in which an agronomist most wants a reliable reading and in which NDVI is least trustworthy. Once the canopy has closed, the soil term stops mattering and SAVI converges back toward NDVI, so on a dense mid-season crop the choice barely matters.

What SAVI does not fix is as important as what it does. Saturation at high leaf area index is untouched — SAVI flattens out at dense canopy just as NDVI does, because the ceiling comes from near-infrared scattering physics and has nothing to do with soil. And SAVI does absolutely nothing about calibration. A soil correction applied to raw digital numbers from an uncalibrated consumer camera corrects nothing, because $L$ is defined in units of reflectance and a DN is not one: adding 0.5 to a number that is not a reflectance is arithmetic without meaning. SAVI computed that way is exactly as incomparable between flights as NDVI computed the same way. Reflectance panels, per-band radiometric calibration and a downwelling irradiance sensor are the price of a comparable time series, and no index formula substitutes for them.

One last practical caution: SAVI values computed at different $L$ are not comparable with each other, and published figures do not always state which was used. If you are reading somebody else's index and you also have the bands, you can solve for the $L$ they must have used — which is worth doing before putting their numbers beside your own.