NDVI vs NDRE in perennial crops: when each index lies to you

What each index measures

A healthy leaf absorbs almost all red light (chlorophyll uses it) and reflects a great deal of near infrared (NIR), because the internal structure of the leaf scatters it. Every vegetation index lives off that contrast. NDVI measures it with red: (NIR − Red) / (NIR + Red). NDRE measures it with the red edge, the 730 nm band that sits right at the transition: (NIR − RedEdge) / (NIR + RedEdge).

The difference between them is not the formula, it is depth. Red is exhausted in the first leaves of the canopy: once the surface is covered in green leaf, more leaf underneath no longer changes the signal. The red edge penetrates further and keeps responding after red has given up. That is where the whole practical rule comes from: NDVI is better with sparse cover, NDRE with dense cover.

Two more show up in any report. GNDVI swaps red for green and is somewhat more sensitive to nitrogen. SAVI is an NDVI with a constant that damps the effect of bare soil; that constant assumes reflectance from 0 to 1, a fact worth keeping in mind for the last section of this guide.

Where NDVI lies

It saturates. In a mature closed-canopy orchard, NDVI over almost the whole plot falls between 0.85 and 0.92, and that range holds both the excellent tree and the one that already started to decline. The map comes out evenly green and the grower concludes his orchard is uniform; what is uniform is the index. It is the most common mistake of all, and stretching the color ramp does not fix it: the information is no longer in the data.

It mixes soil and plant. In young or widely spaced plantings, much of each pixel is soil. A low NDVI can mean a weak tree or simply that the pixel held more dirt than leaf. When cover is sparse, measure per canopy —not per average pixel— and use SAVI if you are going to look at the full map.

In the comparator below, try the closed-canopy case: two plants in clearly different condition separated by 0.06 of NDVI and by three times as much with NDRE. Then try the young-plant-over-bare-soil case, where the opposite happens and NDVI is the one that separates better. That crossover is the whole guide in two clicks.

Where NDRE lies

It needs a real red-edge band. RGB cameras do not have one, and neither do RGB cameras modified with a red filter; those produce something similar to NDVI, with a channel that is not truly NIR. If your equipment has no 730 nm band, NDRE does not exist for that flight no matter what checkbox the software offers.

It is more fragile. The contrast between NIR and red edge is much smaller than between NIR and red, so the same calibration noise weighs more: without a calibration panel photographed before and after, or with clouds changing the light mid-mission, NDRE moves for reasons that have nothing to do with the plant.

And it does not rescue sparse cover. With a lot of soil in the pixel, NDRE is contaminated just like NDVI. It is not a better index: it is an index for a different regime.

The costliest mistake: swapped bands

This is the one we have seen most often, and the most expensive because it is invisible. The order of the bands inside the file does not have to be the order of the sensor: processing reorders them, adds the transparency channel as one more band, and does not always write down their names. If your software computes NDVI assuming band 4 = NIR and band 4 is actually the red edge —or the alpha channel, which is constant— you will get a smooth, plausible, false map.

It happened to us with files where the alpha channel landed in the NIR position: NDVI came out near +1 across the entire plot. Nobody questioned it for weeks, because a healthy plot in uniform green is exactly what one expects to see.

How to check it in a minute, with no exotic tooling:

• Read the description of each band in the GeoTIFF metadata instead of deducing it from the band count; the count includes the mask or alpha and it lies
• Look at a dirt road or a patch of bare soil inside the plot: its NDVI should land near 0 (never 0.9). If the road is as green as the crop, the bands are wrong
• Look at a roof, a tarp or water: water must come out negative
• Check the index histogram: if the whole plot fits in a 0.05 range it is not homogeneous, you are looking at a constant

Absolute scale versus relative scale

An index can be computed over calibrated reflectance (physical values from 0 to 1, corrected with a panel or with the sensor model) or over the raw values the camera wrote. Both produce a map that ranks worst to best correctly within that flight. Only the first produces a number comparable with last month's flight.

With 8-bit products the problem is harder still: the absolute scale is no longer recoverable from the file, so the index only supports a relative reading —this plot against the one next door, today. That is why, when the source does not let you state the scale, the honest thing is to say so in the map legend instead of labeling 'low' and 'healthy' over numbers that do not support it.

The same happens with satellite, in a very concrete way: since 2022 Sentinel-2 scenes encode reflectance with an offset of 1000 (the file value is reflectance × 10000 + 1000). Anyone computing NDVI without subtracting that offset gets a biased series, and the bias changes with the scene date, which is the worst way to be wrong: it looks like a crop trend.

Which index for which problem

The operational summary, with what we actually do in each case:

• Vigor inside a mature, closed orchard: NDRE. NDVI no longer discriminates there.
• Young planting, wide spacing, plenty of visible soil: NDVI (or SAVI for the full map), measured per canopy rather than per pixel.
• Telling crop from weeds: NDRE if you have multispectral; with RGB only, a greenness index over chromaticity works surprisingly well because it is invariant to illumination.
• Comparing seasons: same index, same source, calibrated reflectance. If any of the three changes, it is not a comparison, it is a coincidence.
• Water stress: neither of them, really. The good signal combines surface temperature against vegetation with the area's rainfall balance, which is what satellite monitoring does; the vigor index arrives late, once the damage is done.
• Detecting harvest, pruning or plant loss: again not the absolute value, but the step in the time series against what the rest of the area did.