The NDVI is used quite a lot to estimate biophysical variables like LAI or fAPAR and others. How can one estimate the uncertainty of these variables without knowing the absolute error on the NDVI? How does one proceed in the error estimation of the NDVI knowing the standard deviations of RED and NIR reflectance?
I've seen that ground-based observations retrieved from "spectrometer device" can be beneficial in order to have the most accurate NDVI. If you have, you can use them as benchmarks and calculate absolute error. Besides, it can pave the way for atmospheric effects mitigation and sensor calibration. Yet, if we are willing to use just standard deviation of two signals, I think, we can only get standard deviation of NDVI using propagation of errors:
(sigma NDVI)^2=(dNDVI/dRED)^2*(sigma RED)^2+(dNDVI/dIR)^2*(sigma IR)^2
I've seen that ground-based observations retrieved from "spectrometer device" can be beneficial in order to have the most accurate NDVI. If you have, you can use them as benchmarks and calculate absolute error. Besides, it can pave the way for atmospheric effects mitigation and sensor calibration. Yet, if we are willing to use just standard deviation of two signals, I think, we can only get standard deviation of NDVI using propagation of errors:
(sigma NDVI)^2=(dNDVI/dRED)^2*(sigma RED)^2+(dNDVI/dIR)^2*(sigma IR)^2
Spectrometer will best give a point observation how well it will correlate with entire pixel espically in desert region where the vegetation is scanty?
NDVI suffers from atmospheric effects. That's why we also use blue band in order to make NDVI more atmospherically resistant. When you have spectrometer (an accurate signature) you can retrieve the most precise indices from it (no matter what type of class is there). It is rigid to say we can get absolute error of something, since spectrometers also have uncertainty, but I think it is the most accurate. Now you can use "Empirical line calibration" in order to mitigate atmospheric effects for whole images; besides, we are able to compare vegetation indices with spectral mixture analysis. But about other pixels, some of the ground-based observations should be used as benchmarks and some of them for check points in order to be sure that pixels existing among them have been precisely calibrated. Although it is somewhat an easy and superficial method, I also have a thesis that someone used spectrometer in order to model the atmosphere and sensor itself using physical approaches (instead of empirical).
Although I am not aware what Dr. Veroustaete's research objective is, it is typical to use a relationship between LAI and vegetation indices. Firstly you should mitigate atmospheric effects (using 6S or ModTRAN) and then, compare the mean, standard deviation and correlation of coefficients in this model.
There are two useful research papers:
http://www.sciencedirect.com/science/article/pii/S0034425711002926
http://www.sciencedirect.com/science/article/pii/S0034425712001630
You should take the fact that NDVI is used to get time series into account. The standard deviation of the red and nir bands will only give you an insight on the absolute error around each NDVI measurement (and it will probably be overestimated, since those standard deviation may include relevant information, and not only random errors).
But NDVI series are usually filtered, and fitted.
If you're interested, I shared an article which deals with this issue (for phenology detection, not LAI estimation).
The best, and maybe the only way to properly assess the error on faPAR and LAI estimations is to compare them to ground truth.
Thanks Gabriel,
I am very interested to read your artciles on the topic you mentioned.
Thanks beforehand,
Frank
- Badges
- Science topic
- Similar topics
- Geoscience
- Remote Sensing