When Alex [Goetz] asked me to speak at this symposium, I knew I wanted to present something special instead of doing an overview. I decided to look at variation between spectrometers to assess stability and variability of soil reflectance measurements with different ASD instruments under different conditions in the lab. [Information on different types of portable ASD Fieldspec spectrometers is at: http://www.asdi.com/products/instrumentation/portable]
Soil spectroscopy is well studied and used in many applications. Dozens of soil attributes can be estimated from NIR and reflectance data, but people use different kinds of data and methods making this process less robust. In order to reduce variability, we need a single spectrometer to create and validate the model.
Every measurement needs an internal standard, especially those used in the field of scientific study. There are many soil spectra libraries. The World Spectral Group database, as an example, contains 7000 samples, five attributes, 60 instruments, 80 users and 40 protocols. To add validity to soil spectroscopy, we must standardize the process and be able to take reference material into the field so we can reduce variability.
There are a few problems associated with this, though: Water is hard to control, and soil moisture can affect the spectra markedly. Different instruments in source and destination labs can also cause variations in the soil spectra. Other causes of variability include type of device, sample preparation, measurement protocol, humidity, optics and more. We must assume that operator error affects the standard, too. Protocol calls for reference measurement at the beginning and end of each sampling run to account for this.
Standards need testing. We were very careful with the lab equipment, including testing chambers and stabilization. The objectives for this study are to compare three ASD spectrometers under controlled conditions and to analyze instrument protocols and conditions for each.
We measured various soils with several ASD devices. In the first experiment, we used all three ASD devices in the same lab at same time. This produced generally low variation despite the fact that our aggregated soil samples contained various grain sizes.
In a second experiment, we sent the ASD instruments to three different labs in Israel to re-measure the samples. Each lab’s humidity level differed, making for higher variation than in the first experiment. The handheld probe was an especially large source of variation.
Protocol for this experiment was an 11-step method. We found some variations in the second experiment, even in the average environment. The largest of these variations occurred in the least humid lab, because the soil samples tended to dry over the course of the experiment. Good protocol minimizes the need to correct light sources and geometry, but variation is still a factor.
So, how do we correct for variation? Analytical chemistry uses certified internal standards. We can borrow that model to correct the spectra. The solution must be inexpensive, available around the world, spectrally stable, materially inert, stable chemically and similar in particle size to soil. Industry standards include formica, ground glass and bleached sand. Even with these correction factors in place, we found some variation in the instruments.
We set out to design a correction via a normalized additive factor. We found bleached sand to be the most favorable standard. We then tested effects of using different standards on false alarms for classification of soils. Formica was the worst of these standards. The effects of these additives were not visible in correlations of soil indices, but we found sand to work best in our calculations for finding soil content.
Though variation sources are numerous, including instrument, operator and protocol, we found that measurement protocol and external conditions dominate variation in this environment. The instrument accounts for less variation (up to 10% variation across spectrometers). As variation is not an issue of ratio indices, but is an issue of partial least squares, internal reference standards help validate the model.
We propose that this method enables the variable spectra of a given soil sample to be translated into a common denominator. Thus, it should facilitate the exchange of spectral information among scientists worldwide and allow creation of a robust soil database for diverse applications.