Testing a National VisNIR Spectral Library with Texas Soils.

There is a growing interest in building proximal VisNIR spectral libraries for in situ and lab soil characterization. Conceptually, these libraries should contain both compositionally and spectrally diverse calibration samples that can represent a wide range of soil samples that may be encountered.  A diverse spectral library was previously created by Brown et al. It contains about 4200 soils from the NRCS National Characterization lab in Lincoln Nebraska. The soils were collected from all over the Unites States.  In this study, we will test the usefulness of this library in characterizing soils from different areas in Texas and primarily focus on clay, soil organic carbon and inorganic carbon content.  Three sets of test samples are used: (1) a set of ~500 floodplain soil samples collected from Quemado, Texas, (2) a set of ~250 soil samples from Central Texas, representing from various parent materials and land uses, (3) a set of ~ 2000 soil samples collected from all over Texas and archived by the Texas A&M University’s Soil Characterization Lab.

 

First, calibration models built on the entire library will be applied on the test samples.  Second, methods such as Mahalanobis Distance and unsupervised classification will be attempted to compare the spectral similarity between the library samples and the test samples.  Calibration models that are built on a subset of the library samples (which are spectrally similar to the test samples) will be also tested.  We postulate that (ad-hoc) subset models may outperform the full model due to the spectral similarity between the calibration and test samples.  Lastly, we also study the possibility of the expansion of this soil spectral library by including new samples that are spectrally dissimilar from all existing library samples.  Continuous expansion and refinement of the soil spectral libraries are necessary to improve their applicability and accuracy.