Open AccessEstimating coefficient of linear extensibility using Vis–NIR reflectance spectral data: Comparison of model validation approaches
Author(s) -
Rehman Hafeez Ur,
Arthur Emmanuel,
Tall Andrej,
Knadel Maria
Publication year - 2020
Publication title -
vadose zone journal
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.036
H-Index - 81
ISSN - 1539-1663
DOI - 10.1002/vzj2.20057
Subject(s) - calibration , coefficient of determination , correlation coefficient , mean squared error , statistics , predictability , linear regression , mixed model , feature selection , near infrared spectroscopy , selection (genetic algorithm) , model selection , mathematics , remote sensing , computer science , artificial intelligence , geology , physics , quantum mechanics
The coefficient of linear extensibility (COLE) is used to classify soils according to their swell–shrink potential, and its estimation is crucial for engineering and agronomic applications. The aims of the study were (a) to develop a visible–near‐infrared spectroscopy (Vis–NIRS, 400–2,500 nm) calibration model to estimate COLE, (b) to compare two model validation approaches (mixed data and country‐wise), and (c) to test if a variable selection method improves the estimation accuracy of the calibration models. For this purpose, partial least square regression (PLSR) was used on the spectra of 53 soil samples from Slovakia and 24 samples from the United States. First, a calibration model based on 70% of the entire dataset (including samples from both locations) was developed and validated with the remaining 30% (mixed data approach). Second, a calibration model based on the Slovakian samples was validated with the U.S. samples (country‐wise approach). Higher predictability for COLE with standardized root mean square error (SMRSE) of 0.099 was obtained for the mixed data approach than for the country‐wise validation with SRMSE of 0.279. Furthermore, using interval PLSR (iPLSR) as a variable selection method did not improve the estimation accuracy of the mixed data approach (SRMSE of 0.099), and rather resulted in a twofold increase in SRMSE (0.560) for the country‐wise validation approach. Overall, the good estimation of COLE from Vis–NIRS was attributed to the high correlation of COLE with clay content and spectrally active clay minerals.