Visible and Infrared Lab Spectroscopy for Soil Texture Classification: Analysis of Entire Spectra V/s Reduced Spectra

Soil, directly or indirectly, impacts seven of the Sustainable Development Goals set by the United Nations. Soil texture is a fundamental physical property that controls other properties ranging from soil stability, erodibility, compaction to water holding capacity, nutrient availability and fertility. While Visible Near-Infrared and Shortwave-Infrared (VNIR/SWIR) laboratory spectroscopy has been already used for soil texture classification, Mid Infrared (MIR) region and unison of the VNIR/SWIR and MIR region warrant quantification. Thus, the objectives of this work were: 1) to investigate the critical spectral regions for soil texture classification, 2) to compare the use of entire spectral bands v/s subset of spectral bands obtained via the feature selection method, and 3) to quantify the degree of misclassification in the neighbouring classes for a given textural class. This study utilized an ICRAF–ISRIC global soil spectral library consisting of 3643 VNIR/SWIR, MIR soil spectra and the USDA soil textural classes.Two classifiers were used for soil texture classification using the three databases (VNIR/SWIR, MIR, VNIR+MIR) with entire spectral bands and three databases with only the Partial Information Correlation (PIC) selected bands (VNIR/SWIR_PIC, MIR_PIC, VNIR+MIR_PIC). The mean confusion matrix, overall accuracy and kappa were calculated to evaluate classifiers' performances over 100 iterations. Two additional measures were proposed for partitioning inaccuracies between neighbouring and far classes: Correct-Neighbouring-Far classes distribution matrix and added neighbourhood accuracy. This work highlighted that: 1) the combined VNIR/SWIR+MIR region provided the best texture classification, followed by MIR and then VNIR/SWIR region, 2) the use of PIC selected bands provided lower classification performances when compared to using all bands but a massive reduction in the number of bands allowing to reduce model complexity, 3) the misclassifications were predominantly in the neighbouring classes rather than far classes, for a given texture class. Finally, the textural classes with poor classification performance had low areal and sample representation. These insights may launch a number of soil texture classification studies worldwide based on global soil spectral databases, further helping map soils and the associated ecosystem services it provides.