Predicting USCS soil texture classes utilizing soil spectra and deep learning

Soil texture identification is vital for various agricultural and engineering applications but generally involves rigorous laboratory work, especially for estimating USCS (Unified Soil Classification System) soil texture classes. Soil texture influences soil water storage capacity, soil fertility, compaction characteristics, and soil strength. Soil spectroscopy offers a reliable approach that is non-destructive, rapid, and cost-effective to estimate several soil properties including texture. For engineering applications, the USCS soil texture classes are preferred, but very few studies have focussed on estimating USCS soil texture using soil spectroscopy or remote sensing data in general.

Two large soil spectral libraries (SSLs), viz., Kellog Soil Spectral Library (KSSL) and Open-source Soil Spectral Library (OSSL), as well as three deep learning algorithms (VGG-16, ResNet-16, and Swin transformers), were used in this study to predict six USCS soil texture classes and three USCS soil texture groups. The USCS soil texture classes and groups were derived by grouping clay, sand, and silt fractions that are closely associated with the corresponding USCS soil texture classes.

The results indicate that the Swin transformer model performed the best with an accuracy of 67% for six USCS soil texture class predictions and 81% for three USCS soil texture group predictions. Cohen’s kappa value implies a moderate agreement (0.55) for soil texture class predictions and a substantial agreement (0.64) for soil texture group predictions.

The proposed methodology offers a novel approach for USCS soil texture class predictions utilizing SSLs and deep learning techniques.