Prediction of Soil Colour using Vis-NIR Spectroscopy and Machine Learning Models

Abstract

Soil colour is a critical indicator of soil properties and conditions, influencing various agronomic and environmental factors. A total of 2216 surface soil samples (0-15 cm) were collected from the Kymore Plateau and Satpura Hill zone of Madhya Pradesh, using Global Positioning System (GPS) for precise location. Soil colour parameters were measured in the field using the Munsell soil colour chart, while chemical analysis was conducted in the laboratory following standard procedures. Additionally, spectra of the soil samples were recorded using a spectroradiometer under laboratory conditions. The results showed that the soil colour hues ranged from 10R, 10YR, 2.5Y, 2.5YR, 5Y, 5R, 5YR, to 7.5YR, Values and Chroma varied from 2 to 7 and 1 to 8, respectively. Correlation analysis revealed negative correlations between the RGB components and organic carbon, with r values of -0.114**, -0.071**, and -0.101* for R, G, and B, respectively. Polynomial models showed the best fit for the relationship between the value and chroma of the colour parameters and soil organic carbon (SOC), with equations Y = 0.086x² - 0.860x + 7.528 (R² = 0.982) and Y = 0.018x² - 0.249x + 6.126 (R² = 0.948), respectively. A linear relationship was observed between chroma and available phosphorus (P), with the equation Y = -0.873 + 13.92 (R² = 0.922). In addition, machine learning models, including PLSR, SVM, Random Forest, ANN, XGBoost, LightGBM, CatBoost, and ELM algorithms, were used to predict soil colour parameters. Among these, the Random Forest and XGBoost models demonstrated the best performance in predicting soil colour parameters (L*, a*, b*, R, G, and B), with model accuracies of 83.6%, 80.9%, 83.0%, 84.3%, 83.7%, and 83.4%, respectively. soil colour variation depicted in the maps generated using GIS can also serve as covariates for  mapping, offering comprehensive insights into the soil's properties.