Prediction of Soil Organic Matter Using Visible–Near‐Infrared Spectroscopy Combined With Textural Hand‐Feel Determination

ABSTRACT

Timely and accurate estimation of soil organic matter (SOM) is vital for monitoring soil fertility and enabling precision management. This study investigates how combining soil texture data with visible and near‐infrared (Vis–NIR) spectroscopy can improve SOM prediction under both laboratory (dry) and in situ (moist) conditions. We also identify which soil properties cause the most significant differences between dry and moist spectral readings. Four modelling strategies were compared: (1) Pure spectra (using Vis–NIR data alone), (2) texture class and (3) texture merged (integrating quantitative clay, silt and sand content) and (4) texture dummy (integrating categorical texture classes determined by hand feel). We used the Cubist algorithm to assess the contribution of different soil properties, while partial least squares regression (PLSR) and random forest (RF) models were used for SOM prediction. Results show that soil texture was the most significant factor driving the differences between the moist and dry spectra. Texture was also the primary variable explaining the prediction errors (residuals) when using spectra alone. Critically, incorporating simple, field‐assessed texture data via the Texture Dummy strategy significantly improved SOM prediction accuracy compared to models relying on pure spectra, with the highest
R
2
increasing from 0.65 to 0.78 and from 0.79 to 0.86 and the root mean square error decreasing from 5.73 to 4.56 g kg
−1
and from 4.47 to 3.62 g kg
−1
for wet and dry samples, respectively. The Texture merged strategy provided a modest improvement (
R
2
increasing from 0.63 to 0.66 using in situ spectra), while the Texture class strategy showed the lowest performance, with slightly reduced
R
2
values for both dry and in situ spectra. The accuracy achieved using the Texture dummy strategy with in situ spectra (
R
2
 = 0.76, LCCC = 0.86) was comparable to that of the general model developed from laboratory spectra (
R
2
 = 0.79, LCCC = 0.88). Our findings demonstrate that soil texture is the primary factor driving the spectral differences between in situ and laboratory soil samples, except within the known water‐absorption wavelengths; incorporating the texture class determined by hand‐feel can effectively guide spectral techniques, enhancing the potential of vis–NIR spectroscopy in rapid estimation of SOM. However, further verification is needed using soils from a wider range of areas and larger spectral libraries, as well as other spectroscopic techniques across different wavelength ranges.