Vis-SWIR spectral prediction model for soil organic matter with different grouping strategies

The effect of soil organic matter (SOM) on the spectral characteristics of soil differs among soil samples of different soil types with various mineral contents or mechanical compositions. SOM prediction models with visible and near-infrared and shortwave infrared (Vis-SWIR) spectral variables for grouping soil samples are more accurate than global modeling methods. However, the optimal grouping method must be explored. We measured the Vis-SWIR (400–2500 nm) spectral reflectance of 274 soil samples from the Songnen Plain and applied different grouping strategies, including traditional soil grouping, spectral clustering grouping and decision tree grouping considering both traditional grouping information and spectral information. The original reflectance, continuum removal, and first-derivative reflectance were analyzed with competitive adaptive reweighted sampling (CARS), and the results were used as inputs to establish a random forest model for SOM prediction. The following results were obtained: (1) the decision tree grouping method is the optimal grouping strategy for the SOM prediction model; (2) among different inputs, first-derivative reflectance is optimal; and (3) CARS is an effective way to reduce the number of inputs and improve the prediction accuracy of SOM. This study provides an effective soil grouping methodology for SOM estimation at the soil great group level.