Objective The aim was to explore the potential of mid infrared spectroscopy in predicting soil organic carbon (SOC) content in farmland in Inner Mongolia region.

Methods This study focused on farmland soil in Inner Mongolia, China. Soil samples (411) were collected from the main farmland distribution areas in eastern, central and western Inner Mongolia as test samples. Based on different preprocessing combinations, SOC prediction models of partial least squares regression (PLSR) and support vector machine regression (SVR) were established to compare the accuracy of mid infrared spectroscopy in predicting OSC in the region.

Results ①From the overall prediction performance, among the different preprocessing methods corresponding to PLSR, the best prediction accuracy was achieved through normalization (R²=0.8360, RMSEP=1.7928 g kg⁻¹, RPD=2.4816); while the best preprocessing combination for SVR was multivariate scattering correction (MSC) + MA smoothing + centralization (R²=0.7557, RMSEP=2.1881 g kg⁻¹, RPD=2.0332). ② From the perspective of prediction performance in different regions, both modeling methods showed the best prediction effect on SOC in eastern farmland (SVR is better than PLSR), followed by the central region, and the worst in western farmland (PLSR is better than SVR), mainly due to differences in soil type and SOC content. In addition, We found that the SVR was more suitable for predicting SOC in high SOC farmland in the east, while PLSR had a more accurate prediction effect on SOC in western, central farmland, and overall soil.

Conclusion The farmland soil type, the difference of SOC content and the selection of pretreatment methods have a great impact on the prediction effect of mid infrared spectrum. Normalization PLSR quantitative prediction model based on mid infrared spectrum technology has a good prediction effect on regional farmland SOC (R²>0.80), which can provide important theoretical support for the development of precision agriculture in this region.