This study used field-collected soil samples as test subjects to conduct experiments including soil organic matter (SOM) content determination, hyperspectral data acquisition, and preprocessing. Six spectral transformations were applied to the preprocessed spectral data: absorption depth (Depth), first derivative of log-reflectance (FD-lgR), second derivative of log-reflectance (SD-lgR), second derivative of reflectance (SD-R), second derivative of reciprocal reflectance (SD-1/R), and second derivative of reciprocal log-reflectance (SD-1/lgR). Partial least squares regression (PLSR) models for SOM estimation were established under different spectral transformation forms to analyze the correlation between spectral transformations and SOM content, as well as their impact on model accuracy. The results showed that all 6 transformations exhibited bands significantly correlated with SOM content, with FD-lgR having the highest number of significantly correlated bands (71). The FD-lgR model achieved a determination coefficient (R²) of 0.995, a root mean square error of calibration (RMSEC) of 0.063, a cross-validation R² of 0.775, and a relative percent difference (RPD) of 2.681, all of which were among the highest values across all transformations. The scatter plot of predicted versus measured values indicated that the FD-lgR model’s estimates were close to the actual values, with an R² of 0.872. Overall, the regression model based on FD-lgR demonstrated high accuracy and good stability. These findings provide a reference for subsequent hyperspectral data preprocessing and estimation model construction for soil organic matter.