Moisture-robust soil organic carbon prediction from vis-NIR hyperspectral imagery using external parameter orthogonalization and transfer learning

Digital monitoring of soil organic carbon (SOC) is increasingly needed to support climate-smart agriculture, carbon accounting, and precision soil management. However, soil moisture substantially alters near-infrared (Vis-NIR) reflectance signatures, limiting the transfer of laboratory-developed SOC prediction models to field conditions and applications. This study evaluated a moisture-robust SOC prediction pipeline combining imaging Vis-NIR hyperspectral spectroscopy (500–998 nm) with transfer learning from the LUCAS topsoil spectral library, External Parameter Orthogonalization (EPO) for moisture correction, and soil texture (sand, silt, clay) as an auxiliary input. The pipeline was applied to 142 agricultural soil samples from different field trials in Germany, each imaged at four gravimetric moisture levels (0, 5, 10, and 15%). EPO reduced the moisture-induced spectral distance between paired dry and wet spectra of the same sample by 95.9 to 98.2% across moisture levels, while five principal components explained 99.7% of the moisture-induced variance. The EPO-corrected random forest model achieved R 2 = 0.82, 0.81, and 0.83 at 5, 10, and 15% moisture, respectively (Root-Mean-Square Error, RMSE = 0.120–0.124% SOC; Ratio of Performance to InterQuartile distance, RPIQ = 3.92–4.04), exceeding the accuracy obtained under dry conditions (R 2 = 0.71; RMSE = 0.157%; RPIQ = 3.09). Feature importance analysis revealed a shift from texture-dominated prediction under dry conditions (52% cumulative texture importance) to spectrum-dominated prediction under EPO correction (76% spectral importance), with key wavelengths concentrated in the green-yellow visible region (528–552 nm), associated with organic matter chromophore absorption. These findings demonstrate that accurate moisture-robust SOC prediction can be achieved using Vis-NIR wavelengths alone, providing an advanced methodological foundation for proximal soil sensing on autonomous field platforms operating under heterogeneous soil moisture conditions.