Harmonizing Global MIR Soil Libraries with Deep Learning for Improved SOM Modelling

Mid-infrared (MIR) spectroscopy coupled with deep learning offers a scalable route to fill gaps in global soil property databases. Critical for modelling soil organic matter (SOM) dynamics across agricultural and natural ecosystems. This study utilized a global dataset of approximately 57,000 soil samples with MIR spectra. Laboratory measurements of carbon (C) and organic matter (OM) were available for 51,079 and 5,750 samples, respectively. Spectral preprocessing comprised normalization, outlier detection, inter-laboratory harmonization, and physically constrained prediction ranges. For each target property, convolutional neural networks (CNNs) were trained on available paired spectral–laboratory data and then used to impute missing values, producing a hybrid dataset of measured and predicted values with accompanying prediction-error estimates. Model performance was strong for C (R² = 0.96, RMSE = 15.41 g kg⁻¹, MAE = 6.62 g kg⁻¹, RPIQ = 1.38; prediction SD = 40.08 g kg⁻¹; 5,708 imputed samples) and moderate for OM (R² = 0.79, RMSE = 19.15 g kg⁻¹, MAE = 11.76 g kg⁻¹, RPIQ = 1.05; prediction SD = 10.06 g kg⁻¹; 51,037 imputed samples). The resulting hybrid dataset increases spatial and depth-wise coverage for SOM-related variables, supporting more robust parameterization, calibration, and validation of SOM models across ecosystems. Despite lab and spectral limitations, MIR and deep learning improve global soil data by providing soil property estimates and uncertainty information, supporting SOM studies and applications in agriculture, conservation, and carbon accounting.