Tabular Foundation Models Excel in Soil Spectroscopy Predictions

Researchers have conducted a comprehensive study on using visible and near-infrared (vis-NIR) and mid-infrared (MIR) spectroscopy for predicting soil properties, a method known for its speed and cost-effectiveness. The challenge lies in accurately translating high-dimensional spectral data into reliable predictions, especially with machine learning. The study compared several regression models, including a tabular foundation model (TabPFN), a convolutional neural network (CNN), and traditional methods like Random Forest and Partial Least Squares Regression (PLSR), across 85 regression tasks from open benchmark datasets. The findings indicate that TabPFN consistently delivered superior performance across both field-scale and large-scale global soil spectral library tasks, even when applied directly to full spectra without explicit dimensionality reduction. While TabPFN alone surpassed classical baselines, further improvements were observed when combining it with features derived from Partial Least Squares (PLS) latent variables. This suggests that modern tabular foundation models and established chemometric techniques can complement each other effectively, offering robust guidance for selecting spectroscopic calibration models across various operational scales.