The demand for cost-effective, high-throughput soil analysis is growing and can be addressed using diffuse reflectance spectroscopy in the visible–near-infrared (Vis-NIR) and mid-infrared (MIR) regions. This study compares the predictive performance of raw and pre-processed (baseline correction and standard normal variate) spectra from both MIR and Vis-NIR for estimating 11 soil properties: organic carbon (OC), total carbon (TC), total nitrogen (TN), cation exchange capacity (CEC), clay, sand, total solids (TS), pH, potassium (K), bulk density (BD), and nitrate-nitrogen (NO₃⁻-N). A dataset of 8,304 samples common to both spectral domains was selected from the USDA-NRCS Kellogg Soil Survey Laboratory library. The MIR spectra (2,500–16,260 nm) were collected using a Bruker Vertex 70 FTIR with HTS-XT, and Vis-NIR spectra (350–2,500 nm) using an ASD LabSpec. Eight machine learning models were evaluated. For MIR, spectral pre-processing improved prediction performance for all soil properties. The largest R² gains were observed for K (0.59 → 0.74), NO₃⁻-N (0.56 → 0.69), BD (0.53 → 0.59), and pH (0.84 → 0.87). Already strong models for OC, TC, TN, CEC, TS, and clay show further improvement (e.g., TN: 0.93 → 0.95; OC: 0.98 → 0.99). For Vis-NIR, preprocessing yielded more significant improvements. R² increased notably for clay (0.64 → 0.74), sand (0.50 → 0.67), silt (0.43 → 0.63), K (0.38 → 0.56), and pH (0.68 → 0.73). OC and TC maintained high predictability (R² > 0.90), while BD and NO₃⁻-N remained challenging (R² ≤ 0.56). Overall, MIR outperformed Vis-NIR across all properties, and preprocessing notably enhanced model performance, for both spectral regions. ANN and CatBoost emerged as the most robust algorithms across both spectral regions and pre-processing conditions. These findings support the strategic use of MIR spectroscopy and pre-processing techniques to improve the reliability of soil property estimation in large-scale applications.