Temperature-dependent IR (TIR) spectroscopy is highly informative for complex samples. This approach increases the informativeness of soil IR spectra, which are characterized by low-intensity characteristic bands, especially those of soil organic matter (SOM). Cluster analysis with machine learning (ML) was used with attenuated total-reflection TIR. Dimensionality reduction algorithms are PCA, NMF, t-SNE, and LDA. Agricultural land-use chernozems (native, fallow, cropland, and shelterbelt) were subjected to granulometric fractionation to obtain a broad range of soil particles and microaggregates. Silt (size, <2 µm), dust (2–5 µm), cutoff (<20 and <50 µm), and narrow (20–30, 40–50, 50–100, and 100–200 µm) fractions were selected as characteristic. Different thermal behavior of bands assigned to SOM and inorganic matrix was found. Among them are hydrogen-bond region (4000–3200 cm^–1), aromatic/aliphatic fragments (3100–2800 cm^–1), carboxylic acids, carboxylates, and other SOM functional groups (1800–1200 cm^–1), and bands associated with phytoliths and quartz (850–150 cm^–1). While individual IR spectra do not provide enough information for differentiation due to broad and unresolved bands, changes in band frequencies and integral intensities in TIR (heatmap analysis) separates land-use samples based on the vegetation and agrogenic loads. Main contributions are changes in biogenic quartz and phytoliths. However, direct TIR analysis cannot distinguish fine fractions (dust, silt, and <20 µm). To the contrary, ML-TIR-based approaches, especially t-SNE, provided the separation of fine-size samples with increased contributions from SOM due to dimensionality reduction, which cannot be employed in direct approaches.