Comparative Assessment of NIR, MIR Transmission (Cuvette and Pearl Cell) and MIR-ATR Spectroscopy Coupled with LASSO–OLS Regression for Fatty Acid Profiling of Olive and Sunflower Oils

Fatty acid composition determines the nutritional value, authenticity and frying stability of edible oils, but its reference determination by gas chromatography is slow and reagent-intensive. Near-infrared (NIR) spectroscopy and three mid-infrared (MIR) sampling configurations—transmission cuvette, Pearl transmission cell and attenuated total reflectance (ATR)—were compared for predicting thirteen fatty acids methyl esters (FAMEs) in 142 oil samples from 14 olive and sunflower oil categories, including hydroxytyrosol-supplemented and deep-fried oils. Sparse LASSO–OLS calibrations were evaluated by fully nested leave-one-out (LOO) and leave-one-variety-out (LOVO) cross-validation (CV) and benchmarked against partial least squares (PLS) and ridge regression. Linoleic acid was predicted on all platforms with R2CV = 0.997–0.998 (RPD 17.1–21.4; LOVO R2CV ≥ 0.986) using 21–28 wavelengths, and models calibrated without fried oils predicted fried ones accurately. Oleic acid reached R2CV = 0.884–0.942, but 0.52–0.76 within olive oils and 0.55–0.72 under LOVO. Palmitic and palmitoleic acids reached screening-level accuracy that did not extend to new categories. Platform differences were small for the main fatty acids and are confounded with spectral range and path length. LASSO–OLS performed close to full-spectrum models with 2–6% of the variables, supporting band-selective sensors for authenticity screening and at-line monitoring of frying oils.