The mid-infrared photoacoustic spectra of CaCO3 was determined and characterized, and multi-calibration methods of principal component regression (PCA), partial least squares regression (PLSR), and GRNN artificial neural network were applied to quantitative analysis of soil carbonate. The results showed that abundant absorptions were found in the mid-infrared photoacoustic spectra of CaCO3, especially the very strong band at the wavenumber of 1 450 cm-1, in which there was few interferences, and could be used as spectral indicator of soil carbonate; the calibration results were good or excellent with the three chemometric methods, in which PLSR and GRNN modeling were excellent with a R2 more than 0.9, and PCA modeling was good with a R2 of 0.847; the validation results showed that PLSR and PCA modeling were excellent with higher R2 values (>0.9), and GRNN was also very satisfied with a R2 of 0.882. Totally, PLSR modeling was the best with RPD values more than 3.0, indicating its strong potential in the prediction of soil carbonate.