In order to improve spectral stability and quantitative inversion accuracy of heavy metal in soil by laser induced breakdown spectroscopy (LIBS), hemispherical space constraint structure has been designed, the relative standard deviation (RSD) and relative error of prediction (REP) of LIBS of heavy metals are studied. Multiple average and internal standard method are used to process spectral data, the RSD of spectrum has been reduced 1.59% and 3.65% respectively by increasing the number of pulses and internal standard method, and the spectral stability has been improved. The REP of sample concentration is comparatively analyzed under univariate and patial least squares (PLS) method, REP of univariate method of four characteristic spectral lines CrI: 357.869 nm, CrI: 359.349 nm, CrI: 425.435 nm and CrI: 427.480 nm are 9.42%、8.43%、 6.16% and 7.79% respectively, the REP of PLS under four variable combination are 3.70%、4.10% 、4.83% and 4.67% respectively. The REP of sample is reduced 5.72% by PLS method, and the accuracy of quantitative inversion is improved. The above results provide data and method support for improving analytical ability of LIBS of heavy metals in soil.