Under the dual influences of high geological background and human activities, the soil of karst mining area is seriously polluted by heavy metals and is an important source area of heavy metal migration in the region. Rapid monitoring of heavy metal content can provide a basis for the study of regional heavy metal migration and transformation. In this study, the content of As and Sb in soil samples are determined simultaneously, and hyperspectral data are collected. By comparing nine hyperspectral data preprocessing methods, five optimal preprocessing methods are selected, and two hyperspectral inversion models of machine learning (partial least squares regression (GA_PLSR) and random forest (GA_RF) are established based on genetic algorithm to screen spectral feature bands. The results are as follows: (1) The multi-scattering correction+first-order differential (MSC+FD) pretreatment method is the optimal spectral pretreatment method for soil As and Sb; (2) The optimal inversion model of soil heavy metal As in karst mining area is partial least squares regression model, the R² (decision coefficient) of the training set and the test set of this model are greater than 0.9, and the RPD (residual predictive deviation) is greater than 3.0. The best inversion model of soil Sb is the random forest model, the R² of the training set and the test set are both greater than 0.9, and the RPD of the model is greater than 3.0; (3) The spectral characteristic bands of soil as from the karst mining area were distributed in the range of 446.1~468.4, 534.7~561.2, 685~728.3, 735.7~766.6, 817~839.7 and 852.2~1077 nm. The characteristic bands of soil heavy metal Sb are mainly 343.6~359.6, 671.5~686.5, 746~753.4, 921.2~998.6, 1024.3~1037.7 nm. The results can serve as a basis for rapidly obtaining the content of As and Sb in soil within karst mining areas.