Abstract: To solve shortcomings of sampling methods for constructing training dataset of surrogate models in current inverse analysis, this study introduces the CV-Voronoi adaptive sequential sampling method for generating training samples. This sampling method does not require the determination of the sample size in advance and can adaptively add new samples based on the information of the existing sample points. Moreover, the sparrow search algorithm (SSA) is employed to optimize the twin support vector regression (TSVR) model, and the SSA-TSVR surrogate model is then generated. Based on the adaptive sequential sampling method and SSA-TSVR model, using SSA as the optimization algorithm, a new inversion technique is proposed. Using the shear strength parameter inversion of the Baishuihe landslide slip zone soil as an example, the new inversion method is validated through engineering application. The effects of different sample generation methods (adaptive sequential sampling, orthogonal design, and uniform design) and surrogate models (SVR, TSVR, SSA-SVR, and SSA-TSVR) on the inversion results are compared. The results show that the adaptive sequential sampling method has a significant advantage, reducing the inversion error by more than half. This method not only significantly improves the inversion performance, but also achieves higher accuracy with fewer samples. The SSA-TSVR surrogate model offers higher inversion accuracy and computational speed, providing a new approach for inversion analysis of geotechnical engineering mechanical parameters.