Abstract: Level scaling of expansive soil is of significant importance in practical engineering. In this study, a support vector regression (SVR)-based predicting model of expansive soil level was proposed. Using the laboratory test dataset of expansive soil in Kenya, two training models with different predictors were built. Seven parameters, including liquid limit, plastic limit, plastic index, particle percentage of three different particle sizes (< 0.075、0.075～0.25、0.25～0.5), and soil type, were used in model I, while five parameters, including liquid limit, plastic limit, plastic index, percentage of particles with size < 0.075 and soil type, were used in model II. For each model, four kernels namely, Linear, Polynomial, RBF and Sigmoid, were used. The result had shown that all training models become stable when the sampling number reached 1000.The results show that with the increase of the number of randomly selected training samples, when the number of predictions reaches 1000, the prediction accuracy obtained by the SVR models with 4 different kernel functions basically stabilizes. The prediction effect is better when the RBF function and Linear function are used, followed by the Sigmoid kernel function. The prediction accuracy of the above three is more than 70% in scheme one, and prediction accuracy of the above three is more than 70% in scheme two. When Linear function, Sigmoid function, and RBF function are used as kernel function models to predict 44 groups of unknown expansive scale soil samples, the number of identical prediction results accounts for 73% in scheme one. The prediction scale of the remaining sets of soil samples is the same or adjacent, there is no "crossing" phenomenon. The number of identical prediction results accounts for 73% in scheme two, and there is no "crossing" phenomenon. The results of the study can provide a basis for the prediction and treatment of the scale of the expansion soil in the construction of Kenya and other regions.