To address the issue of low efficiency in extracting numerical values from figures, based on deep learning, a technical framework to extract numerical values from figures in soil literatures was proposed. Firstly, the common figure elements and their symbols were sorted out, and some figures were collected and manually labelled to form a training dataset. Secondly, using YOLO v8 base model, which uses the global image to detect multiple targets through one-time process, an optimized model suitable for the detection of figure elements in soil literatures was trained through several rounds of training. Thirdly, to convert the identified figure elements to real values, an algorithm was designed to automatically calculate the numerical values in 2D scatter and histogram figures. Using figures that were not involved in the training, the results showed this technique could effectively extract the figure elements and the numerical values were in high agreement with the manually extracted values (the linear regression coefficient of determination R² > 0.99). Therefore, the technical framework proposed in this study has strong feasibility, which provides a new approach for the efficient use of figure data in soil literatures.