To improve the inversion accuracy of Electrical Resistivity Tomography (ERT) and overcome the limitations of traditional linear methods, this paper proposes an im-proved Shuffled frog leaping algorithm (SFLA). Firstly, a balanced grouping strategy is designed to balance the contribution weights of each sub-group to the global optimum solution, thereby suppressing the local optimum traps caused by the dominance of high-quality groups. Secondly, an adaptive moving operator is constructed to dynam-ically adjust the search step size, enhancing the guidance effect of the optimal solution. Synthetic data tests of three typical geoelectric models (including 5% random noise) show that, compared to the least squares method (LS) and standard SFLA, the im-proved algorithm increases the accuracy of anomaly boundary recognition by ap-proximately 2.3 times and reduces the root mean square error by 57%. In the engi-neering validation at the Baota Mountain mining area in Jurong, the improved SFLA inversion clearly reveals the undulating bedrock morphology. At 55m along the survey line, the bedrock depth is 14.05m (ZK3 verification value 12.0m, error 17%), and at 96m, the depth is 6.9m (ZK2 verification value 6.7m, error 3.0%). The bedrock depth, which is deeper in the south and shallower in the north, is highly consistent with the terrain and drilling data (RMS = 1.053). This algorithm provides reliable technical support for the fine detection of ERT in complex geological structures.