The long-offset transient electromagnetic method (LOTEM) offers a large depth of investigation and high sensitivity to subsurface resistivity variations, making it valuable for deep resource exploration, oil and gas prospecting, and engineering investigations. As exploration targets move to greater depths, increasingly undulating surface and subsurface interfaces and more complex structural settings make LOTEM responses more difficult to relate accurately to the true subsurface resistivity distribution. Under such conditions, 3D LOTEM inversion faces greater challenges in background-structure representation, target-anomaly recovery, and result reliability. To address the inadequate representation of background resistivity structures by conventional initial models under complex backgrounds, and the resulting impact on inversion reliability, this study proposes an improved 3D LOTEM inversion strategy based on initial-model optimization. Prior information capable of reflecting the characteristics of background resistivity distribution and structural relationships is used to guide the construction of the initial model, thereby providing a more suitable model basis for the early stage of inversion. In synthetic examples, a background initial model is first constructed from station-wise 1D inversion results, and a horizon-constrained initial model is then further established by incorporating horizon information. The results show that the proposed strategy can significantly improve the recovery of both background structures and target anomalies. Finally, the strategy is applied to LOTEM field data for deep karst investigation in a shale gas exploration area in southwestern China, where a deep low-resistivity anomaly is identified near the target interval. Combined with the regional geological setting, karst development patterns, and well-seismic constraints, the anomaly is inferred to represent a deep karst-developed zone and provides geophysical support for drilling planning. Overall, the proposed strategy improves the representation of background resistivity structures, reduces the influence of background structures on target recovery, and enhances the recovery performance and reliability of inversion results, thereby providing technical support for methodological improvement of 3D LOTEM inversion and for deep-target detection in complex resistivity settings.