Abstract: Agricultural environments were generally characterized by weak geometric saliency, repetitive vegetation structures, and frequent dynamic disturbances, which posed significant challenges to conventional simultaneous localization and mapping systems that primarily relied on geometric feature matching. Under such conditions, unstable data association and unreliable residual estimation were often produced, leading to degraded pose accuracy and accumulated localization drift. Therefore, this study was conducted with the objective of improving localization accuracy and robustness for agricultural robots by introducing semantic information into a tightly coupled LiDAR–IMU SLAM framework, such that feature association and optimization could be guided by category-dependent reliability rather than geometry alone.In the proposed framework, LiDAR point clouds and inertial measurement unit data were jointly utilized. Raw LiDAR scans were first processed by motion distortion correction and outlier removal to mitigate the effects of sensor motion and noise. Subsequently, a semantic segmentation network (RangeNet++) was employed to assign semantic labels to each point, generating semantically annotated point clouds. To balance computational efficiency and semantic preservation, a semantic-aware downsampling strategy was adopted, in which different semantic categories were sampled with different densities according to their spatial distribution and relevance to pose estimation. During LiDAR odometry estimation, curvature-based edge and planar features were extracted and associated between consecutive frames. Semantic information was explicitly embedded into the optimization process by assigning category-specific weights to residual terms. Moreover, a dynamic semantic weighting mechanism based on historical residual statistics was introduced. Residuals accumulated within a sliding window were statistically analyzed to adaptively update semantic weights, so that categories producing larger residuals were down-weighted, while categories exhibiting stable matching behavior were reinforced. On the back end, LiDAR odometry constraints and IMU pre-integration constraints were jointly formulated within a factor-graph framework. Keyframes were selected according to translational and rotational motion thresholds, and all constraints were optimized through nonlinear least-squares optimization.Extensive experiments were carried out on both public benchmark datasets and self-collected agricultural datasets to evaluate the effectiveness of the proposed method. On the KITTI benchmark, the method, referred to as LID-SLAM, was evaluated on Sequences 05, 07, and 10 and compared with LIS-SLAM, LIO-SAM, and LeGO-LOAM, with loop closure disabled for fair comparison. On Sequence 05, which contained frequent turns, occlusions, and dynamic traffic participants, the influence of unreliable semantic categories was effectively suppressed. As a result, a mean absolute trajectory error of 2.48 m and a root mean square error (RMSE) of 2.88 m were achieved, outperforming LIO-SAM (mean 2.88 m, RMSE 3.41 m) and significantly surpassing LeGO-LOAM (mean 9.36 m, RMSE 10.1 m). On Sequence 07, representing a relatively static and open environment, stable performance was maintained, with a mean error of 0.78 m and an RMSE of 0.82 m, which were comparable to those of LIO-SAM and markedly better than those of LIS-SLAM and LeGO-LOAM. On the more challenging Sequence 10, which involved complex dynamics and higher motion speeds, peak and long-tail errors were substantially reduced. The maximum error was lowered to 3.77 m, and the RMSE was reduced to 1.39 m, compared with 2.22 m for LIO-SAM and 7.92 m for LeGO-LOAM.Ablation experiments were further conducted to assess the contribution of the dynamic semantic weighting mechanism. When the dynamic weighting module was removed, larger local deviations and increased error fluctuations were observed, although the overall trajectory shape was preserved. In contrast, introducing dynamic semantic weights led to consistent improvements, reducing the maximum error by approximately 19.23% and decreasing RMSE by up to 14.9% on representative sequences. On a self-collected agricultural dataset featuring dense vegetation, mixed terrain, and narrow roads, the proposed method consistently produced trajectories that closely aligned with the ground truth. Even after long straight motions followed by sharp turns, no obvious drift or trajectory break was observed. Quantitatively, the RMSE was reduced to 0.084 m, compared with 0.102 m for LIO-SAM and 0.205 m for A-LOAM, corresponding to relative reductions of 17.65% and 59.02%, respectively.In conclusion, a semantic-assisted, LiDAR–IMU tightly coupled SLAM framework was presented in this study. By combining semantic-aware downsampling, category-weighted residual modeling, and residual-feedback-driven dynamic semantic weight adaptation within a factor-graph optimization backend, data association reliability was improved and error accumulation was effectively suppressed in weak-geometry and dynamic environments. Experimental results on benchmark and real-world agricultural datasets demonstrated that the proposed approach significantly enhanced localization accuracy and robustness, providing a practical and effective solution for high-precision autonomous navigation of agricultural robots.