Abstract: LiDAR-inertial simultaneous localization and mapping (LiDAR-inertial SLAM) is a fundamental technology for autonomous navigation of mobile robots, which achieves accurate pose estimation and environmental mapping by fusing LiDAR measurements and inertial measurement unit (IMU) information. Existing LiDAR-inertial SLAM methods usually adopt fixed keyframe insertion strategies and fixed voxel grid resolutions for front-end state estimation and local map maintenance. These methods can achieve satisfactory performance in environments with stable feature distributions. However, when the mobile platform moves through scenes with different structural complexity and observation conditions, such as open areas, regular structured spaces, narrow corridors, and transition regions, fixed parameters are difficult to adapt to changing feature constraints. This may cause redundant or insufficient keyframe constraints, mismatched map representation scales, increased computational burden, and decreased pose estimation accuracy. To address these problems, this study proposes an adaptive parameter adjustment method for LiDAR-inertial SLAM based on multi-scale environmental perception. Firstly, an omnidirectional curvature measurement model based on multi-level spatio-temporal neighborhoods is constructed to extract geometric features from panoramic LiDAR point clouds. Combined with a sector-based spatial geometric balancing mechanism, the average observation depth, the number of effective feature points, and the average reflection intensity are statistically calculated in each sector to construct a directionally balanced environmental observation vector. Compared with conventional single-feature environmental description methods, the proposed observation vector provides a more comprehensive representation of environmental characteristics under different structural conditions. Secondly, Haar wavelet decomposition is introduced to analyze the environmental observation sequence in a sliding window. The low-frequency component is used to extract the evolution trend of the environment, while the high-frequency component is used to characterize local disturbances and abrupt changes. Based on this multi-scale representation, a trend-disturbance dual-channel environmental state discrimination mechanism is designed. The trend channel identifies open, normal, and narrow environmental states through low-frequency information, while the disturbance channel detects sudden scene changes by calculating high-frequency disturbance energy. Therefore, the proposed method can continuously track environmental states and avoid unstable parameter switching caused by short-term observation noise. Finally, a mapping relationship between environmental states and the keyframe insertion distance threshold as well as voxel grid resolution is established. Based on the identified environmental state, the corresponding target parameters are generated and updated through a parameter smoothing strategy to achieve adaptive parameter adjustment. The adjusted parameters are synchronously applied to keyframe management and local map construction, thereby improving the adaptability of the LiDAR-inertial SLAM system under complex environmental changes. Experimental results on the KITTI and M2DGR public datasets show that, compared with LIO-SAM, FAST-LIO2, and Adaptive-LIO, the proposed method reduces the absolute trajectory error (ATE-RMSE) by approximately 38.35% on KITTI Seq.05 and 31.94% on Seq.07, while maintaining a lower standard deviation (STD) of trajectory error on most sequences. These results demonstrate that the proposed method effectively improves the global map consistency and pose estimation accuracy of the LiDAR-inertial SLAM system under complex environments.