Precise estimation of shrub above-ground biomass (AGB) in arid regions is crucial for carbon cycle research and ecosystem assessment. Unmanned aerial vehicle (UAV) -borne light detection and ranging (LiDAR) has become a key tool for quantifying three-dimensional vegetation structure and estimating AGB. However, the short stature of arid zone vegetation, combined with sparse and low-quality point clouds acquired by UAV, limits high-accuracy shrub AGB estimation. To address this issue, this study selected Caragana korshinskii, a typical psammophytic shrub in Ordos City, as the research object. By integrating UAV-based multispectral and LiDAR data, a biomass estimation method based on a novel Shrub Structure Index (SSI) was proposed. The SSI workflow reconstructs the three-dimensional shrub structure under sparse point cloud conditions and improves AGB estimation accuracy. This workflow comprises Object-based image analysis (OBIA) classification for individual shrub extraction, Delaunay linear up-sampling, voxel-based partitioning, and dynamic stratification by height percentiles. Experimental results demonstrate that: (1) The individual shrub extraction method utilizing the large-scale mean shift (LSMS) segmentation algorithm and support vector machine (SVM) classification achieved a total quadrat segmentation accuracy of over 90.61 %, an overall classification accuracy of 91.51 % (Kappa = 0.86). (2) In SSI construction, the height-percentile stratification thickness, point-cloud sampling, and voxel edge length together set Caragana korshinskii stratification accuracy and density scale; the 5 % height percentile interval, a sampling size of 100 points, and 0.04 m voxel edge length proved optimal. (3) Comparative experiments showed that the three-dimensional feature integrated SSI significantly outperformed single-feature, two-feature, traditional allometric equation, and random forest (RF) models, with the SSI-based model achieving R-2, RMSE, MAE, and rRMSE of 0.90, 529.01 g, 432.58 g, and 26.54 %, respectively. These results indicate that SSI more effectively captures shrub spatial structure and improves AGB prediction under sparse UAV-LiDAR conditions.