Abstract
Hyperspectral image target detection algorithms for remote sensing, surveillance, and environmental monitoring require large quantities of high-quality test data featuring authentic target spectral, spatial, and illumination characteristics to enable rigorous validation and performance assessment. However, acquiring real-world hyperspectral imagery with precise ground-truth annotations is extremely costly and logistically challenging, leading to a widespread reliance on synthetic data generation. Traditional synthesis approaches primarily embed targets through two-dimensional image stitching, i.e., superimposing target images directly onto background scenes. These techniques struggle to model three-dimensional geometry, correct perspective projection, or the directional reflectance behavior of surface materials, resulting in inherent deficiencies in illumination consistency and material realism. The resulting images often lack critical visual cues such as cast shadows and occlusion boundaries, which may bias the performance evaluation of detection algorithms. To overcome these limitations, this paper introduces a 3D model-based target insertion method that integrates Blender’s physically-based rendering engine with the MATLAB data processing platform to form an automated generation pipeline. The framework converts detailed 3D models into targets with realistic material properties and precisely embeds them into hyperspectral background data cubes, while supporting dynamic simulation of aircraft motion states including variations in attitude and trajectory. The core insertion process consists of three stages: (1) establishing a real-time bidirectional TCP/IP communication channel between MATLAB and Blender for pose control and perspective-correct rendering; (2) automated material segmentation via clustering in a perceptually uniform color space to enable independent spectral processing; and (3) pixel-wise weighted fusion of material spectra using physically-based modulation coefficients derived from textures, thereby constructing multi-material hyperspectral cubes with realistic spatially varying spectral signatures. Experimental results demonstrate that the proposed method preserves geometric perspective cues and heterogeneous material appearance. Compared with conventional two-dimensional stitching, the generated data exhibit improvements in visual realism and reliability, thus providing a more trustworthy benchmark. Moreover, its dynamic simulation capability can produce temporal image sequences that closely resemble real aircraft maneuvers, offering operationally meaningful test samples for the rigorous evaluation of moving target detection and tracking algorithms in dynamic, realistic scenarios.