A Deep Learning Framework Based on PointNet++ for RCS Prediction of Aircraft Targets

In this paper, we propose RCS-PointNet++, a novel deep learning framework for predicting the radar cross section (RCS) of complex 3D structures. Built upon the PointNet++ architecture, the model effectively captures both local and global geometric features from point clouds, enabling accurate and efficient RCS prediction. RCS-PointNet++ demonstrates strong robustness to variations in point cloud resolution and generalizes well across different target configurations. Inference on 25 aircraft models across 360 directions completes in under 52 seconds. Extensive experiments on two independent datasets comprising over 2,125 3D models validate the proposed method’s accuracy and generalizability, achieving a root mean square error (RMSE) of 3.09 and anR2of 0.92. Furthermore, enhancing the network’s perception of scattering direction reduces prediction error by approximately 6%. While initially designed for single-frequency prediction, the framework can be extended to multi-frequency scenarios by incorporating electromagnetic priors. This work offers a significant advancement in fast and accurate RCS prediction for complex 3D targets.