HE2LM-AD: Hierarchical and efficient attitude determination framework with adaptive error compensation module based on ELM network

Satellite attitude determination, as a crucial pre-processing technology in remote sensing earth observation, is significantly associated with the geometric accuracy of satellite stereo images. To compensate for the observation errors in raw attitude data for achieving accurate results, the conventional Kalman filter-based and filter-smoothing combination frameworks provide insufficient access to statistical features of data, resulting in inferior practical system performance. To conquer these drawbacks, a hierarchical and efficient attitude determination framework is proposed called HE2LM-AD, which consists of a simplified adaptive Kalman filtering module, a neural network-based system error compensation module, and a weighted attitude smoothing module. To be more specific, an adaptive noise covariance matrix update is exploited in the Extended Kalman Filter (EKF) iterations, and a simplification scheme based on Cholesky decomposition is employed to obtain coarse filter outputs efficiently. Following that, a dedicated Extreme Learning Machine (ELM)-based error compensation network is designed to capture the pattern of observation errors, providing robust compensation at lower overhead. To further pursue high-performance attitude determination results, we fused the outcomes of the aforementioned two components according to the weights achieved from adaptive learning. Experiment results on in-orbit satellite attitude simulation datasets demonstrate and validate the feasibility and effectiveness of our framework.