Robust reinforcement learning multi-baseline phase unwrapping algorithm with JPEG-LS

Multi-baseline (MB) phase unwrapping (PU) breaks through the limitations of the phase continuity assumption, but existing MBPU algorithms cannot adapt well to noisy environments. A robust reinforcement learning integer programming multi-baseline PU algorithm with JPEG-LS (RL-MBPU) is proposed. First, the cutting plane–pure integer programming cut-selection process is formulated as a sequential reinforcement-learning problem. An efficacy-guided imitation strategy is used to initialize an MBPU-structure-aware single-cut policy, which is further optimized by an actor–critic method using Cut-PAR2-based trajectory returns to reduce the root-node proof cost. Finally, the initial ambiguity number solution is corrected using dynamic median edge detector based on JPEG-LS, and the PU is completed by using the prediction of the left, top, and top-left neighbors of the current pixel and combining it with the local gradient threshold determination. In both simulation and real data experiments, the RL-MBPU algorithm demonstrated stable PU performance, including computational efficiency and noise robustness for interferograms of different sizes. Compared with the minimum-cost flow and cutting plane–pure integer programming algorithms, the RL-MBPU algorithm reduces the root mean square error by approximately 61.2% and 14.5%, respectively, while improving efficiency by approximately 54.1% and 42.4%, respectively. This can further support high-precision interferometric synthetic aperture radar applications.