To further increase the data rate and rate flexibility of next-generation passive optical networks (PONs), we propose a scheme jointing probabilistic constellation shaping (PCS), constellation diagram identification, and coherent detection. The rate of the coherent PON is dynamically regulated with time-variant entropy, and monitored by the image classifier in real-time. The transmitted cascade frames allocated to different optical network units (ONUs) from the optical line terminal (OLT) can be synchronized, identified, and distinguished by the labeled entropy values. The proposed graphic monitoring scheme is experimentally demonstrated based on a dual-polarized coherent optical transmission system. A 0.5 interval value of entropy is precisely tuned to achieve a 96.13% constellation identification accuracy for entropy/rate by a fast normalized cross correlation coefficient (NCC)-based image classifier. A real-time data rate from 350 to 550 Gb/s with PCS-64-quadrature amplitude modulation (64-QAM) format for a single ONU is achieved which can maximally support five independent ONUs with single wavelength and two polarization states and the entropy varying from 3.5 to 5.5 with 0.5 intervals, respectively.