Global cloud phase monitoring from FY-3G multi-angle polarimetric observations using the new MAPLE-CP framework with independent multi-site lidar validation

Abstract

Global cloud-top phase monitoring is constrained by an observing gap: passive imagers provide broad coverage but limited vertical phase information, whereas active sensors provide stronger physical constraints only along narrow tracks. Here we develop MAPLE-CP, a Multi-Angle Polarimetric Latent-Enhancement framework for FY-3G PMAI cloud-top phase retrieval. MAPLE-CP links shortwave-infrared multi-angle polarimetric scattering information, DQ-1 ACDL cloud-top lidar labels, physics-guided feature screening, compact latent augmentation, and independent lidar validation. From approximately 140 PMAI variables, 50 physically interpretable predictors were selected and enhanced with six self-supervised latent descriptors. Using a 142,362-sample PMAI–ACDL cloud-top phase corpus, MAPLE-CP achieved internal accuracy, macro-F1, and AUC of 0.946, 0.897, and 0.991, respectively, on the original internal test split. Grouped split diagnostics clarified the evidence boundary of this internal benchmark, and independent validation using 6,433 ground-based lidar samples from 31 Chinese sites gave an agreement of 0.788 and macro-F1 of 0.780 within this external network. On an availability-limited shared-collocation subset with MODIS, Himawari-9, and ground-based lidar, MAPLE-CP showed higher agreement and macro-F1, suggesting complementary PMAI polarimetric phase information rather than a global product ranking. Overall, MAPLE-CP supports scalable cloud-top phase mapping and exposes residual mixed-phase uncertainty.