Bias characteristics and estimation of the FORMOSAT-7/COSMIC-2 radio occultation bending angle in the lower troposphere with a machine learning algorithm

FORMOSAT-7/COSMIC-2 radio occultation (RO) measurements have great potential for monitoring the lower troposphere and offering crucial insights into the Earth’s planetary boundary layer. However, due to limitations of the retrieval technique, assumptions used in the algorithm, and atmospheric influences, RO data retrieved from the lower troposphere can be severely biased under highly moist environments and strong refractivity gradients. Under these conditions, the retrieved bending angle profile exhibits pronounced fluctuations or spike-like deviations, respectively. This study examines the characteristics of the RO bending angle bias (BAB) and thereby proposes a machine learning algorithm based on a multilayer perceptron neural network, which is trained with different input proxies to assess region-dependent BAB. The results show that the BAB model can accurately predict the BAB in the lower troposphere in different regions. This research highlights the promise of advanced methodologies for improving RO retrieval and promotes data applications in the lower atmosphere.