Regional Atmospheric Profile Retrieval Driven by Multistation Collaboration based on Geoscience-Informed Deep Learning

Accurate retrieval of atmospheric temperature and humidity profiles is essential for improving numerical weather prediction and climate simulations. However, the uneven distribution of ground-based radiometer stations creates data gaps, limiting the accuracy and spatial adaptability of existing inversion models. To address this challenge, we propose CTDformer (Convolutional Neural Network [CNN]–Transformer–DualBranchMLP), a multimodel learning framework that combines convolution-based feature extraction, attention-based sequence learning, and a dual-branch multilayer perception to retrieve vertical profiles over the Xi’an region. CTDformer incorporates spatial features such as station encoding, geographic coordinates, and interstation distances to enhance representational capacity and generalization. Experiments at 4 stations demonstrate that CTDformer outperforms individual models, reducing temperature and humidity profile errors by 7.64% and 5.31%, respectively. Incorporating spatial information into networked training further decreases errors by more than 7%. On 2021 May 15, the results indicate that CTDformer is more sensitive in capturing the buildup of convective energy, yielding substantially higher convective available potential energy than both the fifth-generation European Centre for Medium-Range Weather Forecasts atmospheric reanalysis and the single-station scheme prior to precipitation. It better aligns with the observed convective development and highlights its potential for severe convection monitoring and forecasting. This study addresses the challenges of sparse ground-based radiometric observations and limited model spatial generalization. By leveraging multistation collaboration, it achieves high-precision atmospheric profile retrieval of region, thereby enhancing diagnostic and early-warning capabilities for extreme weather, smog, and precipitation events. It can also offer reliable data support for applications in intelligent meteorology, short-term forecasting, and climate monitoring.