Atmospheric temperature and humidity profiles are key variables linking observations, models, and forecasts. They are used in weather diagnosis, convective assessment, pollution transport simulation, and climate research. Profiles come from radiosondes, satellites, and reanalysis, but ground-based microwave radiometers offer continuous monitoring. However, station distribution is uneven, and most retrieval models use single-station inputs, causing spatial discontinuities and poor generalization in sparse or heterogeneous networks. Geographic characteristics such as latitude, longitude, and station distance are often ignored. Based on these challenges, in-depth research on multistation collaborative, spatially informed atmospheric profile retrieval is needed.
Developed by researchers at Wuhan University and collaborators, and published (DOI: 10.34133/remotesensing.1063) in Journal of Remote Sensing on 28 August 2026, CTDformer (Convolutional Neural Network [CNN]-Transformer-DualBranchMLP) targets these gaps. The framework uses networked ground-based microwave radiometer observations to retrieve regional vertical profiles, aiming to improve numerical weather prediction initialization, severe-convection monitoring, pollution simulation, and climate services. It explicitly embeds station identity and geographic relationships into the learning process rather than treating each site independently.
CTDformer combines CNN, Transformer, and DualBranchMLP base learners with an XGBoost meta-learner. At four stations, it outperformed individual models, reducing temperature and humidity profile errors by 7.64% and 5.31%, respectively. Spatial information lowered errors by more than 7%. Networked training improved accuracy by 17.39% for temperature and 5.69% for humidity versus single-site retrieval. Before the 15 May 2021 rainfall, CTDformer captured stronger convective available potential energy buildup than ERA5 and single-station schemes, offering earlier instability signals.
The study used 22-channel brightness temperatures from ground-based microwave radiometers in Weiyang, Chang'an, Lintong, and Yangling, with Jinghe as the target station. Radiosonde profiles provided reference labels. After temporal matching and quality control, 529 samples were used. Inputs included station one-hot encoding, latitude, longitude, and distance to the target station. Profiles covered 89 vertical layers. Temperature retrieval reached R˜0.99; humidity R˜0.75. On 15 May 2021, CTDformer estimated pre-rain CAPE above 1,500 J/kg, whereas ERA5 showed only hundreds of joules per kilogram, with sharp increases 1-3 hours before rainfall.
The research team noted that CTDformer improves numerical accuracy and preserves physical consistency, providing reliable support for severe convection monitoring and short-term forecasting. They see multisource integration and physics-informed constraints as key next steps.
Microwave radiometer brightness temperatures were matched with radiosonde profiles at synoptic times. Rain-affected samples were removed using near-surface humidity thresholds; channel-wise IQR filtering removed outliers. Data were normalized to [0.1,0.9] and split 7:1:2. CNN, Transformer, and DualBranchMLP were trained with 10-fold cross-validation; XGBoost fused their outputs. Spatial priors were appended to input features.
Future work will integrate radiometers, radar, radiosondes, and satellite products, and validate across complex terrain. Cross-attention and graph-based fusion could better model interstation relationships. Physics-informed losses and thermodynamic constraints may improve robustness under extreme conditions, supporting intelligent meteorology, short-term forecasting, and climate monitoring.