Improving Retrievals of Temperature and Humidity Profiles From a Microwave Radiometer and the Error Analyses

Thermodynamic profiles measured from the ground‐based microwave radiometer (MWR) have been widely utilized to enrich the atmospheric vertical observations, which however, are characterized by certain retrieval errors. Temperature and humidity retrievals from an in‐situ MWR are assessed against the corresponding radiosonde observations. Calibration experiments are conducted using methods of Kalman filter (KF), random forest (RF), neural network (NN), and time‐involved NN (TiNN), which are subsequently examined across various metrics. The significant errors of MWR retrievals vary with heights and are reduced to certain extents, via these calibration models. KF and RF both offer considerable calibrations on the MWR retrievals but perform inconsistently across different atmospheric layers. The TiNN model is demonstrated with the optimal performance, producing stable and accurate retrievals with root mean square errors (RMSEs) of no greater than 2.0 K and 20% across all heights for temperature and humidity, respectively. Corresponding improving magnitudes reach upto 60% relative to the raw MWR. Such superiority of the TiNN model demonstrates potential advantages of integrating associated auxiliary features, such as time tags into calibration models. The challenges in retrieving relative humidity compared to temperature might be linked to the lower signal‐to‐noise ratios in the moisture retrievals and greater sensitivity to environmental conditions due to indirect and complex nature of the measurements.