Ground Passive Microwave Remote Sensing of Atmospheric Profiles Using WRF Simulations and Machine Learning Techniques

Abstract

Microwave radiometer (MWR) demonstrates exceptional efficacy in monitoring the atmospheric temperature and humidity profiles. A typical inversion algorithm for MWR involves the use of radiosonde measurements as the training dataset. However, this is challenging due to limitations in the temporal and spatial resolution of available sounding data, which often results in a lack of coincident data with MWR deployment locations. Our study proposes an alternative approach to overcome these limitations by harnessing the Weather Research and Forecasting (WRF) model’s renowned simulation capabilities, which offer high temporal and spatial resolution. By using WRF simulations that collocate with the MWR deployment location as a substitute for radiosonde measurements or reanalysis data, our study effectively mitigates the limitations associated with mismatching of MWR measurements and the sites, which enables reliable MWR retrieval in diverse geographical settings. Different machine learning (ML) algorithms including extreme gradient boosting (XGBoost), random forest (RF), light gradient boosting machine (LightGBM), extra trees (ET), and backpropagation neural network (BPNN) are tested by using WRF simulations, among which BPNN appears as the most superior, achieving an accuracy with a root-mean-square error (RMSE) of 2.05 K for temperature, 0.67 g m⁻³ for water vapor density (WVD), and 13.98% for relative humidity (RH). Comparisons of temperature, RH, and WVD retrievals between our algorithm and the sounding-trained (RAD) algorithm indicate that our algorithm remarkably outperforms the latter. This study verifies the feasibility of utilizing WRF simulations for developing MWR inversion algorithms, thus opening up new possibilities for MWR deployment and airborne observations in global locations.