Abstract
To address the insufficient retrieval accuracy of temperature and humidity profiles from ground-based microwave radiometers, this study develops an improved retrieval method based on intelligent optimization and attention mechanisms under clear-sky conditions. The research first constructs a brightness temperature dataset by performing forward modeling of ERA5 reanalysis data for the Haikou region using the MonoRTM model. Subsequently, five common machine learning algorithms—Ridge Regression, XGBoost, Multilayer Perceptron (MLP), Random Forest, and Support Vector Machine (SVM)—are employed to reconstruct atmospheric temperature and humidity profiles. A systematic evaluation of the retrieval results reveals distinct performance characteristics across the five machine learning models. For temperature retrieval, the Wolf Pack Algorithm (WPA) is introduced to optimize SVM parameters, resulting in a WPA-SVM model that reduces the overall Root Mean Square Error (RMSE) by 13.2% compared to conventional SVM. For humidity retrieval, an innovative MLP model integrated with an attention mechanism is proposed. By incorporating adaptive weighting and oversampling strategies, this model significantly improves retrieval accuracy at high altitudes, achieving a 15.1% reduction in overall RMSE compared to traditional MLP. The hybrid retrieval framework developed in this study integrates swarm intelligence and machine learning, under clear-sky conditions, providing a reliable technical pathway for high-precision and high-robustness atmospheric parameter retrieval. These findings not only hold significant value for advancing high-accuracy atmospheric parameter remote sensing systems, but also establish a foundation for precisely retrieving atmospheric refractivity profiles to resolve atmospheric ducts, thereby carrying important implications for marine wireless communications.