A Spatial Resolution Enhancement Method of Microwave Radiation Imager Data Based on Data Matching and Transformer

Abstract

The Microwave Radiation Imager (MWRI) onboard the Fengyun-3D satellite (FY-3D) can provide valuable observation data in many fields such as meteorological research and weather forecasting. However, its coarse spatial resolution limits data application. Recently, image super-resolution (SR) methods based on deep learning have been introduced into the spatial resolution enhancement of radiometers and achieved better results than traditional methods. Most of them use the degradation model to generate dataset and build model based on convolutional neural network (CNN). However, the dataset generation method based on degradation model may impair the information in the original brightness temperature (BT) data. Moreover, CNN-based SR methods often struggle to model longdistance dependencies, which may impact the spatial resolution enhancement of BT data. To address these issues, we propose a dataset generation method based on BT data matching, and introduce a SR network model based on the Transformer structure. We refer to it as the SR Transformer BT Data Matching (SRTBTM) method. Results indicate that the method significantly improves the spatial resolution of MWRI data over current methods and exhibits strong generalization for long-term data outside the training time range.