A Deep-Learning-Based Microwave Radiative Transfer Emulator for Data Assimilation and Remote Sensing

Abstract

In this article, we introduce a fully connected deep neural network algorithm to emulate the Community Cadiative Transfer Model (FCDN_CRTM) simulation of brightness temperatures (BTs) from the Advanced Technology Microwave Sounder (ATMS) channels for clear-sky cases over ocean surfaces. The FCDN_CRTM fine-tuned through three sensitivity experiments with respect to sample-size determination, model separation, and introduction of novel features toward improving the accuracy of the model. In addition to the BT simulation, we evaluated the Jacobians with respect to surface and atmospheric parameters. Atmosphere profiles from the European Centre for Medium-Range Weather Forecasts, sea surface temperature from the Canadian Meteorology Centre, and ATMS sensor data records were used as FCDN_CRTM inputs. In comparison to CRTM, the FCDN_CRTM minus CRTM mean biases were within several hundredths of a Kelvin (K), and the corresponding standard deviations (SDs) were between 0.05 and 0.15 K for all ATMS bands. The accuracies for both mean bias and SD were consistent throughout the evaluation period, which spanned approximately 1 year beyond the period of the FCDN_CRTM training dataset. Furthermore, the model Jacobians generally compared well with CRTM Jacobians in terms of surface temperature, wind speed, air temperature, and (log) water vapor. The performance of the FCDN_CRTM forward and Jacobian model indicate potential for use in data assimilation and physical retrieval systems, such as the NOAA operational Microwave Integrated Retrieval System.

Keywords

• Advanced Technology Microwave Sounder (ATMS)
• Community Radiative Transfer Model (CRTM)
• data assimilation
• deep learning (DL)
• Jacobian
• remote sensing