Applying Deep Learning to Clear-Sky Radiance Simulation for VIIRS with Community Radiative Transfer Model—Part 2: Model Architecture and Assessment

Abstract

A fully connected “deep” neural network algorithm with the Community Radiative Transfer Model (FCDN_CRTM) is proposed to explore the efficiency and accuracy of reproducing the Visible Infrared Imaging Radiometer Suite (VIIRS) radiances in five thermal emission M (TEB/M) bands. The model was trained and tested in the nighttime global ocean clear-sky domain, in which the VIIRS observation minus CRTM (O-M) biases have been well validated in recent years. The atmosphere profile from the European Centre for Medium-Range Weather Forecasts (ECMWF) and sea surface temperature (SST) from the Canadian Meteorology Centre (CMC) were used as FCDN_CRTM input, and the CRTM-simulated brightness temperatures (BTs) were defined as labels. Six dispersion days’ data from 2019 to 2020 were selected to train the FCDN_CRTM, and the clear-sky pixels were identified by an enhanced FCDN clear-sky mask (FCDN_CSM) model, which was demonstrated in Part 1. The trained model was then employed to predict CRTM BTs, which were further validated with the CRTM BTs and the VIIRS sensor data record (SDR) for both efficiency and accuracy. With iterative refinement of the model design and careful treatment of the input data, the agreement between the FCDN_CRTM and the CRTM was generally good, including the satellite zenith angle and column water vapor dependencies. The mean biases of the FCDN_CRTM minus CRTM (F-C) were typically ~0.01 K for all five bands, and the high accuracy persisted during the whole analysis period. Moreover, the standard deviations (STDs) were generally less than 0.1 K and were consistent for approximately half a year, before they significantly degraded. The validation with VIIRS SDR data revealed that both the predicted mean biases and the STD of the VIIRS observation minus FCDN_CRTM (V-F) were comparable with the VIIRS minus direct CRTM simulation (V-C). Meanwhile, both V-F and V-C exhibited consistent global geophysical and statistical distribution, as well as stable long-term performance. Furthermore, the FCDN_CRTM processing time was more than 40 times faster than CRTM simulation. The highly efficient, accurate, and stable performances indicate that the FCDN_CRTM is a potential solution for global and real-time monitoring of sensor observation minus model simulation, particularly for high-resolution sensors.

Benefits of Stochastic Weight Averaging in Developing Neural Network Radiation Scheme for Numerical Weather Prediction

“…An NN emulator that can be used in the RTM was developed some time ago (Chevallier et al., 1998) and was applied to the data assimilation system of the numerical weather prediction (NWP) model (Chevallier et al., 2000). The emulation studies in the RTM are still actively performing (Bue et al., 2019; Liang & Liu, 2020; Stegmann et al., 2022), eventually targeting to the aircraft‐satellite data assimilation in relation to the improvement of forward operator. Recent RTM emulator studies based on clear‐sky simulations have shown a of 1.87–10.88‐fold speedup (Liu et al., 2020) when used with the Rapid Radiative Transfer Model for GCMs (RRTMG; Iacono et al., 2008), and 1.8–3.5‐fold (Ukkonen et al., 2020) and up to 4‐fold (Veerman et al., 2021) for the RRTMG—Parallel scheme (Pincus et al., 2019).…”

Improving ATMS Imagery Visualization Using Limb Correction and AI Resolution Enhancement

“…However, the relativistic loss aims to predict the probability that a real image is relatively more realistic than a fake one, as shown in Eq (5), which aids in learning shaper edges and more detailed textures. Furthermore, ESRGAN removes batch normalization (BN) [15], used in SRGAN, to further minimize artifacts in the generated image and employs a residual-in-residual dense block (RRDB) instead of the original residual block (RB) to improve training stability. All these advantages enable ESRGAN to effectively enhance image resolution at larger upscaling factors.…”

“…Deep Learning, in particular, has become one of the most widely used AI methods. Its applications in remote sensing and numerical weather prediction are currently being investigated [12][13][14][15][16][17]. The remote sensing image processing using a state-of-art convolutional neural network (CNN) [18][19][20][21][22][23] method has also become popular [24,25].…”

Machine Learning for Light Sensor Calibration

“…Here, we used a multilayer neuron to calibrate the input features provided by the low-cost light sensors against the data provided by the Minolta reference sensor [ 35 , 36 , 37 ]. A multilayer neuron (MLP) is a class of feedforward artificial neural network (ANN).…”