Fast Radiative Transfer Models and Data Assimilation Systems

The RRTM and eCrad codes compute only the band-integrated and total fluxes and are usually embedded in global models that call them billions of times within a single model run. Thus, while more inaccurate, these codes are, by far, faster than both RTTOV and σ-IASI/F2N.

Applications and know-how

Using the characterized radiative transfer models described earlier, we are able to build the end-to-end (E2E) chain (from data acquisition to Level 2 geophysical products) for the simulation of atmospheric passive remote sensing missions. The E2E simulation allows to characterize the product quality as a function of the measurement characteristics, thus it is extremely useful to set up the requirements in the initial phases of a new mission.

Conversely, given the accuracy, the precision and the geometrical specifications of a future or a already running mission, we can assess the information contained in the measurements and the possibility to derive new geophysical products.

Artificial Intelligence methods

Due to the large volume of data expected from next-generation sensors, new machine learning (ML) approaches have been proposed. Although these approaches must rely on physical models during the training phase, the algorithms themselves, typically based on architectures such as neural networks, learn a functional relationship directly from previously generated results of the same or similar problems.

In principle, ML methods can replace any full-physics procedure. However, in the field of remote sensing, research has mainly focused on three application areas:

• Scene classification, such as distinguishing between clear and cloudy conditions, or estimating indices related to scene homogeneity and cloudiness.
• The forward model, i.e. predicting a spectrum from a given atmospheric state (radiative transfer, RT).
• The inverse problem, i.e. retrieving the atmospheric state from measurements (retrieval).

Scene classification is important because the presence of clouds requires the use of radiative transfer models that account for multiple scattering. Modern instruments are often complemented by ancillary sensors for analyzing the field of view, and some instruments are specifically designed for cloud detection and characterization.

A major challenge for ML approaches applied to the RT problem is the curse of dimensionality. An atmospheric state may be described by hundreds of parameters, while the corresponding spectral radiance can consist of thousands of channels. In full-physics models, correlations between spectral channels are inherently represented. In contrast, ML models must learn these correlations from the training data set, which is a demanding task.

For the retrieval problem, the main difficulty lies in the ill-posed nature of the inversion. Full-physics retrieval methods address this issue through regularization techniques, such as reducing dimensionality using Principal Component Analysis (PCA), or constraining solutions towards a climatological prior. ML approaches must incorporate similar strategies, either by learning appropriate regularization from the training dataset or by projecting both atmospheric states and radiances into lower-dimensional spaces (latent spaces) and learning relationships between these representations (e.g. latent twin approaches).

Although ML methods offer very fast inference once trained, their retrieval accuracy still generally falls short of that achieved by full-physics approaches. Ongoing research aims to improve their precision, potentially making them a viable alternative to more computationally expensive methods.