Deep Learning Techniques for the Generation of Satellite-borne Thermal Infra-red Spectra and Their Applications to Remote Sensing Retrievals.

Generative Artificial Intelligence is a branch of Machine Learning algorithms. Generative Models such as Generative Adversarial Networks (GANs) and Autoencoders (AE) models can be used to generate data and manipulate images (e.g. Goodfellow et al., 2014). The generated data allow to produce large amount of data at reduced computing costs with respect to traditional Forward Models (FM) and to tests new configurations of variables. In the frame of the ESA DeepLIM project, GANs and AE are used to simulate Thermal Infra-Red (TIR) spectra acquired by satellite-borne instruments. The spectra are generated starting from a dataset of spectra simulated with a Radiative Transfer Model (RTM) or measured during an aircraft campaign. The simulated dataset consists on both Top Of the Atmosphere (TOA) and Bottom Of the Atmosphere (BOA) radiances. The generated dataset is composed by the same fields. To simulate the satellite measurements, both original and newly generated radiances are convolved with spectral response functions in the 5 TIR channel of the Sentinel Land Surface Temperature Monitoring (LSTM) mission (at 8.6, 8.9, 9.2, 10.9, 12.0 ?m) and converted into Brightness Temperatures (BTs). Then, the extended dataset is processed to retrieve Land Surface Temperature (LST) in the five channels. The retrieval is performed using both "state-of-the-art" algorithms (e.g. TES algorithm (Gillespie et al., 1998)) and Neural Network (NN) We will present the characteristics (e.g. average values, standard deviations) of the generated dataset and compare them to the original ones. Then the results of LST retrieval will be presented to asses the usability of generated spectra for remote sensing applications. Although in DeepLIM project the generated spectra are used to retrieve surface parameters, the fact that the underling radiative properties of the original spectra are preserved suggest that the same features can be used to retrieve atmospheric variables. This highlights the capability of deep learning techniques to support missions and retrieval algorithm development. References Gillespie, A. R., S., Rokugawa, T., Matsunaga, J., S., Cothern, S., Hook, and A. B. Kahle (1998), A Temperature and Emissivity Separation algorithm for Advanced Spaceborne Thermal Emission and Reflection radiometer ASTER images. IEEE Transactions on Geoscience and Remote Sensing, 36, 1113-1126. Goodfellow,I., J., Pouget-Abadie, M., Mirza, B., Xu, D., Warde-Farley, S., Ozair, A., Courville, and Y. Bengio (2014), Generative adversarial nets, Advances in neural information processing systems, 2672-2680.