P008-01
Atmospheric Parameters Retrieval Using Machine Learning In Resource Limited Spacecraft Remote Sensing
Monday, 7 December 2020: 05:30
Virtual
Nicolas Gorius1, Thanh Nguyen1, Anthony Mamakos2, Grace Zimmerman3, Thuc Phan1, Dat Tran1, Shahid Aslam4, Valeria Cottini5, Tilak Hewagama4 and George Nehmetallah1, (1)Catholic University of America, Washington, DC, United States, (2)University of California Los Angeles, Los Angeles, United States, (3)University of Central Arkansas, Conway, AR, United States, (4)NASA Goddard Space Flight Center, Greenbelt, MD, United States, (5)Italian Space Agency, Rome, Italy
Abstract:
We present initial results of our neural network model for retrieval of planetary atmospheric parameters applicable to space-based spectroscopic observations. The advent of low power [<10W] specialized application-specific integrated circuit (ASIC) devices for neural network inference [e.g. Google Edge TPU] has enabled the implementation of efficient neural networks in embedded devices for Earth and planetary exploration. Our goals were to develop (1) a retrieval methodology for identification and quantification of atmospheric constituents, (2) a reasonable model footprint [<60MB], (3) instrument level integration, (4) leverage existing hardware and opensource frameworks, and (5) provide real or in near-real time results and feedback [> 0.5 fps]. The training and validation of our current neural network, using an ensemble of spectra from radiative transfer forward calculations, can be completed in a single day on a single GPU, enabling the deployment of model updates based on the actual data while in operation. Derived information could then be used to significantly reduce the telemetry data volume or prioritize/optimize data to be downloaded from the spacecraft. Such protocols will enable optimization of remote operations.
We envision that the training methodology is applicable to atmospheric studies in the ultraviolet, visible, infrared and microwave domains for a wide range of instruments and scientific targets. Other applications include mineralogy characterization using reflectance and/or Raman spectra.
We gratefully acknowledge publicly available datasets for training and testing purpose [1] and compared our results with published models [2]. We will discuss accuracy, execution time, and compare results with other published models. We will also discuss limitations of using unsupervised machine learning in the non-linear limit, and address constraints of this kind of approach.
References:
[1] ExoAI/TauREx data, I. Waldmann, https://osf.io/hn8uk/
[2] ExoGAN, T. Zingales and I. Waldmann, https://osf.io/6dxps/
