Principal Investigator: Jianwu Wang, University of Maryland, Baltimore County
Clouds cover about two thirds of Earth's surface and play a critical role in our climate system, with fundamental influence on the energy, water, and biological cycles. Currently, satellite-based remote sensing is the only way to observe clouds on a global scale. For these reasons, cloud observations have always been a major task of NASA's Earth Science endeavor.
In the latest NASA Decadal Survey, cloud observations have been given top priority for NASA's missions. Numerous satellite sensors have been developed to observe and retrieve cloud properties. They can be largely divided into two groups: active sensors such as those used by the Cloud Aerosol Lidar and Infrared Pathfinder Satellite Observation (CALIPSO) and CloudSat missions, and passive sensors such as the Moderate Resolution Imaging Spectroradiometer (MODIS), the Visible Infrared Imaging Radiometer Suite (VIIRS), and Advanced Baseline Imager (ABI).
The advantages of active sensors include their capability of resolving the vertical location of cloud layer and better performance during nighttime and over polar regions. On the other hand, passive sensors have a much better spatial sampling rate.
Machine learning (ML)-based algorithms have brought revolutionary changes to almost every aspect of our lives. ML is also increasingly used in NASA's satellite remote sensing algorithms. For most machine learning, especially deep learning (DL)-based algorithms, high-quality training datasets are critical.
The overarching goal of our proposed project is to develop an extensible platform that combines collocated satellite observations, numerical simulations, and deep learning methods to generate a highly accurate cloud property training dataset for NASA, NOAA, and the broad science community to develop and benchmark algorithms for passive satellite cloud remote sensing.
This project will deliver:
- Novel deep-learning-based domain-adaptation algorithms to retrieve passive satellite remote sensing cloud bulk properties (e.g., cloud mask and thermodynamic phase) by leveraging one or more available active sensing data.
- A novel hybrid approach combining advanced 3-D radiative transfer simulations based on collocated global satellite observations and deep learning based multi-pixel cloud microphysical and optical property retrieval.
- Scalable data processing and analytics services in a public cloud computing environment (i.e., Amazon Web Service) for the above components/capabilities.
- Comprehensive data quality evaluation of the training datasets (where retrieved cloud properties are data labels) to be delivered from multiple aspects including statistics, climatology, ground observation, and ad hoc case studies.
In particular, we will generate four-year (2017–2020) labeled cloud property training data from the joint NASA/NOAA Suomi National Polar-orbiting Partnership (Suomi NPP) satellite and the geostationary GOES-16 satellite.
The outputs of this project will greatly help NASA scientists and the broader community to independently or collectively develop machine-learning based cloud remote sensing algorithms, compare and evaluate the cloud retrieval products.