Machine Learning Datasets for the Earth's Natural Microwave Emission

Principal Investigator: Carl Mears, Remote Sensing Systems of California Corporation

The purpose of this project is to recast the 32-year record of satellite observations of the Earth's natural microwave emission into datasets that are compatible with machine learning.

The microwave spectrum contains a wealth of information on our ever-changing planet. From providing information on sea ice, to ocean winds and temperature, to atmospheric moisture, to soil and vegetation cover, the microwave spectrum is unique and vital to understanding Earth. Records of these geophysical parameters can all be calculated using a combination of observed microwave radiances from satellites and relevant ancillary geophysical parameters associated with their corresponding Earth scenes.

In order to assemble and use the entire record of satellite microwave observations, however, measurements from a number of disparate instrument types need to be combined. These instruments often differ in seemingly arcane measurement details, such as the exact frequency of the measurement and the angle at which the Earth is observed. This presents a challenge as measurement differences lead to differences in the observed radiances that need to be characterized and removed before the data can be combined or used in a common retrieval algorithm. 

Another challenge is that radiance data are typically available in “swath” form, with elliptical measurement footprints that are irregularly spaced in latitude and longitude. This makes it difficult to collocate the measurements taken by microwave imaging radiometers with other measurements, or with known characteristics of the Earth's surface. Because of these challenges, algorithm development has generally been undertaken by remote sensing specialists, instead of attracting the attention of the large and growing machine learning community.

This project will perform the specialized work that currently serves as a barrier to more widespread study of microwave measurements. This will lead to the creation of easy-to-use datasets which will subsequently be made available to researchers and other users worldwide. The following actions will be performed for a large collection of microwave radiometers:

  • Adjusting the measurements from the disparate instruments so that they are referred to a common set of frequencies at a single measurement angle.
  • Resampling the measurements (using a minimal-noise technique) onto a regular, latitude and longitude Earth grid.
  • Making available (on an identical grid) a set of collocated ancillary variables that we deem likely to aid algorithm development. Examples of these data are: 1) Meteorological information, including atmospheric profiles of temperature, moisture, the presence of rain, surface temperature, and wind; 2) Information about Earth's surface, such as open water fraction and land use/land cover; and 3) Informative quality flags to aid the user in culling measurements that are inappropriate for their purpose.
  • Retrieving surface emissivity for each frequency using atmospheric parameters from reanalysis on an identical regular Earth grid.
  • Using these emissivities and forward microwave models of the atmosphere to produce a second set of training data constructed using diverse, exactly-specified atmospheric conditions.

All of this will be freely provided in easy-to-use formats via popular web interfaces to maximize the potential for widespread use. All together, 150 satellite-years of data will be provided, spanning more than 35 years of recent Earth history, by the end of the project.

Page Last Updated: Sep 24, 2020 at 1:09 PM EDT