Cooperative Institute for Research in the Atmosphere

ML Overview

---

ML Activities Related to Satellites

---

CIRA’s long history of working with satellites

Satellite Earth Station

With NOAA support, CIRA has operated a Geostationary Operational Environmental Satellite (GOES) Earth Station since 1980.  Today our collection capability handles three simultaneous GOES transmissions, NOAA polar data and European MSG.  CIRA also plays an important role in each new GOES GVAR satellite as one of the primary test sites for initial transmissions and sensor verification.

(On right:  Satellite dishes at CIRA headquarters, Fort Collins, providing real-time satellite data feeds.)

Satellite Products

Many of CIRA’s satellite products are publicly available here:
Check out SLIDER at https://rammb-slider.cira.colostate.edu/
(For starters, switch “sector” from “Full disk” to “Conus”, then choose a “Product” to view different satellite products, e.g., from GOES-16, in real time.)

Tropical Cyclones

CIRA and the Atmospheric Science Department at CSU also have a long and strong history of research on tropical cyclones which we leverage for some of the satellite activities described above

ML Activities Related to Numerical Weather Prediction (NWP) Models

---

CIRA’s ML activities related to NWP models fall into two categories:

1. Extracting information from satellites for assimilation into NWP models.
   For more information, see activities outlined for satellite tasks above.
2. Using ML to speed up NWP models by emulating those physics equations that are particularly slow, namely radiative transfer, by much faster ML approaches.
   For more information on this topic see ML4RT.
3. Using ML to combine satellite data and NWP model output to improve vertical profiles of temperature and dewpoint.
   For more information, see ML4Soundings.

Our ML Approaches and Expertise

---

Using machine learning to inform life-and-death decision making, e.g., to make decisions on which regions to evacuate, is very different from using machine learning for standard computer science applications. In particular, in order for machine learning approaches to be accepted into operational use for weather forecasting, they must be as robust and interpretable as possible.

We believe the best way to achieve robustness and interpretability is to make ML models as simple as possible, using any of the following approaches:

• Integrating as much expert knowledge as possible into the ML algorithms by
  • Extensive feature engineering from expert knowledge
  • Developing custom loss functions that better model what experts are truly interested in
• Exploring tools from other fieldsto extract relevant additional features, such as
  • Optical Flow
  • Traditional image processing techniques (pre-defined filter banks, e.g., Gabor filters)
  • Harmonic analysis (Fourier or wavelet transforms in image space)
  • Topological data analysis (persistent homology)
• Simplifying neural networks by
  • Performing ablation studies to test which model elements are truly important
  • Fitting simpler machine learning models to more complex ones
• Providing well-calibrated uncertainty estimates along with the central predictions to supply information on how much to trust the model’s prediction, by
  • Implementing approaches to capture aleatoric and epistemic uncertainty
  • Evaluating the estimates using a variety of graphical and scoring metrics
• Applying Explainable AI (XAI) methods to learn about the strategies used by ML algorithms, such as
  • Permutation importance tests
  • Shapley values
  • Attribution maps for neural networks (e.g., Layer-wise Relevance Propagation, input*gradient)

Use of an attribution method, namely Layer-Wise Relevance Propagation (LRP), to analyze a neural network model that seeks to translate imagery from the Geostationary Operational Environmental Satellite into Multi-Radar Multi-Sensor (MRMS) imagery. LRP results (bottom row) show where in the input channels the neural network is focusing when predicting the MRMS value for a single (central) pixel.  See Ebert-Uphoff and Hilburn (2020) for details.

CIRA Resources for the Community

---

A) Short courses for Atmospheric Scientists to get started in Machine Learning

CIRA Short Course on Machine Learning for Weather and Climate:

• Available material: recorded video lectures, GitHub, and Jupyter Notebooks
• 6 sessions, 90 minutes each.   Dates: 16 September – 21 October 2020
• Instructors:  Ryan Lagerquist (CIRA Boulder and NOAA GSL) and Imme Ebert-Uphoff (CIRA Fort Collins and Colorado State University)
• Link: CIRA Short Course ML for W&C website.
• Or go directly to Youtube playlist (this link only gets you to the videos, not the other course material)

CIRA/AI2ES Short Course on Explainable AI:

• Available material: recorded video lectures, each accompanied by slides and a Colab notebook with open-source Python code.
• All Python examples focus on weather prediction.
• 6 sessions, 90 minutes each.   Dates: 10 May – 21 June 2021
• Instructor: Ryan Lagerquist (with guest lecture by Imme Ebert-Uphoff)
• Course Information and Recordings

The Roebber Lectures:

• Instructor:  Paul Roebber
• Organized by NOAA’s VLab
• 4 sessions (Feb 9 & 15, April 14 & 22, 2022)
• Description:  The National Weather Service (NWS) has a long history of leveraging available data in support of weather forecasting efforts. These efforts are ongoing and with the advent of more advanced techniques (e.g., machine learning), the NWS is in the process of determining where and how to apply them.  This series of four talks is designed to provide some background on these techniques. No assumptions will be made regarding the statistical background of participants. The following table contains information on each talk including a link to register for the presentation.
• Course Information and Recordings

B) Papers / Reports on Foundational Topics:

• CIRA Guide to Custom Loss Functions for Neural Networks in Environmental Sciences–Version 1.  (2021)
  Ebert-Uphoff, I., Lagerquist, R., Hilburn, K., Lee, Y., Haynes, K., Stock, J., Kumler, C. and Stewart, J.Q.
  https://arxiv.org/abs/2106.09757
• Machine Learning for Clouds and Climate – Invited Chapter for the AGU Geophysical Monograph Series” Clouds and Climate (2021)
  Beucler, T., Ebert-Uphoff, I., Rasp, S., Pritchard, M. and Gentine, P., 2021.
  https://www.essoar.org/doi/abs/10.1002/essoar.10506925.1
• The Need for Ethical, Responsible, and Trustworthy Artificial Intelligence for Environmental Sciences.  (2021)
  McGovern, A., Ebert-Uphoff, I., Gagne II, D.J. and Bostrom, A., 2021.
  https://arxiv.org/abs/2112.08453