Skip to content


Machine Learning for Soundings (Vertical Profiles of Temperature and Dewpoint)

Soundings for atmospheric conditions supporting high CAPE. Red (blue) shading have been added for CAPE (CIN). [a] Example of UNet improving over RAP. [b] Example of UNet degrading RAP.

To accurately forecast thunderstorms, it is crucial to have frequent and accurate vertical profiles of the temperature and dewpoint, both of which affect the instability of the atmosphere.  This instability is typically measured by the quantity convective available potential energy (CAPE), and  the likelihood that the instability will be realized is given by the quantity convective inhibition (CIN).  CAPE and CIN are highly sensitive to the temperature and dewpoint in the lowest layer of the atmosphere, usually from the surface to 1-2 km in altitude.

Radiosondes (weather balloons) provide the best thermodynamical profiles, as they measure atmospheric variables with high resolution while ascending in the atmosphere.  However, unfortunately balloon releases are conducted routinely only twice a day and are separated by several hundred kilometers.  As a result, forecasters leverage vertical profiles generated by numerical weather prediction (NWP) models, which have high spatial and temporal resolution.  Unfortunately, these models are not perfect, and this work uses machine learning (ML), with input from satellite data, to improve vertical profiles of temperature and dewpoint produced by an NWP model.

As part of this work, we

For the ML model, we explored the use of different ML model architectures, including linear regression, fully-connected neural networks, convolutional neural networks, and deep residual UNets.  Using data from 2017-2020, we found that all ML model architectures improve the NWP profiles; however, the UNet models are the best-performing since they can fully-utilize the vertical spatial information.  The ML models also substantially improve the CAPE and CIN predictions.

Temperature (solid) and dewpoint (dashed) profiles when the RAP (pink) and the ML model (purple) perform poorly. Observations are black, and the shading shows the 95% confidence as predicted by the ML model.  Four examples are shown at different stations and times. 

In addition, ML model’s are capable of predicting uncertainty estimates that correctly match the corresponding error for the majority of the predictions without suffering any degradation to the temperature and dewpoint profiles (see figure).  Having uncertainty estimates at all vertical levels helps identify where the ML model may have the largest errors, providing forecasters not only with a heads-up for potential errors, but also with information regarding where in the profile these errors are occurring so that they may use their domain expertise to help overcome these errors.  The uncertainty estimates are well-calibrated, meaning that removing the most uncertain samples results in improved model performance; thus, if specific applications require more accurate results, separate thresholds on both temperature and dewpoint can be set up to show only the most-certain profiles with reasonable confidence.

NOAA-Funded Future Work


Primary Contact

Katherine Haynes (CIRA)