Training

Slide 1: UAH GOES-R CI

Welcome to the UAH GOES-R CI product training module, produced by the University of Alabama in Huntsville and NASA’s Short-Term Prediction Research and Transition Center, or NASA-SPoRT. This tutorial covers basic methodology and examples for the newest version of the UAHuntsville GOES-R CI experimental operational product. The product was produced by UAHuntsville and is transitioned to operations and testbeds by NASA-SPoRT. The module is about 12 minutes long.

Slide 2: Introduction

The purpose of this training is to help end-users (primarily forecasters) develop familiarity with GOES-R CI, and to show a few examples of how the product appears in operations.

The GOES-R CI Team at UAHuntsville aims to partner with forecasters and other end-users to produce satellite-based multi-dataset products that solve forecast challenges.

NASA-SPoRT transitions a number of NASA and NOAA datasets to operations with the objective of improving short-term predictions.

Slide 3: GOES-R CI Product Background

This is a GOES-R future capabilities product.

It fuses Numerical Weather Prediction and geostationary data into a single product in order to produce a 0-2hour nowcast of convective initiation. Part of the product’s capabilities are performed “under the hood”, so to speak. For example, there are algorithms for tracking cloud objects and identifying cloud types and properties. There are other mathematical methods for identifying those cumulus clouds that are most likely to become convectively active, based on compared training datasets from other regions and seasons. All of that occurs “under the hood” in a way that the end user will not be aware of each of these steps. Some of these steps might even be incrementally modified in future versions. What the end user sees in their decision support system is the cloud objects themselves colored on a map by their probability of convective initiation, much like a radar reflectivity display. This display looks like the image on the right and will likely not experience much change in future versions of this product.

Slide 4: GOES-R CI Application Example

This initial example demonstrates the general purpose of the GOES-R CI product, providing lead time and locations of convective initiation, defined here as radar echoes of 35 dbz or greater. GOES-R CI shows a map of cloud objects that are colored by their probability of becoming convectively active in the next 2 hours, with warmer colors corresponding to higher probabilities of CI.

At 1815 UTC, the GOES-R CI output on top shows a few cloud objects with probabilities of CI of 50, 60, and 70% embedded within lots of other cloud objects with almost no probability of CI. These higher probability objects are worth paying attention to over the next few hours. Indeed, the radar image below, from 1845 UTC, show the first 35 dbz echoes for those objects advected downstream which were shown on the CI display.

Later on, at 1915 UTC, there has been substantial development along the dryline. GOES-R CI screens out mature convection, however, and only looks for convective initiation, picking up on a few cloud objects with a 90% probability of becoming convective in the next 2 hours in the southern section of the line. The radar image at 2040 UTC verifies this prediction, showing the new development occurring.

Slide 5: GOES-R CI Product Details

There are some details that are helpful to keep in mind when using this product. For example, the spatial resolution is 1 km in the current GOES instrument in daytime operations and 4 km at night. The temporal resolution for most scans is 15 minutes, but can be as frequent as 5 minutes for GOES rapid scan mode or as infrequent as 30 minutes for full disk scans. This product operates on both the GOES-East and –West domains. The input data for the product includes 9 interest fields from the GOES satellite which indicate cloud motion, growth, glaciation, and height and their associated trends; 15 fields from the RAP Numerical Weather Predication Model which help indicate environmental information such as CAPE, CIN, Lifted Index, and several others; and information from the SNODAS ground snow cover map.

The 24 fields of satellite and NWP data are given to a logistic regression equation. This type of regression technique uses a training dataset of variables and a binary outcome, in this case whether or not a cloud object becomes convective, and produces an equation to predict the probability of the outcome’s occurrence, given the values of the input variables. The resulting output is a probability of convective initiation within the next two hours for the identified cloud objects. This probability only tells the likelihood of convection, and does not in itself correspond to intensity of convection.

Slide 6: Snow Contamination

The snow coverage map information is used separately to identify areas of snow contamination. Snow covered ground, and clouds moving over the snow can cause the snow to be incorrectly identified as cloud objects by the cloud typing algorithm. For this reason, snow contaminated objects are labeled pink in the display, as shown below.

Slide 7: Cirrus Contamination

Cirrus clouds are another source of potential problems for the IR processing portion of this product, because they obstruct the view of the growing cumulus clouds that are being assessed for convective properties. Cirrus clouds above or optically touching other tracked cloud objects will usually be removed by the cloud masking algorithm and none of the cirrus contaminated area will be included in the processing, as shown in the top part of the image. If cirrus clouds are inadvertently missed, they appear very cold and therefore abnormally “convective-likely” to the CI product, as shown on the bottom portion.

Slide 8: Cirrus Contamination and Cloud Typing

Cirrus contamination can be a pretty tricky hazard for the GOES-R CI product, causing some unexpected results. Here is an example in which a cloud object was initially masked by some thin cirrus clouds, and by the time the object had “detached”, it was already a mature convective cloud, and therefore it was never given a probability of CI by the algorithm. We’ll walk you through it.

First off, recall that the algorithm identifies cloud types and only gives cumulus clouds a probability of CI. In the first timestep, the image on the bottom at 1602 shows the cloud typing algorithm. The blue arrow indicates where our cloud object is, and in the cloud typing algorithm, it is within a large cirrus shield, which is all the green color you see. If it were a cumulus cloud and CI-likely, it would be one of the blue objects on the cloud typing panel. The CI and VISIBLE image to the lower left shows our object a little better, as a slightly thick cloud in the VISIBLE. The radar image from 1607 shows it as a distinct cloud object.

At 1632 and 1633, the radar actually identifies this object as having a 35 dbz echo within it. But as the thin cirrus clouds are still lingering around it, as seen in both the VIS and in the cloud typing image, the cloud typing algorithm still identifies the object as part of a cirrus cloud and removes the whole section from processing.

At 1655, the object now has an overshooting top, according to the cloud typing algorithm.

And by 1725, it is obviously a mature convective cell. It is plain to see that if this object developed in your CWA and GOES-R CI provided no warning, you’d probably wonder why it missed such an obvious convective event, particularly if you were looking at the radar image alone (you never see the associated cirrus on the radar image). So the rule of thumb is, no matter how thin the cirrus is, if you can see it on the VIS, it’s enough to get the whole section masked out, and objects optically touching it, from that scan.

Slide 9: Weaknesses

Here we state some caveats of the GOES-R CI product and its usage. Again, the presence of thin cirrus close to growing cumulus clouds, might occasionally net some unexpected results, as previously mentioned.

Another potential problem is data flow. In order to track cloud objects, and therefore reach the steps to produce a probability of CI, the product needs three consecutive satellite scans. When there are data interruptions, it takes three scans to produce the next CI image. Therefore, even brief data interruptions can cause long delays in the GOES-R CI product as the processing “catches up”.

Another caveat to mention is that the training dataset is still growing, which is used by the logistic regression equation “under the hood”. Data from new regions and seasons are being added to it, but in the meantime, some regions might appear to have better or more consistent results than others. For this reason, it can be important to self-calibrate, or become familiar with how the various probabilities correspond to actual convective events in your area.

The final caveat to mention is that the resolution for nighttime processing is 4 km, vs the daytime resolution of 1 km. This means that nighttime images appear blocky in comparison to daytime images. Also, as shown on the images at the right, the jump from day to nighttime processing can even result in cloud objects becoming “lost” between scans.

Slide 10: Strengths

There are several strengths to this product as well. For one, being satellite-based, this product helps provide data in regions that are generally data-sparse, like off-shore, over mountains, or in other locations where there is little other instrumentation. Also, being data-fused with NWP had greatly improved the False Alarm Ratio of this version of GOES-R CI over previous versions. In the GOES-R era, the improved spatial and temporal resolution will continue to improve the utility of this product for end-users.

Slide 11: Application Example

In this application example, we demonstrate how GOES-R CI can provide additional guidance in data-deprived locations. Mountainous terrain causes considerable radar beam blockage in the western US in places like northwestern New Mexico, as shown in the radar coverage map to the right. In the series of convective cells that developed over the course of a few hours ending at 2200 UTC, none of them were visible on radar until they were in western Rio Arriba County (that’s the second county to the left on the northern border, just outside our orange circle). The GOES-R CI product gave advanced warning of the convection, at 1845 and 1915 UTC, and also indicated a very high probability that convection was likely to occur in San Juan County, where our radar coverage was most compromised.

Slide 12: Acknowledgments and Notes

The UAH GOES-R CI product development team deeply appreciates the time, effort, and thoughtful feedback provided by the forecasters and various end-users of this experimental product. It is through these collaborative efforts that our work is made possible.

For more in-depth information about this product, including peer-reviewed publications, and real-time data feeds, visit the website at nsstc.uah.edu/uahci

For more information about NASA-SPoRT, SPoRT products and product training, or to see the NASA-SPoRT blog, visit the website at weather.msfc.nasa.gov/sport

Slide 13: Thank You!

Thank you for your time and attention!