• While Stormtrack has discontinued its hosting of SpotterNetwork support on the forums, keep in mind that support for SpotterNetwork issues is available by emailing [email protected].

ChatGPT for Weather Analysis?

JamesCaruso

Staff member
Joined
Jul 5, 2009
Messages
1,961
Location
Newtown, Pennsylvania
Playing around with ChatGPT to see if it can analyze a satellite image for the setup in the southern Plains this evening, 05/24/25, from a storm chasing perspective. Below is the image I uploaded, my prompt, and the response. Curious how you think ChatGPT performed. Also post about ways you have experimented with it or actually used it for chasing.

IMG_2601.jpeg

Prompt:
Can you interpret this satellite image? Identify the types of clouds you see in each region and what they might indicate (for example, “in the southwestern quadrant of the state of X, there are clouds of the type Y, which indicates Z…” Based on your interpretation, where would you want to be right now as a storm chaser?

Response:
“This is a visible satellite image centered on Oklahoma and surrounding states, with the timestamp showing late afternoon (6:42 PM CDT) on Saturday, May 24. Here’s a regional breakdown of cloud types and what they suggest:

Northeastern Oklahoma:
  • Cloud type: Extensive altocumulus and stratocumulus, some showing gravity wave patterns.
  • Interpretation: These mid-level clouds and wave features often indicate stable conditions aloft, likely a capping inversion, but with moisture and lift in the lower levels. Storm development may be suppressed here unless the cap breaks
Western Oklahoma & Texas Panhandle:
  • Cloud type: Isolated cumulus growing into towering cumulus, especially in the far eastern Texas Panhandle and western Oklahoma.
  • Interpretation: These are bubbling cumulus clouds indicating rising air and possible convective initiation. Some are showing vertical development suggestive of early thunderstorm updrafts. This area is most promising for storm chasing right now.
Southwestern Texas (near the Big Bend and northwest of it):
  • Cloud type: Larger convective cells with overshooting tops casting shadows — indicative of ongoing thunderstorms or strong development.
  • Interpretation: Storms are already mature or rapidly intensifying here, possibly severe.
Eastern Texas & Southeastern Oklahoma:
  • Cloud type: Shallow cumulus fields, more scattered.
  • Interpretation: Daytime heating is generating instability, but there’s little vertical development — likely lower chances of organized convection for now.
New Mexico and the High Plains:
  • Cloud type: Stratiform and thicker mid- to high-level clouds in the northern parts; cellular convection starting in the southeast.
  • Interpretation: Upper-level dynamics might be moving in, helping to destabilize this region slowly, though storms are still less organized compared to western OK/TX.
Best Spot for Storm Chasing Right Now:

Target Area: Western Oklahoma into the eastern Texas Panhandle, especially around the I-40 corridor west of Clinton to east of Amarillo.

Why:
  • Visible signs of vertical development (towering cumulus).
  • Clear skies elsewhere support surface heating and instability.
  • Proximity to likely surface boundaries or dryline.
  • Good visibility and road network for chasing.
Want me to check SPC outlooks, mesoscale discussions, or radar to pinpoint an exact chase target?”
 
Last edited:
Turn GPT onto itself using a constant AI ping for HRRR/RAP/RADAR/SATELLITE/SPC and it could be interesting.. and scary. I dont know what its skill would be really though in determining TOR warns .. my guess it would probably meet and not exceed the same skill as all of the models mentioned and looking at yesterdays Bust in OK, GPT wouldn't have done me any favors.. but acting as a Co-pilot to put me in the right area?, it might take some burden off of analysis for solo chasers. So is it helping? or making me more lazy.. or providing the inexperienced with those smooth responses of like.. Who needs weather experience.. I just ask GPT! (I stayed at a HIE last night, I'm an Expert!) or does society not need expertise anymore, just results, because they designed their way out of a job. You see where I'm going with this, lol.
 
Oh I definitely agree with you Jason! I would never just want “the answer” on a target from ChatGPT. Even before ChatGPT was a thing, I was posting on here that chasing would lose its appeal if forecasting skill ever got to the point of being able to say where and when a tornado would occur. As much as I complain about the frustration and disappointment of busting, I am well aware that it’s all the failures that make the successes so magical. In fact, I have come to realize that it’s all about the *process* for me, and with an easy AI answer there is no “process.”My purpose of posting this was more about testing what ChatGPT *could* do. I would definitely be interested in using it to learn, including by seeing how it explains things that I may be unable to understand or interpret myself. With this post I was more interested in whether people thought it interpreted the satellite image properly.
 
Pretty interesting stuff there. If all the models were punched in I wonder if it would have told me not to chase the NAM 3K ghost of storms that never materialized in SE Colorado on Friday. NAM and HRRR had a big disagreement and HRRR won. I wonder if it has enough information to be able to tell which models were more accurate with different scenarios. I personally think we’re just at the tip of the iceberg with this whole AI thing…
 
Wow...that response seems shockingly reasonable! I wonder what else it could do. Can you have it respond to a prompt showing a model forecast image?
 
I was surprised as well .. how its interpreting an image like that. We've been working on AI image capture analysis for a little bit, apparently others are to, for it to be able to distinguish area, stable vs. unstable, is pretty impressive.
 
This is cool, but I will caution that ChatGPT will hallucinate things to justify an assertion. I've asked various LLMs to analyze images, to determine where it was taken is one example use. ChatGPT will make assumptions, "this architecture is common in Italy so this is likely Italy," but then it will hallucinate evidence to support its assumption. It will pointblanc claim "the Italian license plate in the photo confirms this is likely Italy." There is no Italian license plate in the photo. It made it up. This has happened repeatedly in my experience. So really, any conclusions you ask ChatGPT to draw, you need to confirm, and if you're putting in the effort to analyze the image anyways, why ask ChatGPT in the first place?
 
Back
Top