TWM Contributor: Dan Newcomb
GIS Manager

Let’s Talk about AI

There’s a lot of discussion about “Artificial Intelligence” (AI) lately with the rise in popularity of tools like ChatGPT, DALL-E, Midjourney, Stable Diffusion, Jasper, Google’s Bard, and other applications. What you might not know is that many of these aren’t true artificial intelligence. Artificial Intelligence has become a bit of a generic term for machine learning and related areas of study. Bard, Jasper, and Chat GPT are all “Large Language Models” (sometimes called transformer models), and while they may seem intelligent, they’re essentially just more conversational search engines backed up by massive statistical models.

Image tools such as DALL-E, Midjourney, and Stable Diffusion aren’t really AI either except in the more popular use of the term, although they can produce some truly fascinating images. Nearly all of these are evolutions of machine learning. At the end of the day, it all comes down to statistics. Given enough pictures of a cat, samples of language, or paintings by Matisse, a computer will eventually understand those concepts quite well (a large language model like Chat GPT can have trillions of parameters).

We’re often surrounded by these algorithms every day, but we don’t think about them when they’re designed and implemented well. Toyota (and other automotive manufacturers) has an option available on some of their cars that uses a front-facing camera to recognize speed signs, putting it up on the dash or a heads-up display for you to see. Other examples are the spam filtering on your email, removing a background from a photo, the voice to text feature on your phone, when your video doorbell tells you someone or a package is on your porch, or maybe an app that scans a picture of a plant to tell you whether it’s a flower or a weed you should pull. All of these are examples of machine learning in action.

These algorithms require training. When you see one of those captchas on a web page that says, “Click all the images with a traffic signal”, you’re helping to train an ML model or algorithm. Large companies can also train their models by basically just feeding them internet content. For example, if you wanted to train a camera to recognize “dogs”, you could theoretically just point it to google or another search engine and ask it to return all the images of dogs.

If you’re wondering what happens when a copyrighted or trademarked image or text is integrated into an AI algorithm, then produces a result for someone else, you’re not alone. There have already been and will continue to be lawsuits over copyright and trademark infringement as well as licensing violations. The laws are not well prepared for this, so it will likely take some time to settle. This also brings about a valuable warning: anything you feed back into one of these tools is potentially visible to the public. As these models learn and interact with users, they can use your feedback elsewhere. If you ask Chat GPT to write an executive summary of a contract document that you provide, pieces of that document have the potential to appear in someone else’s executive summary. There are already companies providing firewalled AI’s to corporations and governments, preventing any data from going outside their control. Expect to see more of that in the future as the privacy concerns of these technologies are better understood.

The Power of Spatial Intelligence

In the GIS world, versions of these AI tools and methods have been leveraged for many years, although recent improvements in both software and hardware have greatly improved their efficiency and efficacy. The integration of these two fields is unlocking novel opportunities for analysis, visualization, and decision-making.

Building footprints from imagery using Deep Learning from one of TWM’s tests.

Even so, training a model can still take considerable time. Running a test training routine on one of our top workstations with significant amounts of RAM and a powerful GPU, easily took 72 hours of continuous iteration with 10 square miles of sample imagery. We’re fortunate that ESRI, the company that provides the GIS software that we use, has a library of pre-trained models available as well that can provide some excellent starting points for further refinement.

Creating a special purpose-built model is obviously quite intensive for time and computing resources. For this reason, custom models only make sense for larger projects. Using pre-trained models can help lower that barrier substantially. Many valuable insights can be extracted from GIS datasets. They’re getting quite good at extracting features from imagery and LiDAR like cars, pools, sidewalks, buildings, trees, etc. Expect these tools and methods to become more viable for smaller projects soon.

It isn’t all about imagery or LiDAR though. These same statistical methodologies are being applied within GIS to areas like asset management to predict maintenance and lifecycles (our software partner DOT does this, for example), hydrologic modeling, traffic prediction, change detection, and multi-sensor inputs (mashing up radar, LiDAR, and imagery for example).

The convergence of AI and GIS mapping include:

  • Enhanced data analysis and insights
  • Satellite imagery and object detection
  • Environmental monitoring and conservation
  • Predictive modeling and forecasting
  • Spatial data management
  • Real-time monitoring and response

Fusion for the Future

Artificial intelligence, machine learning, deep learning, and all their various names and iterations aren’t terribly new, but they are reaching the point of being very useful and user-friendly. As AI technologies continue to advance, we can expect more sophisticated applications that will allow decision makers to be better informed about planning, resource allocation, and risk assessment. We are very near a point where algorithms can not only react based on modeling, but reason as well. It’s wise to be mindful about who’s using the data you provide and where it could potentially end up. It’s also important to apply sound data security practices. Expect some of these tools to be integrated into the software you use every day – if it isn’t already!

To find other resources, you can browse ESRI’s existing Deep Learning models here: https://www.arcgis.com/home/search.html?q=deep+learning+package

Decision Optimization Technology:
https://www.bettercapitalplanning.com/

Deep Learning in ArcGIS:
https://www.esri.com/arcgis-blog/products/api-python/analytics/deep-learning-models-in-arcgis-learn/

Flood inundation mapping with radar and AI:
https://www.esri.com/arcgis-blog/products/arcgis/imagery/flood-inundation-mapping-using-radar-ai/

Data privacy:
https://ai.googleblog.com/2023/05/making-ml-models-differentially-private.html

AI, GIS, and disaster resilience:
https://storymaps.arcgis.com/stories/784114d3bf5c406d9ec9e170425caa6f

Read more about our Geospatial and GIS Mapping services.