Turning Imagery into Action with GeoAI

When a major disaster strikes, every minute matters. Roads may be blocked, buildings damaged, and infrastructure disrupted, yet aid must still reach hospitals and affected communities as quickly as possible.

At this year’s Developer and Technology Summit plenary, Priyanka Tuteja demonstrates how GeoAI can turn post-disaster imagery into operational insight, rapidly mapping roads, assessing damage, generating optimized response routes, and accelerating decision-making.

To illustrate this workflow, Priyanka focuses on a critical response task: delivering medical aid from a helicopter landing zone to two priority locations—a hospital and a relief tent in Westmoreland—while navigating a transportation network disrupted by Hurricane Melissa, a Category 5 storm that swept through Jamaica.

Reconstructing the Road Network from Imagery

With pre-disaster data often unreliable, Priyanka shows how GeoAI allows responders to reconstruct the transportation network directly from post-disaster imagery. She uses a road extraction model from ArcGIS Living Atlas of the World to automatically detect roads from imagery captured after the hurricane. The resulting network reflects the real-world conditions on the ground.

However, detection alone isn’t enough, as responders need to know which roads are usable.

Classifying Road Conditions with a Vision-Language Model

Once roads are extracted, Priyanka uses the Classify Objects Using Deep Learning tool to run the Vision Language Context-Based Classification pretrained model, which categorizes each road segment based on post-disaster conditions:

• Completely blocked
• Major debris
• Minor debris
• Clear

By providing context—specifying that the imagery is from a Category 5 hurricane—the model interprets visual patterns more accurately. Roads classified as blocked or heavily obstructed can then be used as polygon barriers in routing, ensuring that response teams avoid dangerous or impassable routes.

Generating Routes for Rapid Response

With an updated road network in place, Priyanka performs routing analysis starting from the helicopter landing zone, including the hospital and relief tent as stops, to generate optimal routes that bypass blocked roads.

In just minutes, responders move from raw imagery to actionable routing, enabling faster delivery of critical medical aid.

Mapping Debris with a Foundation Model

Removing debris is essential for effective disaster response, helping responders direct equipment, personnel, and aid to the areas that need it most. For this, Priyanka fine-tunes a foundational remote sensing model, Dynamic One-For-All (DOFA), to generate a debris layer.

Foundation models are pre-trained on massive amounts of imagery, giving them an understanding of spatial patterns in landscapes, spectral characteristics in remote sensing imagery, and relationships between objects and environments.

Because of this, DOFA required fewer than 100 training labels and only about 30 minutes of fine-tuning to produce a detailed debris layer highlighting areas with significant accumulation. This illustrates one of the key advantages of foundation models: they dramatically reduce the amount of training data needed to produce useful results.

Evaluating Building Damage with GeoAI

Assessing structural impacts is equally important. Here, Priyanka demonstrates how GeoAI models from ArcGIS Living Atlas—such as Roof Decking Delineation and Roof Hole Delineation—can be used to quickly identify building damage.

She begins by exploring the models’ item pages, which include detailed descriptions, input requirements, performance metrics, and links to GitHub repositories where the training workflows and implementation details are documented. Using the provided notebook, Priyanka fine-tunes the model on the post-disaster imagery to adapt it to the study area.

The result is a building damage severity layer that highlights structures showing visible signs of impact and categorizes the severity of that damage, helping responders quickly understand where structural damage is most significant.

From Imagery to Actionable Intelligence

When disaster strikes, time is the most valuable resource. Priyanka’s demo illustrates how GeoAI can rapidly turn imagery into operational intelligence, helping responders make informed decisions when every minute matters. How will you use GeoAI to turn imagery into action?