Hi everyone, my name is Gina Girgente and I am finishing up my M.S. GISc degree at the University of Denver in June. My thesis researched how geoAI can be used to understand wildfire-induced building loss, and I think the workflow that was developed can be helpful for emergency management professionals, urban planners, and residents in wildfire-risk zones alike. Thus, I want to share my research with you all to increase accessibility of the subject.
What is Geospatial Artificial Intelligence?
GeoAI, “the application of artificial intelligence (AI) fused with geospatial data, science, and technology to accelerate real-world understanding of business opportunities, environmental impacts, and operational risks” [1], is becoming increasingly accessible thanks to Esri’s contributions and is streamlining geospatial tasks, such as disaster response and emergency management [4, 5].
Object detection versus instance segmentation.
My research used a pretrained model by Esri with data inputs from before and after a wildfire in order to discern building loss influences. Labeled by Esri as the state-of-the-art model for instance segmentation, Mask R-CNN not only predicts bounding boxes, but also object masks [2]. In this case, instance segmentation is better than object detection because the model outputs the actual building footprint, rather than a bounding box around each building it identifies.
Methods for the research
General methodology.
To complete this workflow, you’ll need: satellite imagery before and after a wildfire of interest, access to Esri’s Mask R-CNN pre-trained model and geoprocessing tools, and basic statistical knowledge to extract building and environmental variables, as well as analyze how they relate to building loss.
Data preprocessing
I downloaded imagery from Maxar’s Open Data Program, which was in a geographic coordinate system projection. The model needs a projected coordinate system, so I reprojected it using UTM. I then ensured the imagery was a resolution accepted by the model.
Workflow for AOI creation.
My dataset was large, so I created areas of interest (AOIs). For each area of interest, I needed imagery before the fire, after the fire, and a ground truth layer. A ground truth layer is data of verified houses that I can compare my model to and evaluate the accuracy of its performance. For this, I used a building footprint layer by Microsoft from the year of the fire I studied.
I used the Generate Tessellations tool to normalize the area of each AOI. With each iteration of a tessellation, a building footprint ground truth layer was created by selecting buildings from the original layer if they intersected with the AOI image. If there were 1 or more houses per AOI, the new ground truth layer was saved in its respective geodatabase. Additionally, an image the size of that AOI was saved into a geodatabase. This iteration continued until all tessellations were evaluated.
Building footprint extraction
This workflow partially follows instructions from Esri’s tutorial for the building footprint object detection deep learning package [3]. It is a great place to start learning geoAI!
Workflow for building footprint extraction.
An image file and the deep learning package are input into the Extract Features Using AI Models tool. After running for each AOI, you will have layers of detected buildings pre- and post-fire.
Model evaluation
Next, we need to make sure the model accurately detected buildings. Ideally, this would be done by manual ground truthing (going to your study area in person and verifying the buildings yourself). However, this isn’t realistic temporally, or based on the size of your study area or budget you have. Instead, you can download Microsoft’s free building footprint layer and use Esri’s Compute Accuracy for Object Detection geoprocessing tool. I chose an IoU (Intersection over Union) of greater than or equal to 0.5. The F1 score is the main evaluator of model performance for this project because it acknowledges false negatives, which is important in the case of building loss. For more information on model metrics, read this article by Esri.
From here, building-derived variables can be computed, like building density, average building distance, number of buildings, and percent building loss (per AOI). You can also extract various environmental variables, if appropriate.
Discussion
Ground truth footprints overlayed on imagery, with some footprints missing.
Ground truth footprint shapes may not match the true shape.
Some footprints are cut in half by AOI boundaries.
In regard to model metrics, if you choose to use the same ground truth data I did, your model metrics may report less accuracy than is realistic. Some areas of imagery show buildings that the ground truth layer does not. Additionally, the ground truth layer has some instance of incorrect building shapes. Finally, some buildings are cut in half by AOIs, which created irregular building shapes by models.
With these limitations of the metrics being realized, the model performed relatively well for me. Burnt imagery metrics will naturally be lower than the non-burnt imagery because the ground truth dataset usually more closely mirrors the urban landscape before the wildfire of interest.
Statistical analysis
I used multiple linear regression with the Forward selection method due to the unknown importance of the provided independent variables. My dependent variable was percent building loss per AOI, which was calculated by finding the difference of buildings before and after the fire per AOI.
Don’t want to do the workflow manually? Test this tool in progress!
Linked here is the user manual and geoprocessing tool that will automate the process described above. Note that this tool usage is still new and may not work for you at all, regardless of if you meet the requirements. If you try the tool or the manual workflow, please comment on this blog about what you think and how it can be improved!
References
- esri. (n.d.). What Is GeoAI? | Accelerated Data Generation & Spatial Problem-Solving. Www.esri.com. https://www.esri.com/en-us/capabilities/geoai/overview
- How Mask R-CNN Works? | ArcGIS API for Python. (2017). Arcgis.com. https://developers.arcgis.com/python/latest/guide/how-maskrcnn-works/
- Learn to use AI to extract information from World Imagery. (2025, February 11). ArcGIS Blog. https://www.esri.com/arcgis-blog/products/arcgis-living-atlas/imagery/learn-to-use-ai-to-extractinformation-from-world-imagery
- Singh, R. (2025, February 10). Pretrained Geospatial AI Models Can Streamline Disaster Response | Winter 2025 | ArcUser. Esri. https://www.esri.com/about/newsroom/arcuser/pretrained-geospatial-ai-models-can-streamline-disaster-response
- Speranza, C. (2025, February 11). Leveraging Geospatial Artificial Intelligence in Emergency Management. Industry Blogs. https://www.esri.com/en-us/industries/blog/articles/leveraging-geospatial-artificial-intelligence-in-emergency-management
