FIREMAPS

Project Title: FIREMAPS: Fire Imaging, Recognition, and Evaluation Mapping with AI Processing SAR

Team Members:

Rafael Rubenstein and Matt McMaster

Background:

One of the main benefits of synthetic aperture radar (SAR) imagery is the ability to image day and night and during all weather. The ability to penetrate through thick smoke during an active wildfire emergency highlights a key advantage over normal optical imagery, which is blocked by the smoke. Designing a machine learning (ML) model that inputs a SAR image and can determine the active boundary of a wildfire presents a unique capability for first-responders to mitigate the spread of the fire and provide ample warning to nearby residents. We utilized imagery from the Open Capella SAR Dataset to create a supervised Convolutional Neural Network that will detect the geographic extent of a wildfire. We trained a model using labeled wildfire data. Our main objective was to draw a bounding shape around the burned landscape with our trained model. Our secondary objective was to incorporate elevation data to create a topographical map of the damaged area and its surroundings to further inform the optimal locations for first-responders' response.

Datasets we used:

We used the following datasets:

• https://felt.com/map/Capella-Space-Open-Data-bB24xsH3SuiUlpMdDbVRaA?loc=0,-20.53,1.88z
• https://radiantearth.github.io/stac-browser/#/external/capella-open-data.s3.us-west-2.amazonaws.com/stac/catalog.json?.language=en
• https://data-nifc.opendata.arcgis.com/datasets/7c81ab78d8464e5c9771e49b64e834e9_0/explore?showTable=true
• https://planetarycomputer.microsoft.com/dataset/cop-dem-glo-30

Tools/packages we used:

Download_Data.ipynb requires the following Python packages:

• os
• requests
• json
• urllib.parse
• numpy
• rasterio
• matplotlib

Label_Chip_Data.ipynb requires the following Python packages:

• numpy
• geopandas
• matplotlib
• matplotlib_scalebar
• contextily
• os
• rasterio
• shapely
• warnings

CNN.ipynb requires the following Python packages:

• json
• numpy
• rasterio
• torch
• PIL
• os
• itertools
• sklearn

DEM.ipynb requires the following Python packages:

• pystac_client
• planetary_computer
• os
• rasterio
• geopandas
• shapely
• matplotlib
• matplotlib_scalebar
• contextily
• rioxarray

Methodology:

• Download SAR imagery from Open Capella SAR data, wildfire perimeter data from WFIGS, and Digital Elevation Map data from Copernicus 30m DEM
• Load in SAR data in regions with active wildfires using methods learned in class/labs
• Identify intersections between boundaries of SAR extent and wildfire boundaries
• Chip labeled SAR data to prepare for CNN training
• Train CNN using hyperparameter grid search and evaluate accuracy to choose model
• Apply model to subset of chipped, labeled data to validate accuracy
• Create DEM overlay with SAR/wildfire intersections to gain greater situational awareness regarding topographical features in the area

Goals:

Gain experience with SAR data and ML techniques. Produce a model with greater than 75% accuracy.

Results / Conclusions / Future Directions:

• Results: Successfuly prepared SAR data and trained a model with 77.5% accuracy!
• Conclusions:
  • Predicting active wildfire boundaries with X-band SAR imagery is a realistic task.
  • SAR data is tricky to work with - much of the time was spent figuring out how to feed it into the CNN.
  • CNNs are expensive to train! More time and/or processing power would have helped immensely.
• Future Directions:
  • Build out a binary image classifier to filter out amount of data that get fed into the CNN boundary generator.
  • Integrate DEM, weather data, soil mosture data, etc. to create a more robust model to predict where fires may spread to improve first-responder situational awareness.

References

We used the following references as a guide:

• https://www.nature.com/articles/s41598-019-56967-x#Tab1
• https://www.sciencedirect.com/science/article/abs/pii/S0143622822002259
• https://tema-project.eu/articles/wildfire-mapping-using-sar-derived-nbr-through-generative-adversarial-network
• https://www.diva-portal.org/smash/get/diva2:1557423/FULLTEXT02.pdf
• https://arxiv.org/html/2501.09129v1
• https://www.earthdata.nasa.gov/learn/earth-observation-data-basics/sar
• https://byrd.osu.edu/news/revolutionizing-wildfire-management-sar-technology-enhances-detection-and-response
• https://medium.com/@robmarkcole/deep-learning-with-sar-18b27972a4b5