Author: Magnus Tveit, University of Utah MSGIS Capstone
Community Partner: Tommy Thompson, Utah Division of Forestry, Fire & State Lands (FFSL)
Academic Supervisor: Phoebe McNeally, University of Utah
Utah's wildfire risk is mapped using a Structural Exposure Score (SES), where scores of 7 or 8 represent the highest-risk areas — covering approximately 138,000 km², or roughly 63% of Utah's land area. The High-Risk Wildland Urban Interface (HRWUI) is a subset of this zone, defined as areas where two or more structures exist within 250 meters of each other within the SES 7/8 boundary.
Under HB48, the Utah Division of Forestry, Fire & State Lands (FFSL) is responsible for identifying all structures that meet this criteria and maintaining a complete, accurate inventory of qualifying buildings. This dataset did not previously exist in verified form.
This project develops a two-track approach to build that inventory:
- Manual review of existing Overture Maps Foundation (OMF) building polygon data in ArcGIS Pro
- Deep learning pipeline to identify structures missed by current datasets using 15cm Hexagon aerial imagery
ffsl-hrwui/
│
├── README.md ← this file
│
├── data/
│ ├── inputs/ ← validation input shapefiles
│ ├── predictions/ ← completed DL inference outputs
│ └── outputs/ ← figures and CSV outputs
│
├── deep_learning/
│ ├── deep_learning_README.md
│ ├── ENVIRONMENT_deep_learning.md ← environment setup
│ ├── notebooks/ ← pipeline scripts
│ │ ├── chip.py ← training chip generation
│ │ ├── train.py ← MaskRCNN fine-tuning
│ │ ├── infer.py ← county-by-county inference
│ │ └── infer.slurm ← SLURM submission script
│ └── models/ ← model weights
│
├── SQL/
│ ├── SQL_README.md
│ ├── ffsl_parcel_join.sql ← three-way join script
│ └── parcel_join_strategy.csv ← county-specific parcel ID cleaning rules
│
└── validation/
├── validation_README.md
├── hrwui_validation.ipynb ← validation notebook
└── environment.yml ← conda environment for validation
All large data files live on the University of Utah CHPC at:
/uufs/chpc.utah.edu/common/home/u0972368/FFSL_HRWUI/
They are not stored in this repository. See each subfolder's README for what data is needed and where to obtain it.
Model weights are hosted on Google Drive. Download and place at deep_learning/models/:
| File | Size | Description |
|---|---|---|
best_model.pth |
176 MB | Fine-tuned MaskRCNN weights - use for inference |
usa_building_footprints.pth |
176 MB | ESRI base model weights |
usa_building_footprints.dlpk |
165 MB | ESRI base model package |
Google Drive link: (https://drive.google.com/drive/folders/1oHUUzfBstTIaE6exV0mRdQSz-aAzCwMh?usp=drive_link)
See deep_learning/README.md for full setup and inference instructions.
See SQL/README.md for instructions on running the parcel join for any Utah county.
See validation/README.md for instructions on running the validation notebook.
| Metric | Value |
|---|---|
| Ground truth structures | 572 (manually verified) |
| DL detections (full county) | 2,350 |
| DL detections (SES 7/8 only) | 1,466 |
| True Positives | 534 |
| False Positives | 932 |
| False Negatives | 64 |
| Recall | 89.3% |
| Precision | 36.4% |
| F1 Score | 0.517 |
| Detections outside existing HRWUI | 468 |
High recall is the right outcome for this use case - missing a real structure is more costly than reviewing a false detection under HB48 fee administration.
- Tommy Thompson & Utah Division of Forestry, Fire & State Lands (FFSL)
- Phoebe McNeally, University of Utah
- CHPC (Martin Cuma) for HPC resource allocation
- UGRC for Hexagon 15cm aerial imagery
- Overture Maps Foundation for building polygon data