Automated Post-Disaster Field Image Classification System
An ML-powered tool that integrates directly into field data collection workflows — enabling crews to photograph a site and receive an automatic classification of burned, missing, debris, or intact condition, embedded in the record.
FieldSight emerged from the 2025 Eaton/Altadena wildfire response, when Las Flores Water Company needed to assess hundreds of water infrastructure sites across a fire-damaged landscape — quickly, consistently, and with limited personnel.
The system works inside the data collection tools field crews already use. A technician photographs a site — the meter box, the surrounding lot — and FieldSight automatically classifies what it sees, writing the result directly into the survey record without interrupting the workflow.
Rather than requiring a data analyst to manually review hundreds of images after the fact, every photo becomes a structured data point the moment it is taken.
Hundreds of photos reviewed by hand — slow and inconsistent under crisis conditions
Raw images in Survey123 carry no condition data unless manually labeled
Crews need an ordered dispatch list — not a photo archive to interpret
Only ~20 manually labeled examples available to train the classifiers
FieldSight classifies every field image into one of four structural condition categories — enabling automated population of data collection records without manual review.
Structural or meter components show direct fire damage — charring, melting, or combustion evident in the image. Highest replacement priority.
Expected infrastructure element is not present — meter, connection, or equipment removed, displaced, or destroyed beyond visual recognition.
Infrastructure may be present but is obstructed by rubble, fallen material, or post-fire remnants. Crew access and condition assessment pending clearance.
Infrastructure appears undamaged. Site has been cleared of debris. Standard service restoration procedures apply — lower dispatch priority.
From GPS-tagged field photograph to ranked priority list — FieldSight processes images through a reproducible automated workflow built on PyTorch and ArcGIS Pro.
GPS-tagged photos captured via Survey123 at approximately 60 survey points
ArcGIS Pro processes spatial data; photos exported and organized by site
Approximately 20 survey points receive expert ground-truth labels for training
Two binary PyTorch classifiers trained: meter condition and site condition
Confidence metrics combined into a PriorityScore; uncertain predictions flagged for review
Predictions joined to ArcGIS feature class and displayed in Streamlit
Evaluates each meter photograph to classify whether the water meter unit survived or was destroyed — the primary driver of infrastructure replacement priority.
Assesses the surrounding property — whether the site remains in ruins with debris present or has been cleared — informing crew safety and access logistics.
All 21 evaluated sites ranked by combined confidence score. Sites with a destroyed meter receive the highest priority for crew dispatch.
Each marker represents a surveyed address in Altadena, CA. Color encodes priority score — click any marker to inspect site-level predictions.
Five sites received low-confidence predictions from one or both models. These are automatically flagged for manual inspection before dispatch decisions are made.
| Object ID | Meter Prediction | Meter Conf. | Site Prediction | Site Conf. | Priority Score | Status |
|---|---|---|---|---|---|---|
| #619 | destroyed | 83.9% | cleared | 57.6% low | 57 | Needs Review |
| #1071 | damaged | 85.7% | ruins | 61.4% low | 52 | Needs Review |
| #718 | damaged | 70.9% low | ruins | 59.7% low | 52 | Needs Review |
| #596 | damaged | 63.1% low | cleared | 99.7% | 42 | Needs Review |
| #964 | damaged | 75.9% borderline | cleared | 63.4% low | 42 | Needs Review |
Full inference output from the pilot deployment — including APN, service connection length, confidence scores, and priority ranking for every evaluated site.
| # | Address | APN | Conn. Length (ft) | Meter | Meter Conf. | Site | Site Conf. | Review | Priority |
|---|
Interactive Streamlit application with priority map, review queues, and image inspection panels for all 21 evaluated survey points.
Open DashboardFull source: PyTorch classifiers, ArcGIS Pro workflow, Streamlit application, prediction CSV datasets, and holdout results.
View SourceThe narrative context — tracing the Eaton wildfire impact, the collaboration with LFWC, and how this geospatial documentation gave rise to FieldSight.
Read the Story