Student · Burlingame High School · Bay Area

I’m Daniel Guo. I photograph places from the air, started my high school’s drone club, and research how wildfire shows up in California’s homeless counts.

Bixby Creek Bridge, California, 2025 36.37° N, 121.90° W © Daniel Guo
  1. 01Aerial photography and film 10 frames
  2. 02Wildfire and homeless counts Research
  3. 03Burlingame HS Drone Club Founder

01Aerial

From the air

Coastline, city grids, and open country in California and northwest China, all shot in 2025. Select a frame to see it larger.

02Research

Working draft · updated September 2026

Do wildfires show up in California’s homeless counts?

I matched 18 years of federal homeless counts for California’s 44 regions to CAL FIRE wildfire records, then tested whether AI models could estimate the relationship without seeing the counts.

View the research Code and data
Map of California divided into 44 regions, shaded by 2024 homeless count and overlaid with that year’s fire perimeters in orange. Los Angeles is darkest, at 71,201.
Fig. 1Homeless count by region in 2024, with that year’s fire perimeters. The dashboard version is interactive across every year from 2007 to 2024.
Regions
44HUD Continuums of Care
Years
2007–2024Point-in-Time counts
Fires
1,000+ acresCAL FIRE perimeters
AI models tested
2Claude and Gemini

The question

If an AI model sees only a region’s wildfire history, can it estimate how closely fire tracks the homeless count that HUD’s data actually shows? And if it can’t, are its mistakes random or do they lean one way?

Why this question

It seems obvious that fires which destroy homes would push homelessness up. But homelessness is hard to measure, and the official counts are shaped by how they are defined and collected. I wanted to know whether a model would notice how weak the link is in this data, or fill in the intuitive story.

What I built

A Python pipeline that pulls HUD’s Point-in-Time counts and region boundaries, CAL FIRE perimeters for fires of 1,000 acres or more, and CAL FIRE structure-damage records. It clips everything to California’s border, then assigns each fire’s acreage to the regions it burned, split by area when a fire crosses a boundary.

On top of that is the dashboard: a map built with MapLibre and charts built with D3, laid out as a research report with numbered figures, tables, methods, and downloads.

Decisions

  • Fires are matched to the regions they burned in, not compared as statewide totals.
  • The model gets each region’s name and fire history and nothing else. The counts are withheld.
  • I checked the result several ways: acres burned, structures destroyed, year-by-year correlations, and a fixed-effects model that controls for each region and each year.

What the data shows

Almost nothing.

Pooled across every region and year, the correlation between acres burned and the homeless count is 0.011. For structures destroyed it is 0.003. The fixed-effects estimates are small, positive, and not statistically significant.

What the models did

Every model overstated the relationship. Adding a paragraph that explains what the count measures pulled the estimates down, but no model ever predicted a negative correlation, and none could rank the regions by how strong their real relationship was.

ModelFire history onlyWith a description of the count
Claude Sonnet 50.280.19
Gemini 3.5 Flash-Lite0.130.06
Observed0.050.05

Mean estimate across 41 regions, 10 runs per condition. Observed is the mean of each region’s actual same-year correlation.

Scatter plot of Claude’s estimated correlation against the observed correlation for each region. The points sit mostly between 0.1 and 0.6 regardless of the observed value, and the fit line through them is nearly flat.
Fig. 2Claude’s estimate for each region against the observed correlation. Points on the diagonal would be perfect estimates. The orange fit line is almost flat.

Limits

Point-in-Time counts miss many people displaced by fire. So this does not show that wildfire has no effect on housing loss. It shows that the official count does not pick it up at the regional level.

Status

Working draft. The dashboard, data, and code are public at research.danielguo.xyz.

03Drone club

My school didn’t have a drone club, so I started one.

I wrote the proposal, built a beginner-friendly curriculum, and recruited the first members. The club now has more than 15 members, and we have covered more than five school events.

Role
Founder
Members
15+
Events covered
5+
School
Burlingame High School

04About

I got my first drone at eight.

I have been flying ever since. More than 250 flights later, I still love the shift in perspective: California coastline, Chinese city grids, and a football field under Friday-night lights.

I’m TRUST certified, I edit in DaVinci Resolve, and I post new work on Instagram as @aboveandbelonging.

Kit

A small kit I know well, from flight to final cut.

  • Main camera DJI Mini 4 Pro

    My everyday camera in the air. It is light enough to travel with and capable enough for the landscapes, events, and moving shots on this site. 4K/60fps HDR · 1/1.3" CMOS · 34 min flight · 12.4 mi range · 249 g · Level 5 wind resistance

  • FPV DJI Avata 2

    For tighter lines, low passes, and shots that put you in the middle of the flight. It gives me a more immediate point of view than a traditional drone. 4K/60fps · 1/1.3" CMOS · 23 min flight · 8 mi range (O4) · 377 g · 60 mph top speed

  • Editing DaVinci Resolve

    Where each project comes together: color, sound, pacing, and the small choices that turn footage into a finished story.

  • Ground station MacBook Pro M4

    My editing desk and flight-planning station. It keeps 4K footage moving so I can stay focused on the cut.

05Contact

Say hello.

Have a project, an event, or a good place to fly in mind? I would love to hear about it.

Get in touch