Can image analysis support LA's unhoused population?
Our team's approach to test if machine learning and aerial imagery can be used to locate unhoused populations in high-risk disaster areas.
In the event of natural disasters such as fires and floods, homeless populations face an increased level of risk.
Homelessness is a growing problem in LA County, and so is the risk of extreme weather events and natural disasters.
From 2019 to 2020 there was a 12.7% increase in countywide homelessness, and these counts were pre-pandemic where many lost jobs and housing.
Climate change has doubled the cumulative area of large wildfires in the western US from 1984 to 2015.
This amplifies the need to decrease the risk that vulnerable populations face by ensuring they can be located and contacted in the event of an emergency. A countywide solution for quickly and accurately locating encampments does not currently exist.
This map from 2019 shows a prevalence of high risk for homelessness across large regions of the county.
Locating vulnerable populations is currently dependent on data such as these general areas of risk and annual point-in-time counts.
Point-in-time counts can help to accurately locate some encampments but are costly, time-consuming, and miss many encampments that are hidden from view from the road.
Pictured are Cal Fire's Fire High Severity Zones (FHSZ).
The Los Angeles Fire Department has conducted field surveys in these regions to attempt to find encampments.
With similar issues as point-in-time counts, surveys were an annual, week-long endeavor requiring personnel from 47 different fire stations - and likely still missed many encampments.
Ethics
Ethics
Although efforts have been made to locate these populations, there is a lack of understanding of exactly where encampments are and how best to reach them in case of a natural disaster.
The Plan
- Data Acquisition, Storage, and Hardware Configuration
- Data Analysis and Script Development
- Methodological Testing and Training
- Map Design and Production
Data
Vexcel Imaging data was used with an API that allowed us to apply different image analysis tools with Python and access anywhere in the County with a set of coordinates.
Frequent revisit time of this aerial imagery and the high spatial resolution were well suited to our project goals.
Methods
Where did we focus our study?
At first glance it may be difficult to locate unhoused populations.
We started with images like this one to help train our model.
and used Vexcel's API to obtain imagery
We attempted to use some basic machine learning algorithms such as DBSCAN a density-based object segmentation approach.
This was strong for object segmentation, but not always distinguishing tents from vegetation.
Then we attempted deep learning using ArcGIS Pro.
In our literature review, deep learning was a recommended method, however as we can see in this example, we struggled to achieve that success. The boxes representing our deep learning results selected many vegetated areas, yet missed some very distinct tents.
This led us to focus on a completely different approach: a focus on spectral information and color.
We converted RGB images to the HSV colorspace and applied various masks and filters.
Then we used computer vision, specifically blob detection, to detect tents.
This combined approach of traditional spectral analysis and machine learning was where we found the magic!
Results
Challenges:
- White Tents
Challenges:
- Shadows
Heat Map
A heat map offers a more effective end product for managers:
- Unhoused populations frequently migrate
- Protects privacy of individuals
- False positive (low precision) do not affect users as significantly
Conclusion
In conclusion, the goal is to use the results to inform policy makers to better direct county resources, communication efforts, and disaster response personnel to the areas of encampments in the event of a life-threatening natural disaster.
