Detecting trees over houses

Aaron Berdanier

Problem

• Trees grow over buildings and create a risk of property damage - liability for real estate, concern for insurance.

Solution

Supervised classification can easily distinguish houses from surrounding vegetation (grass, trees). Detecting buildings automatically in aerial imagery will allow rapid ID of problem structures and properties.

Data

Using Portland, OR (because they have awesome imagery data)

• Aerial imagery:

  • 4-channel (R-G-B-Nir) rasters (convert to numpy array)
  • 6-inch resolution over the whole city
  • each approx. 500MB, already downloaded 80 from 2010 (may download repeat scenes from 2012 for change detection)

• House outlines:

  • shapefile, 1.3GB, outlines of each structure

  • 600K buildings, reduced to ~150K 'houses' for analysis

Workflow

• For each image, identify structures in scene
  • For each structure:
    1. Extract raster data for classification
    2. Calculate aggregate measures

Next steps

• Convert example from R to Python
• Download data from USGS and Portland Open Data
• Generate test and training data, demonstrate separation
• Fit classification model with manual test and training data
• Classify all houses in an image
• Run all images (maybe in parallel)

Future extensions

• Get tax parcel boundaries (but not open for download right now)
• Get building permit data for each parcel
• Get repeat imagery for change detection