5. Findings and discussion
We built a machine learning classifier to detect changes in urban extent and density over time, and applied this to Houston, Texas. Having applied the classifier to satellite images from 1999 to 2015, we see a pattern of expanding urban extent, particularly in the north-east suburbs. Areas such as the Highlands moved from non-urban to light urban, or light urban to dense urban.
The classifier achieves 76% accuracy based on a 6-fold cross-validation. However, error rates are higher between certain category pairs. Water is rarely mis-categorized, as expected given its distinctive reflective signature. However, dense urban is frequently miscategorized as light urban, while light urban is frequently miscategorized as dense urban or non-urban. Our classifier performs better if the two urban categories are grouped together, restricting the task to merely distinguishing built-up area from non-built up. Nevertheless, the findings are encouraging and demonstrate a proof-of-concept. Extensions aimed at raising classification performance further would include:
- Incorporating additional satellite imagery bands, such as Synthetic Aperture Radar imagery (an effective input for water detection) and MODIS thermal imagery (helpful to identify urban areas given their higher thermal reflectivity);
- Adding an NDVI layer through calculations based on the red and near-infrared Landsat bands; and
- Implementing a corner detection algorithm to pick up texture in urban environments, such as multiple dwelling roofs in dense urban areas.