Building data –

Building data

Building data

At the core of is an artificial intelligence pipeline that analyses, detects and classifies buildings on satellite images. We always compare at least two images of the same areas at two different points in time. Thereby, we can provide you insights on a per hectare level about how the hectares, wards and constituencies / LGAs have changed and are changing.

Precision, recall and the F-score are metrics to measure how well an AI model performs. More background about these concepts is available on here.

The building data is produced using a multi-step AI pipeline. It contains two major steps:

  1. First, buildings are detected by processing satellite images. The goal of this step is to identify buildings on a satellite image. Besides the underlying trained detector, the results also depend on the specific quality of the satellite image (lighting, clarity, etc.).
  2. Second, all detected buildings are classified into the different building types and the roof colors are detected.
In the first step, the counting recall measure is important as tries to identify most buildings present on an image. The average counting recall measure is between 88 and 94 per cent.
The first step sometimes detects objects that are actually not buildings. In the second step, each detected object is run against a classification model to detect the type of building and its roof color. The precision of this model is between 87 and 92 per cent on average.
The F-score combines recall and precision and is around 90% for the building data. For each individual city, it will be slightly different depending on how the models perform with a specific image.

The annual growth rates are calculated on a per km², administration 2 level (e.g. constituency, Local Government Area, Department) and administration 3 level (e.g. ward,  council, communes). Depending on the dashboard, the different growth rates are available. We always compare the number of buildings on two different point of times and compute the annual growth rates.