This accelerator shows how you can use Visual AI on geospatial data. Instead of deriving numeric features from the georeferenced data, you look at the geospatial data as images. For example, if you have a map of population distribution, instead of extracting the population that corresponds to each row of the main table you can pass the region of the map that corresponds to that row. This provides more information than a raw count of the population would, as it also encodes the distribution within the region (is it uniform or does it concentrate in some areas? what is the shape?, etc.)


The example used to illustrate the approach comes from work done with the Virtue Foundation. As part of the "Data Mapping Initiative", DataRobot has built models to identify suitable locations for new healthcare facilities. By looking at the location of existing hospitals and clinics as a function of several features (road networks, population, terrain, etc.) you find which other areas are suitable in terms of these features (similar to a propensity model).


This work has been peer reviewed and published at the 2023 IEEE International Conference on Imaging Systems and Techniques (IST).


The accelerator is organized as follows:

  1. Dependencies
  2. The data
  3. Pre-processing
  4. Training the model
  5. Scoring
Version history
Last update:
‎02-23-2024 01:51 PM
Updated by: