Building Population Estimation in Boromata Village

Deep Learning Building classification and population estimation for Emergency Response.

Background

Due to the recent attacks in Boromata village in Central African Republic, there is need to estimate the number of buildings that existed in the area prior to the attack to ensure efficient Emergency Response by the ICRC Team.

Data and Method

Study Area

The area of interest (AOI) situated in the Northern part of the Central African Republic (Vakaga), covering approximately 155 Hectares.

Area of Interest

A Natural Colour (RGB) satellite Imagery at 43cm spatial resolution from  DigitalGlobe  - GEO1 was downloaded for the date 19-January-2018 (before the event). Imagery for the study area was clipped from source data and exported into TIFF format.

© 2018 Digital Globe

Workflow and Software

We used Deep Learning tools in ArcGIS Pro to classifying building based on roof shape in this region and estimate their count. The building were places into two categories i.e. Huts (round shaped) and Rectangular (all other regular shapes). Some Building outlines within the study area were manually generated to train the Algorithm.

Results

Buildings were classified as either Huts or rectangular Based on their aerial view shape as seen from the Satellite image. The Rectangular buildings consisted of both with four corners or more. Of the buildings analyzed and classified in the study area, 33.9% were classified as Huts with the remaining 66.1% classified as Rectangular.

Of those classified as Rectangular, 1 was identified as a mosque, 1 as a School and 44 belonging to the Market Place.

Results

After testing numerous iterations of rulesets and Deep Learning Algorithms in ArcGIS Pro, an overall accuracy of 95% was the best that could be produced for this project. The possible sources of error include mis-classification of buildings in incorrect class. The results were visually inspected for validation and correction to fill buildings that were not classified.

Building Classifications

Image Analysis

GIsupport@icrc.org

GEO1 Satellite Image

© 2018 Digital Globe

Results