Using spatial video and deep learning for automated mapping of ground-level context in relief camps

Area of Science:

• Geospatial analysis
• Artificial intelligence
• Disaster response

Background:

• Relief camps pose health risks due to overcrowding and poor sanitation.
• Dynamic changes in camp size and services challenge traditional mapping.
• Geospatial data is crucial but faces granularity and sustainability issues.

Purpose of the Study:

• To develop a deep learning-based solution for dynamic mapping of relief camps.
• To address challenges in spatial data collection for rapidly changing environments.
• To enable automated mapping using spatial video.

Main Methods:

• Trained a convolutional neural network (CNN) on spatial video (SV) data from Goma, DRC.
• Implemented a spatial filtering approach to enhance object tagging accuracy.
• Utilized raster math for longitudinal analysis of camp distribution changes.

Main Results:

• The CNN model achieved high precision and recall in detecting temporary dwellings from SV imagery.
• Spatial filtering identified camp concentrations and enabled exploration via a web tool.
• Longitudinal analysis revealed significant spatiotemporal changes in tent distribution.

Conclusions:

• The study provides a foundation for automated spatial feature mapping from imagery.
• This work supports sustainable data generation for informal settlements and relief camps.
• Future integration of SV, object identification, and mapping promises enhanced situational awareness.