197 — Satellite imagery analysis for operational damage assessment in Emergency situations

Trekin et al (1803.00397)

Read on 05 March 2018
#damage-assessment  #satellite-imagery  #deep-learning  #neural-network  #emergency-response  #mapping  #GIS  #CV 

Speedily determining hotspots of damage can be among the most influential factors in the success of disaster recovery efforts. For example, wildfire damage varies dramatically with geography, and allocating resources equally to all affected zones may not be the most effective response.

Trekin et al propose a deep-learning approach based upon satellite imagery to determine regions heavily affected by a disaster (such as flooding or fire) and map them using common GIS tools. These maps can then be used to triage disaster response efforts by the appropriate commissions or governments.

The technology works by aligning existing satellite imagery with emergency mapping efforts performed by public or private-sector initiatives. Digitalglobe’s Open Data Program provides access to these satellite images with high temporal resolution (which enables high-speed response-time).

Though the image segmentation challenge performed poorly when trained on one fire in one region and then tested on another region with another fire, fine-tuning the model helped ameliorate these weaknesses.

The result is an image segmentation that delimits areas of substantial damage versus regions left undamaged.

As far as I can tell, the code and models are not open-source.