CHALLENGING AI WITH AERIAL & SATELLITE IMAGERY

1: IDENTIFYING AND CLASSIFYING DAMAGE ON AERIAL IMAGERY.

2: DEVELOPING AN ALGORITHM TO FIX ALIGNMENT BETWEEN BUILDING LOCATIONS IN OPEN STREET MAP & BING SATELLITE IMAGERY.

As part of the Netherlands Red Cross, 510 provided the two challenges above for the Hackathon for Peace, Justice, and Security. Multiple international actors provided challenges for the over 120 participants from the fields of AI, Big Data, and Innovation.

Over the 24 hours of the hackathon 7 teams worked on the 2 challenges and found new and innovative approaches to the problems. In which approaches to new CNN architectures as well as retraining of existing networks were considered. Both concepts showed promising results and could make an overview of impact after a disaster more quickly available. We are now looking into possibilities to work with some of the teams to continue the research into their methods and find implementable steps. Below is the outline of the two challenges.

We don’t often participate in Hackathons but we do work on these challenges continuously in our office in The Haage. Want to be a part of our team and solve similar problems? See what skill sets we are looking for and how you can contribute!

CHALLENGE 1: IDENTIFYING AND CLASSIFYING DAMAGE ON AERIAL IMAGERY

Measuring the extent of building damage after a disaster can help humanitarian actors to quickly identify the areas for targeted distributions and people affected. However, doing this assessment manually can be a slow and dangerous process as aid workers are required to go to the affected areas which may still be dangerous or fragile. Remote sensors such as satellite imagery and UAVs are therefore used as alternatives to collect information from disaster areas.

While the remote sensors collect the information, people are currently still being used to analyse it through visual interpretation. Whilst a “Human in the loop” is important in the process depending solely on people makes this method open to mistakes and time pressure.

Identifying and classifying damage has already been tested using two approaches:

A: CNN Convolutional Neural Network approach based on outlines from Open Street Map.

A histogram mapping approach derived from Univariate Image Differencing (As described in this thesis)

 

Expected outcomes:

*Using Computer Vision, detect damage from the pictures directly and classify the level of damage.

*Datasets: geographical subsection of the island of St.Maarten, representative of the damage caused on the island

*All other 34 UAV datasets

*Building outlines from Open Street Map

AND:

B: Interpreting values for the classification of buildings without empirical testing.

In this case, the histogram matching and Univariate Change Detection has already been performed on pre- and post-event imagery and been aggregated on a building level. This has been achieved through the consideration of every layer separately for the RGB and HSV description of the imagery. The results are the median values of these pixels per building.

CHALLENGE 2: DEVELOPING AN ALGORITHM TO FIX ALIGNMENT BETWEEN BUILDING LOCATIONS IN OPEN STREET MAP AND BING SATELLITE IMAGERY

The Red Cross’s Missing Maps project helps map areas where humanitarian organisations are trying to meet the needs of vulnerable people. Missing Maps uses imagery from various sources – mainly from Bing Maps and Digital Globe to plot buildings on Open Street Map. However, these sources of satellite imagery have different geo-references, resulting in misalignment of buildings on the Open Street Map. Misalignment ranges from less than a meter up to dozens of meters. The Red Cross uses Open Street Map data as ground truth for remote sensing techniques and the misalignment therefore creates problems.

Expected outcome: An algorithm that

*Detects misalignment between Open Street Map data and Bing satellite imagery 

*Shifts the Open Street Map building outlines to fit the satellite imagery as accurately as possible.

Datasets:

Training sets deliberately misaligned building footprints, ground truth data, and optical imagery.

 

Comments are closed