Inspiration
In conflict areas, mine are often used to limit troop mobility in a region. As these mines are seldom removed, the number of civilian casualties is considerable over the next years and even decades. Being able to predict where the mines are located would be immensely useful for mine-clearing efforts.
What it does
Our project attempts to predict landmine placement by extracting relevant terrain features.
How I built it
For this project we used open source imagery from the Sentinel 2 satellite constellation, a dataset containing all the landmine incidents reported to NATO from 2004 to 2009 and ground truth information from openstreetmap. We divided our imagery in a 200x200m grid. To train our convolutional neural network we used batches of those small grid images coupled with a binary label (landmine incident, no landmine incident). We can then use our trained model to look at other parts of Afghanistan and predict if the terrain has a high probability of containing landmines.
Challenges I ran into
Working with geospatial data is hard.
Accomplishments that I'm proud of
We actually trained a neural network on this data.
What I learned
Working with different geospatial coordinate system, tif files, ...
What's next for safe-atlas
For each grid image we would like to also feed the neural network information such as distance to waterways, distance to roads, elevation and vegetation. We believe combining these features with the actual image as input of the neural network would improve its performance.
Log in or sign up for Devpost to join the conversation.