Maxar and the New Zealand Space Agency hosted the GIS AI 24 hour Hackathon. I was drawn to the competition by the synergy with GIS + AI. These are two fields of my interest. Our group had 3 domains of expertise; computer graphics, GIS and machine learning. Over 24 hours, the team produced three separate products. With the dream that they would just plug and play together.
Our computer graphics specialist made a JavaScript website. It visualized the rising sea levels over a landscape as an animation. He relied solely on pixel intensity to convey altitude information. This takes the GIS computation out of rendering map data. It was snappy and fast.
The GIS experts produced a flash 3D visualization of the geospatial data using a Mapbox. It provided another web application to visualize the map layers we were provided in a three-dimensional sandbox. Both of these visualisation tools were intended to display the results of the following machine learning model.
I developed a Custom Object Detection model to recognise houses from satellite imagery. It was a (painfully) hand annotated supervised learning model that used the ImageAI library. In the end, it did not perform well, as it lacked qualitative and quantitative training data. Probably since we forgot to integrate the lidar data into our analysis. Despite our limitations that competitions were good practice and an invaluable learning experience.