FEDGEODAY 2023 | Lightning Talk - Todd Bacastow

Summary

Todd Bacastow presents updates on SpaceNet, a collaboration focused on open-source machine learning for geospatial applications. SpaceNet 8, concluded in August 2022, addressed flooded buildings and roads using pre and post-event imagery, introducing attribution of features. The challenge saw a 1.5x improvement over the baseline, employing techniques like data preprocessing, ensemble neural networks, and rules-based postprocessing. SpaceNet 9, launching in winter, aims to enhance pre-processing techniques, focusing on aligning tie points for optical imagery and SAR data to improve downstream analytics for object identification.

Highlights

🛰️ SpaceNet 8 successfully identified flooded buildings and roads with a 1.5x improvement over the baseline.

🌐 SpaceNet, a collaboration founded in 2016, accelerates open-source machine learning for geospatial applications.

🚀 SpaceNet 9, launching in winter, focuses on advancing pre-processing techniques for optical imagery and SAR data.

🏗️ SpaceNet's four pillars include providing labeled training datasets, running prize challenges, open-sourcing winning algorithms, and publishing evaluation metrics.

🌍 SpaceNet aims to improve data wrangling capabilities, particularly in disaster response scenarios, using computer vision for multisource data alignment.

🏆 Prize challenges contribute to innovation, with winning algorithms and datasets available to the community, fostering continuous improvement.

📈 SpaceNet's evolution spans foundational mapping challenges, progressing from building footprints to more complex tasks like identifying flooded areas.