Ferdinand Schenck - 🌳 The taller the tree, the harder the fall. Determining tree height from space..

Learn how deep learning and satellite imagery enable scalable tree height measurement for power line safety, replacing costly LIDAR with automated risk assessment.

Key takeaways
  • Vegetation height measurement is crucial for power line infrastructure risk assessment, as trees falling or growing into power lines cause significant outages and fire risks

  • Traditional LIDAR methods for measuring tree height are accurate but expensive, slow, and logistically complex due to requiring aerial platforms

  • Deep learning combined with stereo satellite imagery provides a scalable alternative for measuring vegetation height, using paired images taken from different angles

  • Model training required carefully matched LIDAR ground truth data and satellite imagery pairs from the same timeframe to account for vegetation changes

  • The team used semi-global matching initially but found deep learning approaches performed significantly better for vegetation height estimation

  • Challenges included dealing with semi-transparent vegetation, sparse trees like eucalyptus, and ensuring accurate height measurements for risk assessment

  • Processing requires handling massive datasets (40k x 40k pixel images) and thousands of image pairs for production deployment

  • Model confidence estimation is critical - the system needs to identify when it’s uncertain about height predictions to maintain customer trust

  • Solution enables global scalability without requiring local contractors or manual measurements for each region

  • System classifies trees into risk categories (no risk, low, medium, high) based on height and other factors to help infrastructure operators prioritize maintenance