Emily Dorne - Using Satellite Imagery to Identify Harmful Algal Blooms and Protect Public Health

Learn how SciFi uses Sentinel-2 satellite imagery & machine learning to detect harmful algal blooms in lakes, helping water quality managers protect public health more effectively.

Key takeaways
  • SciFi (Cyanobacteria Finder) is an open-source Python package that uses Sentinel-2 satellite imagery and machine learning to detect harmful algal blooms in small inland water bodies

  • The system achieves similar accuracy to existing Sentinel-3 based tools while providing 10x more coverage of lakes across the US due to Sentinel-2’s higher resolution (10-30m vs 300-500m)

  • Decision trees (Light GBM) were chosen as the ML model due to their quick training time, interpretability, and strong performance without requiring GPUs

  • The system filters for water pixels only and uses a 2000m bounding box around sampling points to avoid false positives from land pixels

  • Visual examples and an interactive explorer tool help build user trust and understanding of the model’s predictions compared to raw metrics alone

  • SciFi is designed to augment, not replace, manual sampling by helping water quality managers better allocate resources and prioritize high-risk areas

  • The model performs best at detecting low and high severity cases, which aligns well with the primary use cases of ruling out low-risk areas and identifying urgent situations

  • Model validation using out-of-sample datasets was crucial for identifying when the model was getting correct predictions for wrong reasons

  • The system caches satellite imagery locally and can generate predictions within seconds to support regular monitoring

  • SciFi is not meant for fully automated decisions but rather to support human-in-the-loop workflows for water quality management and public health interventions