CyFi

Cyanobacteria are microscopic algae that multiply fast in warm water rich in nutrients, often turning a lake blue or green.² Some make toxins that poison people, pets and livestock.² Water managers usually find them by going out and sampling. That is accurate but slow, and hard to do often.² CyFi, short for Cyanobacteria Finder, gives an estimate from satellite imagery in seconds, so a manager can decide where to sample first and when a health warning may be needed.^1,2

It comes from DrivenData, which ran a machine learning competition called Tick Tick Bloom to detect cyanobacteria blooms in small inland water bodies and classify how severe they are.^1,2 The labels were samples collected on the water by many organisations across the United States.^1,2 CyFi's model is built from the winning entries, reworked to generalise better and run faster.^1,2 The method is written up in the 2024 proceedings of the Python in Science Conference by E. Dorne, K. Wetstone, T. B. Cerquera and S. Gupta.^1,2

You install it with pip or conda and give it a CSV of latitude, longitude and date, or a single point at the command line.¹ For each point it looks for the most recent Sentinel-2 image from the 30 days up to that date with under 5 per cent cloud in a 2 kilometre box around the point.¹ Keeping only the water pixels, it computes summary statistics and colour ratios across the image bands, then adds a land cover class from the European Space Agency's land cover map.¹ A gradient-boosted decision tree model (LightGBM) turns these into a density in cells per millilitre.¹ The imagery comes through Microsoft's Planetary Computer.¹

Each estimate also carries a severity level from World Health Organization thresholds:^1,3

• low, under 20,000 cells per millilitre
• moderate, 20,000 to 100,000
• high, over 100,000^1,3

Users are free to apply their own thresholds instead.^1,3 A companion viewer, the CyFi Explorer, shows the satellite image behind each estimate.^1,3

For a model of a watershed, CyFi puts a bloom estimate on small lakes and reservoirs, working from Sentinel-2 bands as sharp as 10 metres.¹ NOAA's Harmful Algal Bloom Monitoring System and the Cyanobacteria Assessment Network do similar work from Sentinel-3 imagery at 300 metres.² Its training labels all came from the United States.^1,2 The code is MIT licensed, and version 1.2.0 was released in October 2026.^3,4