Applied ML · Forecasting · ETL
Avian Forecast
County-level bird-presence forecasting from large multi-source ecological and weather data.
Question
Can we estimate the odds that a bird species is present in a U.S. county on a given day by combining sighting logs, geography, and long-term weather history?
Method
I built an ETL pipeline over roughly 700 GB of structured and unstructured sources (sightings, county shapefiles, ~20 years of weather), hosted weather retrieval with Docker/Open-Meteo, and produced feature-rich training sets for 500+ species (~39M rows). A Random Forest classifier was tuned with 10-fold cross-validation.
Result
The model reached about 0.83 precision and 0.76 F1 for county-level presence odds, giving birdwatchers actionable probability scores rather than raw occurrence dumps.
Limitation
Ecological presence is noisy and biased by observer effort. Weather and geography features help, but rare species and under-sampled counties remain harder; causal claims are out of scope.