Who Can I Chase

A machine learning project to predict which birds might be observable based on weather conditions, location, and time of year in Andalucía, Spain.

Project Goal

Given a location, date, and weather conditions, predict which bird species are likely to be observable. This could help birders plan their outings by knowing what to expect.

Current Status

This is an early prototype. We have a working model that achieves ~84% recall on detecting bird presence, though precision is low (~15%). In practice, this means the model tends to over-predict (it will suggest birds that may not actually appear), but it catches most of the birds that do show up. There's plenty of room for improvement.

What Works

• Data pipeline from eBird observations through weather data to training
• XGBoost binary classifier for presence/absence prediction
• Basic handling of the severe class imbalance (~22:1 negative to positive)

Known Limitations

• Only covers Andalucía region (due to weather API rate limits)
• Only uses 2025 weather data
• Low precision means many false positives
• No hyperparameter tuning yet
• Model doesn't account for habitat type, elevation, or time of day
• The negative example assumption (species not recorded = not present) is imperfect

Data Sources

• eBird: Bird observation data from the Cornell Lab of Ornithology (2020-2025)
• Open-Meteo: Historical weather data (temperature, humidity, precipitation, wind, cloud cover)

Project Structure

who-can-i-chase/
├── data/
│   ├── raw/                    # Original downloaded files (eBird, weather)
│   └── processed/              # Cleaned, joined datasets
│       ├── unique_coords.csv       # Coordinate pairs for weather fetching
│       ├── weather_andalucia_2025.csv  # Daily weather data
│       └── training_data.csv       # Final ML-ready dataset (~10.6M rows)
├── notebooks/
│   ├── explore_ebird.ipynb     # Initial data exploration
│   ├── merging.ipynb           # Joining bird + weather data
│   ├── prepare_training.ipynb  # Generate negatives, add features
│   └── train_model.ipynb       # Model training and evaluation
├── models/
│   ├── bird_model.pkl          # Trained XGBoost model
│   └── label_encoder.pkl       # Species label encoder
├── src/
│   └── data/
│       ├── fetch_weather.py           # Weather API functions
│       ├── fetch_weather_unattended.py # Batch weather fetcher
│       └── prepare_coords.py          # Coordinate extraction
├── requirements.txt
└── README.md

Data Pipeline

1. eBird data - Download observation data for the region
2. Weather fetching - Get historical weather for each unique coordinate
3. Merging - Join bird observations with weather conditions
4. Training prep - Generate negative examples, add temporal features
5. Training - Train XGBoost classifier

Training Data Details

• 488 species tracked in the dataset
• ~21,700 unique checklists (location + date combinations)
• ~10.6 million rows after generating negative examples
• ~4.3% positive class (bird was seen), ~95.7% negative (not seen)

Features Used

Feature	Description
lat_rounded, lon_rounded	Location (0.1 degree precision)
temperature_2m_mean	Daily mean temperature
relative_humidity_2m_mean	Daily mean humidity
cloud_cover_mean	Daily mean cloud cover
precipitation_sum	Daily precipitation total
rain_sum	Daily rain total
wind_gusts_10m_mean	Daily mean wind gusts
wind_speed_10m_mean	Daily mean wind speed
day_of_year	Day of year (1-365) for seasonality
month	Month (1-12)
species_encoded	Encoded species identifier

Setup

pip install -r requirements.txt

Usage

Fetching Weather Data

1. Prepare the coordinates file:

cd src/data
python prepare_coords.py

2. Run the weather fetcher (this takes a while due to API rate limits):

python fetch_weather_unattended.py

Training the Model

Run through the notebooks in order:

1. explore_ebird.ipynb - Understand the raw data
2. merging.ipynb - Join bird and weather data
3. prepare_training.ipynb - Create training dataset
4. train_model.ipynb - Train and evaluate the model

Making Predictions

import joblib
import pandas as pd

model = joblib.load('models/bird_model.pkl')
le = joblib.load('models/label_encoder.pkl')

# Create input for all species at a given location/weather
conditions = pd.DataFrame({
    'lat_rounded': [37.2] * 488,
    'lon_rounded': [-2.5] * 488,
    'temperature_2m_mean': [12.0] * 488,
    # ... other weather features ...
    'day_of_year': [15] * 488,
    'month': [1] * 488,
    'species_encoded': range(488)
})

probs = model.predict_proba(conditions)[:, 1]
conditions['species'] = le.inverse_transform(conditions['species_encoded'])
conditions['probability'] = probs

# Get top predicted birds
top_birds = conditions[['species', 'probability']].sort_values('probability', ascending=False)

Dependencies

• pandas - Data manipulation
• numpy - Numerical computing
• requests - Weather API calls
• xgboost - Gradient boosting classifier
• scikit-learn - ML utilities
• joblib - Model serialization
• jupyter - Notebooks
• matplotlib, seaborn - Visualization

Future Ideas

• Add more features (habitat, elevation, time of day)
• Try different models or ensemble approaches
• Tune hyperparameters
• Expand to other regions (if API limits allow)
• Build a simple web interface for predictions
• Improve negative example generation (maybe sample proportionally to species rarity)