Rainfall Prediction using Machine Learning

The Indian summer monsoon delivers the overwhelming share of the country's annual rainfall in just four months, and its swings decide farm incomes, reservoir levels and water policy. The India Meteorological Department publishes long district- and subdivision-level rainfall series, and large-scale climate signals such as ENSO (El Niño/La Niña) are known to modulate monsoon strength. Machine learning can exploit these signals: gradient-boosted trees learn non-linear interactions between lagged rainfall, temperature and humidity anomalies and the prevailing climate state. But rainfall is a chaotic system, and intellectual honesty is the core design constraint here — skillful prediction lives at the monthly and seasonal scale, not at "will it rain on Saturday". This project builds a monthly-scale predictor with walk-forward validation and states its predictability limits in the report instead of hiding them.

1. Public IMD monthly/district rainfall series load from data.gov.in; missing months are marked and excluded from supervised rows rather than invented.
2. The feature builder adds lagged rainfall (1/3/6/12 months), month-of-year, region encoding, temperature/humidity/pressure anomalies versus monthly normals, and the season's ENSO phase.
3. A gradient-boosting regressor trains on the feature table to predict monthly rainfall totals; a logistic classifier head predicts the wet-month probability for zero-inflated regions.
4. Validation is walk-forward: the model trains on all years up to Y and is tested on year Y+1, rolling through the record, with skill scored against the climatology baseline.
5. ENSO-conditioned splits re-run the evaluation separately for El Niño, La Niña and neutral years to expose regime-dependent performance.
6. The demo app renders the monsoon board (regional climatology, season progress), the predictor (inputs in, rainfall + probability + driver contributions out), and the climate-analysis view.

Tech stack:

• Python 3.10, scikit-learn (gradient boosting, logistic regression, metrics)
• pandas, NumPy (climate panel wrangling, lag features)
• Matplotlib, Seaborn (climatology charts, validation curves, skill plots)
• Jupyter notebook (buyer-run walk-forward training and evaluation)
• Flask demo app (monsoon board, predictor, analysis views)

• [Monthly rainfall regression] A gradient-boosting regressor predicts expected monthly rainfall (mm) for a chosen region and month from lagged series, anomalies and ENSO phase.
• [Rain/no-rain probability head] A separate classifier predicts P(monthly rain > 75 mm), handling the zero-inflation that breaks pure regression in arid regions.
• [Walk-forward validation] Train on years up to Y, test on Y+1, roll forward — the evaluation mimics real forecasting with no future leakage, and the notebook automates it.
• [Skill vs climatology] Every model must beat the naive "predict the monthly normal" baseline; results are reported as skill scores, not raw RMSE alone.
• [ENSO-conditioned analysis] El Niño, La Niña and neutral years are validated separately, since monsoon teleconnections are regime-dependent.
• [Dry-spell and extreme-event views] A two-state Markov chain estimates dry-spell risk per region, and daily IMD series yield extreme-rainfall-day trends for the analysis section.
• [Driver-contribution display] Each prediction breaks down into climatology, ENSO modulation, anomaly response and persistence, so the report can explain any single forecast.

• Complete source code (data loading, features, training, walk-forward evaluation, demo)
• Jupyter notebook (buyer-run: walk-forward training, metrics, skill plots, analysis)
• Reference trained models from the worked example run
• Project report PDF (background, methodology, validation, honest limits, analysis)
• PPT presentation for final review
• Viva Q&A preparation document (walk-forward, ENSO, skill scores, chaos limits)
• Setup guide (environment, IMD data download, running the notebook, using the demo)

• The system predicts monthly and seasonal tendencies only; day-level rainfall timing is beyond its scope — and beyond machine learning's honest reach without numerical weather models. The report states this boundary explicitly.
• Unprecedented extremes (record-breaking cloudbursts) are out-of-distribution events; the model will underpredict them, and the report documents this failure mode.
• Public IMD series have gaps and station changes over the decades; missing months are excluded, which shortens some regional records.
• ENSO is a modulator, not a switch — El Niño years still contain normal monsoons, so ENSO-conditioned skill varies and is reported per regime.
• This is a research and academic prototype, not an operational forecasting service; it does not replace IMD outlooks for real decisions.