Reservoir Rainfall and Sediment Hotspot Prediction

Machine learning pipeline for river-zone rainfall forecasting, with an optional sediment hotspot prediction module for reservoir monitoring.

The rainfall model is trained from historical Open-Meteo weather data instead of repeated live samples. The sediment module can then consume rainfall predictions alongside satellite imagery, flow, terrain, and land-use features.

Features

• Downloads hourly historical weather data for 14 Indian river zones.
• Builds lag, rolling average, trend, and time-based rainfall features.
• Trains an XGBoost regressor for next-hour rainfall prediction.
• Produces live rainfall predictions from current Open-Meteo forecast data.
• Includes a separate PyTorch-based sediment hotspot package. See SEDIMENT_HOTSPOT.md.

System Architecture

graph TD
    subgraph Data Acquisition
        OM[Open-Meteo API] --> Fetch[fetch_historical_weather.py]
        Fetch --> RawData[(historical_weather.csv)]
    end

    subgraph Data Processing
        RawData --> Eng[feature_engineering.py]
        Eng --> Features[(training_dataset.csv)]
    end

    subgraph Modeling & Prediction
        Features --> Train[train_model.py]
        Train --> Model[rainfall_model.pkl]
        Model --> PredictLive[river_weather_predictions.py]
        OM --> PredictLive
        PredictLive --> Outputs[(rainfall_predictions.csv)]
    end

    subgraph Sediment Module & Dashboard
        Outputs --> Sediment[Sediment Hotspot Prediction <br> PyTorch]
        Sat[Satellite & Terrain Data] --> Sediment
        Sediment --> Dash[Streamlit Dashboard <br> dashboards/]
        PredictLive --> Dash
    end

Loading

Screenshots

Project Layout

reservoir/
├── fetch_historical_weather.py      # Download historical Open-Meteo data
├── feature_engineering.py           # Build ML training features
├── train_model.py                   # Train rainfall model
├── river_weather_predictions.py     # Run live rainfall prediction
├── pipeline.py                      # Orchestrate fetch, feature, train steps
├── sediment_pipeline.py             # Sediment workflow entry point
├── sediment_hotspot/                # Sediment hotspot package
├── scripts/                         # Preprocessing and training helpers
├── dashboards/                      # Streamlit dashboard
├── configs/                         # Example sediment config
└── requirements.txt

Generated data and model outputs are intentionally ignored by Git:

historical_weather.csv
training_dataset.csv
rainfall_model.pkl
rainfall_predictions.csv
data/raw/
data/processed/
data/models/
data/outputs/

Setup

python -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Open-Meteo does not require an API key.

Rainfall Pipeline

Run the full training workflow:

python pipeline.py --all

Or run each stage separately:

python pipeline.py --fetch
python pipeline.py --engineer
python pipeline.py --train

Expected generated files:

historical_weather.csv      # Raw hourly weather data
training_dataset.csv        # Engineered model features
rainfall_model.pkl          # Trained XGBoost model

After training, run live predictions:

python river_weather_predictions.py

This writes rainfall_predictions.csv, which can also be used as an input feature source for sediment hotspot prediction.

River Zones

Krishna:

• K1_Mahabaleshwar - 17.9237, 73.6586
• K2_Sangli_Almatti - 16.8544, 74.5642
• K3_Raichur_Kurnool - 16.2076, 77.3463
• K4_Nagarjuna_Sagar - 16.5750, 79.3167
• K5_Vijayawada_Delta - 16.5062, 80.6480

Narmada:

• N1_Amarkantak - 22.6747, 81.7590
• N2_Jabalpur - 23.1815, 79.9864
• N3_Omkareshwar - 22.2452, 76.1510
• N4_Bharuch - 21.7051, 72.9959

Kaveri:

• C1_Talakaveri - 12.3855, 75.4894
• C2_Kodagu - 12.3375, 75.8069
• C3_Mysuru_KRS - 12.2958, 76.6394
• C4_Mettur_Dam - 11.7870, 77.8008
• C5_Thanjavur_Delta - 10.7867, 79.1378

Model Inputs

Raw weather variables:

• temperature
• humidity
• rainfall
• pressure
• wind_speed
• cloud_cover

Engineered variables:

• 1, 2, and 3 hour lag features for each raw variable.
• 3 hour rolling averages for rainfall, humidity, and temperature.
• Pressure, humidity, temperature, and wind change features.
• Hour, day, and month time features.
• rainfall_next_hour regression target.
• rainfall_class category for no, light, moderate, and heavy rainfall.

Sediment Hotspot Module

The sediment system combines rainfall predictions with:

• Sentinel-2 or Landsat imagery
• River flow and inflow sequences
• DEM terrain features
• Land-use features

See SEDIMENT_HOTSPOT.md for the public data sources, preprocessing commands, training manifest, inference workflow, and dashboard command.

References

• Open-Meteo: https://open-meteo.com/
• XGBoost: https://xgboost.readthedocs.io/
• PyTorch: https://pytorch.org/