Indian Monsoon Rainfall Analysis and Forecasting Tool 🌧️

This interactive web application provides analysis of historical Indian monsoon rainfall data (June-September) for various meteorological subdivisions and generates future rainfall forecasts using machine learning (Random Forest) and statistical (Auto-ARIMA) models.

Live Application: https://rainfall-forecaster.streamlit.app/

Table of Contents

• Overview
• Features
• Data Source
• Methodology
  • Historical Analysis
  • Forecasting Models
• Technologies Used
• Project Structure
• Setup and Usage
• Screenshots
• Future Work
• Contributing
• License
• Contact

Overview

The Indian monsoon is a critical climatic phenomenon, and understanding its patterns and predictability is vital. This tool was developed as a personal learning project to explore the application of data analysis and AI/ML techniques to meteorological datasets. It allows users to:

• Visualize historical rainfall trends for different Indian subdivisions.
• Examine rainfall distributions and key statistics.
• Generate short-term forecasts for total monsoon rainfall using different predictive models.

This project was inspired by the challenges and advancements in Numerical Weather Prediction (NWP) and the potential of AI/ML to enhance forecasting capabilities.

Features

• Interactive Data Exploration:
  • Select from various Indian meteorological subdivisions.
  • Filter historical data by a specific year range.
  • View time series plots for total monsoon (JUN-SEP) and monthly (JUN, JUL, AUG, SEP) rainfall.
  • Analyze rainfall distributions using box plots and histograms.
  • Access summary statistics (mean, median, std. dev., min, max) for selected data.
• Rainfall Forecasting:
  • Choose between two forecasting models:
    1. Random Forest Regressor: A machine learning model using lagged rainfall values as features.
    2. Auto-ARIMA: A statistical model that automatically determines the best ARIMA(p,d,q) orders.
  • Specify the number of future years to forecast.
  • For Random Forest, configure the number of past years (lags) to use as predictors.
  • View forecasted values in a table and plotted alongside historical data.
  • For Auto-ARIMA, view confidence intervals for forecasts.
• Model Details:
  • View key parameters and metrics for the selected forecasting model (e.g., OOB score for Random Forest, ARIMA order for Auto-ARIMA).
• User-Friendly Interface: Built with Streamlit for easy interaction.

Data Source

• Dataset: Sub-divisional monsoon rainfall data for India.
• Source: Aggregated from India Meteorological Department (IMD) historical data archives.
• Period Covered: 1901-2021.
• Parameters Used: Monthly rainfall for June, July, August, September, and the total June-September (JUN-SEP) monsoon rainfall for 36 meteorological subdivisions.
• The data file (data.csv) is included in this repository.

Methodology

Historical Analysis

Historical rainfall data is analyzed using Pandas for data manipulation. Interactive visualizations are generated using Plotly Express, including:

• Line charts for rainfall trends over time.
• Box plots to show the distribution and variability of rainfall for each monsoon month and the total season.
• Histograms to display the frequency distribution of total monsoon rainfall.

Forecasting Models

1. Random Forest Regressor

• An ensemble machine learning model from scikit-learn.
• Features: Lagged values of past 'JUN-SEP' total monsoon rainfall (e.g., rainfall from year T-1, T-2, ..., T-n_lags). The number of lags is user-configurable.
• Process: The model is trained on the complete historical data of the selected subdivision. It then predicts future rainfall autoregressively (using its own past predictions as input for subsequent years' forecasts).

2. Auto-ARIMA

• An automated approach to fit Autoregressive Integrated Moving Average (ARIMA) models using the pmdarima library.
• Process: pmdarima.auto_arima automatically searches for the optimal non-seasonal ARIMA(p,d,q) order based on information criteria (e.g., AIC).
• The model is fitted to the historical 'JUN-SEP' rainfall series for the selected subdivision.
• Forecasts are generated along with 95% confidence intervals.

Technologies Used

• Python: Core programming language.
• Streamlit: For building the interactive web application.
• Pandas: For data manipulation and analysis.
• NumPy: For numerical operations.
• Scikit-learn: For the Random Forest machine learning model.
• Pmdarima: For the Auto-ARIMA forecasting model.
• Plotly & Plotly Express: For creating interactive charts and graphs.
• Git & GitHub: For version control and hosting.

Project Structure

Setup and Usage

To run this application locally:

1. Clone the repository:

   git clone https://github.com/pyandcpp-coder/Rainfall-Forecaster.git
   cd Rainfall-Forecaster

2. Create and activate a virtual environment (recommended):

   python -m venv venv
   source venv/bin/activate  # On Windows: venv\Scripts\activate

3. Install the required packages:

   pip install -r requirements.txt

4. Run the Streamlit application:

   streamlit run app.py

   The application should open in your default web browser.

Future Work

Potential enhancements for this project could include:

• Incorporating more sophisticated time series models (e.g., SARIMAX, Prophet with exogenous variables).
• Adding more detailed statistical tests for stationarity, seasonality, etc.
• Allowing users to upload their own time series data.
• Implementing model evaluation metrics (e.g., RMSE, MAE) on a hold-out set.
• Exploring the integration of climate indices (e.g., ENSO) as exogenous variables for forecasting.
• Improving UI/UX and adding more customization options for plots.

Contributing

This is a personal project, but feedback and suggestions are welcome. Please feel free to open an issue or submit a pull request if you have ideas for improvements.

Contact

Yash Tiwari

• Email: yashtiwari9182@gmail.com
• LinkedIn: https://www.linkedin.com/in/yrevash/
• GitHub: https://github.com/pyandcpp-coder