Rain Prediction Model

A deep learning tensorflow project that predicts rainfall occurrence (rain / no rain) using atmospheric and environmental parameters such as temperature, humidity, wind speed, pressure, UV index, cloud cover, and more.
This repository contains a complete pipeline from preprocessing, model training, evaluation, to metric visualization.

Features

• Full preprocessing workflow with a JSON-based numerization dictionary
• Scalable TensorFlow model architecture
• Binary classification (Rain vs No Rain)
• Performance evaluation:
  • Confusion Matrix
  • Precision, Recall, F1-Score
  • ROC Curve + AUC
• Modular Python scripts (TrainModel.py, EvaluateModel.py)
• Clean and reproducible experiment setup

Dataset

The dataset used in this project was obtained from Kaggle.
It contains weather observations with the following relevant features:

• Temperature
• Humidity
• Wind Speed
• Precipitation
• Cloud Cover
• Pressure
• UV Index
• Season
• Visibility
• Location
• Rain (label: 0 = No Rain, 1 = Rain)

Missing categorical entries are handled through dictionary-based mapping stored in numerization_dict.json.

Training the Model

Run the training script:

python TrainModel.py

This will:

• Load and preprocess the dataset
• Scale numerical features
• Train a binary classifier
• Save the trained model
• Print final loss and accuracy

Example training output:

Model Loss: 0.1070
Model Accuracy: 0.9584

Evaluation

Evaluate the model on the test set:

python EvaluateModel.py

This script generates:

• Confusion Matrix

[[9656  244]
 [ 329 2971]]

• Precision: 0.9241
• Recall: 0.9003
• F1-Score: 0.9120
• ROC Curve + AUC (example: 0.990)

Model Architecture

The model uses a feed-forward neural network with:

• Dense layers
• Mish and swish activations
• Binary crossentropy loss
• Adamax optimizer

Optimized for binary weather classification tasks.

Numerization Dictionary

Categorical features (e.g., Cloud Cover, Season, Location) are converted into numeric form using a mapping loaded from numerization_dict.json. This ensures consistent preprocessing and reproducible results.

Results Summary

Metric	Value
Accuracy	0.9584
Precision	0.9241
Recall	0.9003
F1-Score	0.9120
AUC	0.990

The model demonstrates high performance and strong discriminatory power.

Future Improvements

Potential enhancements:

• Add more meteorological features (dew point, wind direction, solar radiation)
• Time-series forecasting using LSTM/GRU
• Deployment through a Flask or FastAPI service
• Interactive dashboard for predictions
• Automatic hyperparameter tuning (Keras Tuner, Optuna)

License

This project is licensed under the MIT License.