NMME Data Training with CNN for SPI Prediction

This project uses Convolutional Neural Networks (CNN) to evaluate various NMME (North American Multi-Model Ensemble) models for predicting the Standardized Precipitation Index (SPI) based on 30 years of historical precipitation data from 1981-2010. The models being evaluated are:

• CanSIPSv2
• CanCM4i
• CanSIPS-IC3
• GEM-NEMO

The aim of this project is to determine which model performs best in predicting SPI using only precipitation data. Evaluation is done using a Taylor diagram and spatio-temporal visualizations of RMSE, standard deviation, and correlation.

Data

• Time Period: 1981-2010 (30 years)
• Data Used: Historical monthly precipitation data from NMME models (CanSIPSv2, CanCM4i, CanSIPS-IC3, GEM-NEMO).
• Target: Standardized Precipitation Index (SPI) calculated from precipitation data.

Objective

To evaluate the performance of multiple climate models for SPI prediction, focusing on:

• Root Mean Square Error (RMSE)
• Standard Deviation (stdev)
• Correlation Coefficient
• Skill Score

Methodology

1. Data Preprocessing:

   • Precipitation data from each model is preprocessed and normalized.
   • SPI values are calculated for each model based on historical precipitation data.

2. Model Training:

   • A CNN is trained on the precipitation data to predict SPI for each model.
   • The model uses spatial-temporal features from the precipitation data for prediction.

3. Evaluation:

   • A Taylor diagram is used to compare the performance of each model in terms of correlation, RMSE, and standard deviation.
   • Spatio-temporal visualizations are generated for the RMSE, standard deviation, and correlation for a detailed comparison.

Tools & Libraries

• Python: Used for data processing and model training.
• TensorFlow/Keras: Deep learning library for building and training the CNN model.
• Numpy & Pandas: Data manipulation and analysis.
• Matplotlib & Seaborn: For visualizations including Taylor diagrams and spatio-temporal plots.
• Basemap: For spatial visualizations.
• NetCDF4: Handling of NMME model data in NetCDF format.

Results

• Taylor Diagrams: Compare model performance in terms of RMSE, standard deviation, and correlation.
• Spatio-Temporal Visualizations: RMSE, standard deviation, correlation, and skill score plotted over time and space to assess model accuracy and behavior across different regions and periods.

Visualizations

1. Taylor Diagram

   

2. RMSE

   

3. Standard Deviation

   

4. Spatial Correlation

   

5. Skill Score

   

Usage

1. Clone the Repository:

   git clone https://github.com/rizkyngrh23/NMME-CNN-training.git
   

2. Install Dependencies

   pip install -r requirements.txt
   

3. Run Data Preprocessing

   python preprocess.py
   

4. Train the CNN Model

   python training.py
   

5. Evaluate the Model using Taylor Diagram

   python taylor_diagram.py