roshantariq/Nowcasting-Convective-Rainfall-in-West-Africa-using-ConvLSTM
Domaine:
climateType de record:
project
Créateur:
ros
Hôte:
A deep learning project focused on developing ConvLSTM models to predict convective rainfall events in the West African region. This model combines spatio-temporal data from satellite imagery to provide accurate, real-time weather predictions. The repository contains code for model training, evaluation, and visualization of predictions.
# Nowcasting Convective Rainfall in West Africa
## Overview
This project focuses on building a ConvLSTM model for nowcasting convective rainfall in the West African region. The goal is to predict short-term rainfall intensity based on satellite imagery, which is crucial for timely weather forecasting, especially in regions prone to rapid storm formation and where ground-based weather infrastructure is limited.
## Aim
The main objective is to accurately predict rainfall intensity at short intervals (up to 2 hours) using ConvLSTM, leveraging the spatiotemporal dependencies in satellite data to enhance real-time forecasting.
## Methodology
### Model Architecture:
__Input__: Sequential satellite imagery frames representing rainfall intensity over time.
__Layers__:
- ConvLSTM layers to capture spatiotemporal features.
- Conv3D layers to refine and generate predictions for the next rainfall frame.
__Output__: Predicted rainfall intensity for future frames.
__Loss Function__: Mean Squared Error (MSE) - This loss function was chosen to minimize the squared difference between the predicted and actual rainfall intensities, making it ideal for continuous prediction tasks.
__Evaluation Metrics__:
- Fractional Skill Score (FSS): Measures spatial accuracy across various window sizes (3x3, 5x5, 10x10).
- Accuracy & AUC: Overall performance metrics assessed using accuracy and Area Under the Curve (AUC) via ROC curves.
__Data__:
Satellite-derived rainfall imagery specific to West Africa’s mesoscale convective systems is used as input data for nowcasting.
__Computational Resources__:
The ARC4 high-performance computing cluster was used to train the ConvLSTM models, offering the computational power necessary for large-scale data processing and deep learning model training.
__Model Configurations__:
Four different ConvLSTM architectures were designed and tested, each with varying input and output configurations:
- Model 1:
Input: Frames 1, 2, 3.
Output: Frames 4 and 5.
- Mode …
Similaires
Nowcasting convective cores using deep learningNowcasting Flood Impacts of Convective storms in the SahelUsing land surface information to improve nowcasts of mesoscale convective systems in West AfricaPreparing for rainfall nowcasting in the GOES-R era using SEVIRI and polar-orbiting microwave data over AfricaRelation between Convective Rainfall Properties and Antecedent Soil Moisture Heterogeneity Conditions in North AfricaHybrid RF–ConvLSTM Approach for Rainfall Estimation from MSG Data over Northern Algeria