A study of cloud classification with neural networks using spectral and textural features

Area of Science:

• Remote Sensing
• Artificial Intelligence
• Meteorology

Background:

• Accurate cloud classification from satellite imagery is crucial for weather forecasting and climate monitoring.
• Traditional methods often struggle with the complexity and variability of cloud data.

Purpose of the Study:

• To evaluate the effectiveness of neural networks for satellite cloud data classification.
• To compare different feature extraction techniques and neural network paradigms.

Main Methods:

• Utilized singular value decomposition (SVD) and wavelet packet (WP) for spectral and textural feature extraction.
• Compared SVD/WP with gray-level cooccurrence matrix (GLCM) and spectral features.
• Implemented and benchmarked probability neural network (PNN) and Kohonen self-organized feature map (SOM) on GOES-8 data.
• Developed a postprocessing scheme using contextual information to enhance classification accuracy.

Main Results:

• The probability neural network (PNN) demonstrated strong performance in cloud classification.
• Feature extraction methods like SVD and WP effectively captured salient cloud characteristics.
• The developed postprocessing scheme significantly improved classification accuracy.

Conclusions:

• Neural network-based approaches, particularly PNN with advanced feature extraction and postprocessing, offer a promising solution for satellite cloud classification.
• The integration of spectral and textural features enhances the system's ability to differentiate cloud types.