Evolutionary Neural Architecture Search for Remote Sensing Image Classification.

Area of Science:

• Computer Science
• Remote Sensing
• Artificial Intelligence

Background:

• Remote sensing scene classification is crucial for image analysis.
• Convolutional neural networks (CNNs) show promise but require expert knowledge and extensive trials.
• Automating CNN design is needed to overcome these limitations.

Purpose of the Study:

• To propose a novel neural architecture search (NAS) approach for automated CNN design in remote sensing scene classification.
• To reduce reliance on expert knowledge and extensive trial-and-error in CNN development.

Main Methods:

• An evolutionary algorithm (EA) is used to search for and combine well-structured basic modules for CNN architectures.
• A new population generation strategy enhances search diversity and prevents premature convergence.
• A random forest-based selection mechanism identifies high-quality individuals, reducing computational complexity.

Main Results:

• The proposed NAS approach discovers CNN architectures that outperform state-of-the-art methods on benchmark remote sensing datasets.
• The discovered architectures achieve superior performance with fewer parameters.
• The search process demonstrates a lower computational cost compared to traditional methods.

Conclusions:

• The developed NAS approach effectively automates CNN design for remote sensing scene classification.
• This method offers a more efficient and performant alternative to manual CNN design.
• The findings have significant implications for advancing remote sensing image analysis.