CLRS: Continual Learning Benchmark for Remote Sensing Image Scene Classification.

Area of Science:

• Computer Science
• Remote Sensing
• Artificial Intelligence

Background:

• Deep learning for remote sensing image scene classification faces catastrophic forgetting, hindering continual learning.
• Existing datasets lack standardized sequential learning benchmarks, limiting algorithm development.

Purpose of the Study:

• To establish criteria for training batches in three continual learning scenarios.
• To introduce the Continual Learning Benchmark for Remote Sensing (CLRS) dataset.
• To facilitate the development of advanced continual learning algorithms for remote sensing.

Main Methods:

• Defined criteria for partitioning datasets into sequential learning batches for three scenarios.
• Proposed a novel method for constructing large-scale remote sensing image classification databases using target detection pretrained models, reducing manual annotation needs.
• Evaluated mainstream continual learning methods across the defined scenarios.

Main Results:

• The CLRS dataset provides a standardized benchmark for continual learning in remote sensing.
• The proposed database construction method reduces annotation effort.
• Analysis of existing methods provides a baseline for future research.

Conclusions:

• The CLRS dataset and proposed construction method are crucial for advancing continual learning in remote sensing image scene classification.
• The benchmark enables rigorous evaluation and development of new algorithms to overcome catastrophic forgetting.