CLRS

CLRS provides a benchmark database for evaluating continual (lifelong) learning methods on remote sensing image scene classification to assess and mitigate catastrophic forgetting.

Key Features:

• Dataset Composition: The database comprises 15,000 high-resolution remote sensing images across 25 scene classes (airport, bare-land, beach, bridge, commercial areas, desert, farmland, forest, golf-course, highway, industrial zones, meadow, mountain, overpass, park, parking lots, playgrounds, port, railway, railway-station, residential areas, river, runway, stadium, storage-tank), with 600 images per class, uniform size 256×256 pixels, and spatial resolution ranging from 0.26 m to 8.85 m.
• Continual Learning Scenarios: The dataset is structured to support three continual learning scenarios with explicit criteria for partitioning training batches sequentially.
• Database Construction Methodology: Construction leverages a target detection pretrained model to reduce manual annotations during dataset creation.
• Application and Analysis: The database provides baseline results and a benchmark to evaluate and compare mainstream continual learning methods under different scenarios.

Scientific Applications:

• Agriculture: Supports evaluation of continual-learning-based scene classification models applied to agricultural remote sensing data.
• Military operations: Enables assessment of continual learning approaches for scene classification relevant to military remote sensing tasks.
• Environmental monitoring: Facilitates development and benchmarking of continual learning models for environmental monitoring using remote sensing imagery.

Methodology:

Dataset construction leverages a target detection pretrained model to minimize manual annotations, and the database defines criteria to partition training data sequentially into three continual learning scenarios.