Meta-Album RESISC: 45-Class Remote Sensing Scene Images

A collection of 1,800 satellite and aerial images across 45 distinct scene categories, preprocessed to a uniform size of 128x128 pixels. The dataset is derived from the NWPU-RESISC45 benchmark, designed by Gong Cheng, Junwei Han, and Xiaoqiang Lu to present a challenging classification task with high within-class diversity and between-class similarity. It was adapted for the Meta-Album few-shot learning benchmark by Phan Anh VU in March 2022.

Use Cases

• Benchmarking few-shot image classification algorithms based on the 45 scene categories.
• Training models for aerial and satellite image scene recognition based on the described within-class diversity.
• Developing land cover classification systems using the provided remote sensing imagery.
• Studying model generalization on challenging datasets with high between-class similarity as noted in the description.

Strengths

• Contains 1,800 images, providing a standardized scale for few-shot learning experiments.
• Offers 45 distinct scene categories, enabling multi-class classification tasks.
• Images are preprocessed to a consistent 128x128 pixel resolution, simplifying input pipelines.
• Derived from a well-cited benchmark (NWPU-RESISC45) designed for challenging classification.

Limitations

• The dataset is a mini subset; the original RESISC45 contains 31,500 images, suggesting this version is a smaller sample.
• Image resolution has been reduced to 128x128 pixels from the original 256x256, which may limit detail for some applications.
• Description metadata is limited; actual data quality and distribution require manual inspection after download.

Provenance

Source

NWPU-RESISC45 Dataset, created by Gong Cheng, Junwei Han, and Xiaoqiang Lu.

Collection Method

The original dataset gathered 700 RGB images per category from satellite and aerial imagery; this version was resized and prepared for Meta-Album.