SATIN aggregates 27 satellite and aerial image datasets covering 6 distinct tasks. The imagery spans 5 orders of magnitude in resolution and contains over 250 distinct class labels. This collection was presented at the ICCV '23 TNGCV Workshop.
Use Cases
- Benchmark zero-shot image classification across the 6 distinct tasks defined in the metadataset.
- Train models on hierarchical land use tasks using the over 250 distinct class labels.
- Analyze performance variations across imagery spanning 5 orders of magnitude in resolution.
- Evaluate model generalization on the globally distributed imagery for complex scenes tasks.
Strengths
- Contains 27 constituent satellite and aerial image datasets.
- Covers 6 distinct tasks including Land Cover, Land Use, and Complex Scenes.
- Includes imagery spanning 5 orders of magnitude in resolution.
- Provides over 250 distinct class labels.
Limitations
- Specific row counts, column details, and file formats are unknown.
- As a metadataset, users must navigate and integrate multiple constituent datasets.
- The license for the aggregated collection and individual datasets is unspecified.
Provenance
- Source
- jonathan-roberts1 on Hugging Face
- Collection Method
- Aggregation of 27 existing satellite and aerial image datasets.
- Time Range
- null
- Freshness
- null
- Geography
- Globally distributed imagery.