DeepGlobe Satellite Imagery and Semantic Maps at 2448x2448 Resolution

Aggregating 803 satellite image and semantic map pairs for semantic segmentation tasks. It includes 773 training pairs and 30 test pairs, with each image having a resolution of 2448x2448 pixels. The data was contributed by user 'arakesh' and was last updated in April 2022.

Use Cases

• Train a semantic segmentation model on the 773 train_img and train_label pairs to classify land cover from satellite imagery.
• Synthesize new satellite imagery using the provided pix2pixHD-deepglobe-synthesis framework based on the semantic maps.
• Evaluate model performance on the 30 test_img and test_label pairs to assess generalization for land cover mapping.
• Perform instance segmentation by utilizing the instance ids available in the semantic maps to distinguish individual objects.

Strengths

• Contains 803 high-resolution image-label pairs at 2448x2448 pixels.
• Provides a clear split of 773 training and 30 testing samples for model development.
• Includes instance IDs in the semantic maps, enabling both semantic and instance segmentation tasks.

Limitations

• The dataset size of 803 samples is relatively small for training deep learning models from scratch.
• No temporal or geographic metadata is provided, limiting time-series or location-specific analysis.
• The last update was in 2022, which may not reflect current land cover conditions.

Provenance

Source

http://deepglobe.org/

Freshness

Last updated on 2022-04-12.