FLAIR-1-2: Countrywide Land-Cover Semantic Segmentation from Aerial Imagery

The FLAIR dataset contains over 20 billion annotated pixels for land-cover semantic segmentation. It consists of very high-resolution aerial imagery at 0.2 m spatial resolution with 5 channels, including near-infrared and elevation. The data was sampled countrywide over three years and different months by IGNF.

Use Cases

• Training semantic segmentation models based on 19-class land-cover annotations.
• Developing domain adaptation models based on spatio-temporal data collected over three years.
• Analyzing land-use patterns based on high-resolution aerial imagery with near-infrared and elevation channels.
• Benchmarking remote sensing models on a large-scale dataset with over 20 billion annotated pixels.

Strengths

• Contains over 20 billion annotated pixels, indicating a large scale.
• Imagery has a very high spatial resolution of 0.2 meters.
• Data includes 5 channels (RVB-Near Infrared-Elevation), providing multispectral information.
• Sampled countrywide over three years and different months, offering spatio-temporal diversity.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• The full description is hosted externally, requiring a click-through for complete details.
• Data may reflect geographic and temporal bias inherent to the sampling methodology.

Provenance

Source

IGNF

Collection Method

Sampled countrywide from aerial imagery.

Time Range

Acquired over three years and different months.

Freshness

Last updated 2024-10-27 02:13:42

Geography

Countrywide sampling (specific country not stated).