UK Habitat Classification Data for Vision-Language Models Using Sentinel-2 and Lidar

A supplementary document for a study on probabilistic vision-language frameworks for habitat mapping. The research integrates Sentinel-2 satellite imagery, UK National lidar data, and textual ecological descriptors to classify habitats under the UKHab standard for Biodiversity Net Gain assessments. The file, authored by Brianna Pickstone and last updated in April 2026, is 760.8 KB in size and shared under a CC-BY-4.0 license.

Use Cases

• Training vision-language models for habitat classification based on integrated spectral, structural, and textual data.
• Evaluating the impact of ecological text detail on model performance for habitats with limited training data.
• Testing the integration of probabilistic priors from existing land cover maps to improve classification accuracy.
• Supporting statutory habitat assessments like Biodiversity Net Gain with automated, interpretable mapping tools.

Strengths

• The associated study reports high model performance, with a top model achieving 0.93 precision, 0.87 recall, and 98.1% overall accuracy.
• Demonstrated performance improvements for specific habitats, such as a 19.36% increase for Wetland classification when using text.
• Integrates multiple data modalities: satellite imagery (Sentinel-2), structural data (UK lidar), and ecological text descriptors.

Limitations

• The dataset is a 760.8 KB DOCX file, indicating a very limited scope, likely containing methodological descriptions and results rather than primary training data.
• Column-level documentation and sample data are unavailable, making direct assessment of data structure and content impossible.
• The use of probabilistic priors from an existing map was shown to produce inconsistent results, including a 16.13% decrease in accuracy for Wetland.

Provenance

Source

figshare, authored by Brianna Pickstone.

Collection Method

Likely contains results and methodology from a research study integrating remote sensing data, ecological text, and prior maps.

Freshness

Last updated 2026-04-29 11:21:40; freshness should be verified

Geography

Focus on United Kingdom habitats, using UK National lidar and Sentinel-2 data.