Sen2LUCAS Dataset

This repository contains a curated dataset of Sentinel-2 satellite imagery aligned with the Land Use/Cover Area frame Survey (LUCAS) 2018 dataset. LUCAS, an initiative by the European Union, systematically monitors land use and land cover across Europe, providing georeferenced ground-level observations. As part of our work on SenCLIP, a vision-language model designed to bridge the gap between satellite and ground-level perspectives for land-use/land-cover (LULC) mapping, this dataset serves as the foundation for adapting pre-trained vision-language models to remote sensing tasks. By combining Sentinel-2 imagery with LUCAS survey data, the work enables new capabilities for zero-shot LULC classification, improving flexibility and accuracy without requiring labeled satellite data. The LUCAS dataset also provides rich metadata about land use and land cover, which were invaluable for aligning ground-level observations with satellite imagery. Data Retrieval and Processing The data collection and preparation process involved several rigorous steps to ensure its quality and relevance: Geolocation Matching: Sentinel-2 imagery was precisely retrieved based on geolocations provided in the LUCAS 2018 dataset, ensuring spatial alignment with ground-truth survey points. Temporal Consistency: Images were collected from the same months and years as the LUCAS survey to maintain temporal alignment between the satellite imagery and ground-level observations. Cloud Coverage Filtering: To ensure high image quality, scenes were filtered to include only those with less than 10-20% cloud cover, minimizing obstructions for reliable visual and spectral analysis. Sentinel-2 Imagery Specifications Spectral Bands: The dataset contains the Red, Green, and Blue (RGB) bands of Sentinel-2 imagery. Spatial Resolution: Each band is provided at a 10-meter-per-pixel resolution, allowing for detailed analysis of land cover and features. Scene Dimensions: Each image has dimensions of 100 x 100 pixels, representing a 1 km x 1 km ground area. These dimensions are consistent with the granularity of LUCAS geolocations, enabling fine-grained analysis of land-use patterns. By integrating the LUCAS 2018 ground-level observations with high-quality Sentinel-2 imagery, this dataset supports a wide range of geospatial and remote sensing applications. Furthermore, it facilitates the development of multimodal models like SenCLIP, which unlocks new potential for scalable, accurate, and flexible Earth observation tasks.