Sentinel-2 Land-cover Captioning Dataset

The Sentinel-2 Land-cover Captioning Dataset (S2LCD) is a newly proposed dataset specifically designed for deep learning research on remote sensing image captioning. It comprises 1533 image patches, each of size 224 × 224 pixels, derived from Sentinel-2 L2A images. The dataset ensures a diverse representation of land cover and land use types in temperate regions, including forests, mountains, agricultural lands, and urban areas, each one with varying degrees of human influence.

Each image patch is accompanied by five captions exported in COCO format, resulting in a total of 7665 captions. These captions employ a broad vocabulary that combines natural language and the EAGLES lexicon, ensuring meticulous attention to detail.

The creation of this dataset involved three key activities:

Definition of the Reference Taxonomy: Establishing a language for describing scenes extracted from satellite images. This lexicon aims to balance the use of natural language for describing image content, ensuring caption variety and the incorporation of specific land use and land cover domain terms. The reference taxonomy is defined and proposed by the EAGLE group (EIONET Action Group on Land Monitoring in Europe), which serves as a standard lexicon in Europe for soil monitoring analyses based on satellite data.
Extraction Procedure for Individual Images: Sentinel-2 satellite images are distributed in multi-layer JP2 format, with one file for each spectral band and usually huge dimensions. Thus, a critical step in the dataset construction is to extract individual 224 × 224 RGB image patches from raw Sentinel-2 data. A custom Python tool was developed to this end.
Caption Creation: Annotation of each extracted 224 × 224 RGB image patch, ensuring linguistic variety and usage of domain-specific terms with the EAGLE lexicon was accomplished.

This dataset was introduced in our paper/repository: RSDiX: Lightweight and Data-Efficient VLMs for Remote Sensing through Self-Distillation. Please refer to it for further details on data collection, captioning methodology, and use cases.

References:
Coastal hamlet near forest and water body.
Water body surrounded by buildings and forest.
Small coastal built-up area.
Dense forestry coverage and water body.
Abiotic artificial surface and forest coverage.

References:
Rocky areas with spread vegetation and sporadic snow coverage.
Abiotic natural surface with permanent herbaceous and forest areas near snow coverage.
Dense forest and bare soil.
Wood and herbaceous coverage in mountain context.
Forest area next to grasslands and snow coverage.

References:
Lake surrounded by herbaceous coverage and sporadic trees.
Permanent herbaceous coverage surrounding a water body, also close to artificial abiotic surfaces.
Woody areas among the grasslands near the town.
Small areas of agricultural fields between grasslands with trees in line.
Residential area with trees in line near grasslands and woody areas.

References:
Airport areas and a small water body near extensive and intensive cultivated fields
with the presence of scattered and isolated buildings and a low density hamlet.
Isolated built-up surrounded by cultivated fields near an airport.
Small water body next to an airport.
Scattered buildings and a small hamlet surrounded by cultivated fields.

Dataset Structure

The dataset_s2lcd.json file contains caption annotations for Sentinel-2 image patches. Each entry in the images array corresponds to a single image patch and includes:

filename: The name of the image file (e.g., "S2A_L1C_20150925_N0204R065_2329.tif").
imgid: A unique integer identifier for the image.
sentences: A list of five captions describing the image. Each caption is stored in a dictionary under the key raw, containing a natural language description of land use and land cover features visible in the image.

Example structure:

{
  "images": [
    {
      "filename": "example_image.tif",
      "imgid": 0,
      "sentences": [
        { "raw": "Caption 1" },
        { "raw": "Caption 2" },
        ...
      ]
    },
    ...
  ]
}

Other Sources

You can also access this dataset on Kaggle and on our Github repository.

License

This dataset is released under the CC BY-NC-SA 4.0 license.

image imagewidth (px) 224 224