Dataset Card for Sentinel-2 Global Land Cover (S2GLC) by Preligens

This dataset is a modified version of the Sentinel-2 Global Land Cover (S2GLC) 2017 dataset, delivered with a 10-meter spatial resolution and a thematic legend composed of 10 land cover classes. The dataset contains 256x256 pixel images with four bands covering the visible spectrum as well as near-infrared (R-G-B-NIR). It is designed for use in land cover classification tasks and other remote sensing applications.

Dataset Details

Dataset Description

The Sentinel-2 Global Land Cover (S2GLC) dataset was initially developed by the European Space Agency and has been modified by Preligens. It provides high-resolution (10m) satellite imagery and corresponding land cover labels. The dataset contains 256x256 pixel images of four spectral bands: Blue, Green, Red, and Near-Infrared (B, G, R, NIR). These images are extracted from larger Sentinel-2 images covering the European continent.

The dataset includes a segmentation mask for every image, where each pixel is labeled with one of the ten land cover classes:

• no_data (0): Pixels with missing annotations (not present in the dataset but has a dedicated channel in ground-truth masks).
• clouds (1): Very rare class, not informative, often equivalent to “no_data.”
• artificial surfaces and construction (2)
• cultivated areas (3)
• broadleaf tree cover (4)
• coniferous tree cover (5)
• herbaceous vegetation (6)
• natural material surfaces (7)
• permanent snow-covered surfaces (8)
• water bodies (9)

The dataset includes strong class imbalances, with certain classes such as "cultivated areas" being dominant and others like "permanent snow-covered surfaces" being very rare.

• Curated by: Preligens
• Funded by: European Space Agency (ESA)
• Shared by: Preligens
• License: apache-2.0

Dataset Sources

• Repository: https://github.com/preligens-lab/challenge-ens/tree/master

Uses

Direct Use

This dataset is primarily intended for land cover classification tasks using machine learning models, especially deep learning models that can benefit from high-resolution satellite imagery. It is suitable for training, validation, and testing of models aimed at segmenting and classifying land cover types.

Out-of-Scope Use

The dataset is not suitable for tasks requiring precise temporal alignment between imagery and land cover labels due to potential time discrepancies between image acquisition and land cover ground truth. It should also not be used in applications where the class imbalance might lead to biased outcomes without proper handling.

Dataset Structure

The dataset comprises 256x256 pixel images in 16-bit TIFF format for the input data and 8-bit TIFF format for the land cover masks. The input images contain four spectral bands in the following order: Blue, Green, Red, Near-Infrared (B, G, R, NIR). Each pixel in the segmentation mask is encoded as an integer representing one of the ten land cover classes.

Dataset Creation

Curation Rationale

The dataset was created to provide a high-resolution, labeled dataset for training and evaluating land cover classification models. The modifications made by Preligens aim to enhance the dataset's usability in machine learning applications.

Source Data

Data Collection and Processing

The source data was collected from the Sentinel-2 satellite, part of the European Space Agency's Copernicus program. The images were selected and processed to cover the European continent with a focus on land cover during the summer months, maximizing the presence of specific land cover types like permanent snow-covered surfaces.

Who are the source data producers?

The original data was produced by the European Space Agency (ESA) as part of the Sentinel-2 Global Land Cover project. The dataset has been modified and curated by Preligens.

Annotations

Annotation process

The land cover labels were inherited from the original S2GLC dataset, which involved manual and automated processes to classify land cover types based on satellite imagery.

Who are the annotators?

The annotators include remote sensing experts and data scientists from the European Space Agency.

Personal and Sensitive Information

The dataset does not contain personal, sensitive, or private information as it consists of satellite imagery and land cover classifications.

Bias, Risks, and Limitations

The dataset contains significant class imbalances, with some land cover types being overrepresented while others are underrepresented. Users should employ appropriate techniques to address this imbalance in their models.

Recommendations

Users should be aware of the class imbalance and consider techniques such as class weighting, oversampling, or using advanced loss functions to mitigate the impact on model performance.