Sentinel2GlobalLULC: A Sentinel-2 RGB image tile dataset for global land use/cover mapping with deep learning

Area of Science:

• Earth and Environmental Sciences
• Remote Sensing
• Geographic Information Systems (GIS)

Background:

• Land-Use and Land-Cover (LULC) mapping is crucial for climate modeling and urban planning.
• Existing global LULC products show inconsistencies due to regional inaccuracies and varied LULC classifications.
• Advancements in remote sensing and data processing necessitate improved LULC mapping datasets.

Purpose of the Study:

• To introduce Sentinel2GlobalLULC, a novel RGB image dataset for enhanced LULC mapping.
• To provide a high-quality dataset for training deep learning models for precise LULC classification.
• To address the inconsistencies in current global LULC products.

Main Methods:

• Developed Sentinel2GlobalLULC by achieving spatial-temporal consensus from up to 15 global LULC maps using Google Earth Engine.
• Created 194,877 single-class RGB image tiles (224x224 pixels at 10m resolution) across 29 LULC classes.
• Generated cloud-free composites from Sentinel-2 imagery (June 2015 - October 2020) with comprehensive metadata.

Main Results:

• The Sentinel2GlobalLULC dataset comprises 194,877 image tiles with detailed metadata.
• Each tile represents a specific LULC class with high spatial resolution and temporal coverage.
• The dataset is designed to facilitate the development of accurate and robust LULC mapping models.

Conclusions:

• Sentinel2GlobalLULC offers a valuable resource for improving global and regional LULC mapping accuracy.
• The dataset's design supports the training of deep learning models for advanced LULC classification tasks.
• This work contributes to more consistent and reliable LULC data for diverse applications.