A global land cover training dataset from 1984 to 2020

Area of Science:

• Earth Science
• Remote Sensing
• Geographic Information Systems (GIS)

Background:

• High-quality training data is crucial for accurate land cover mapping using machine learning on cloud platforms like Google Earth Engine (GEE).
• Collecting this data is traditionally expensive and time-consuming, hindering large-scale land cover analysis.
• Existing datasets may lack comprehensive geographic or temporal coverage.

Purpose of the Study:

• To develop a large-scale, high-quality global training database for land cover and land cover change (LULCC) mapping.
• To improve the efficiency and accuracy of LULCC studies by providing a robust data resource.
• To support diverse applications including agriculture, forestry, hydrology, and urban development.

Main Methods:

• Utilized Google Earth Engine (GEE) and machine learning algorithms for efficient data collection and quality control.
• Sampled spectral-temporal features from Landsat imagery to ensure broad biogeographic representation.
• Incorporated diverse data sources and strategically augmented the database to reflect regional distributions and disturbances.
• Employed machine learning-based cross-validation to identify and remove mis-labeled training units.

Main Results:

• Compiled a comprehensive global database containing nearly 2 million training units.
• The database spans from 1984 to 2020, covering seven primary and nine secondary land cover classes.
• Ensured data quality, biogeographic representation, and relevance for various ecological and anthropogenic studies.

Conclusions:

• The developed global training database significantly reduces the cost and labor associated with LULCC data collection.
• This resource enhances the accuracy and scalability of land cover mapping and change detection.
• The database serves as a valuable asset for a wide range of environmental and developmental research applications.