Enabling country-scale land cover mapping with meter-resolution satellite imagery

Area of Science:

• Remote Sensing
• Geospatial Analysis
• Artificial Intelligence in Earth Observation

Background:

• High-resolution satellite imagery offers detailed spatial information crucial for complex land cover classification, especially in built environments.
• Existing land cover mapping efforts face challenges due to complex patterns, high costs of data collection, and significant distribution shifts in satellite data.
• Large-scale, detailed land cover mapping using high-resolution imagery remains underexplored due to these limitations.

Purpose of the Study:

• To introduce the Five-Billion-Pixels, a large-scale land cover dataset comprising over 5 billion labeled pixels from Gaofen-2 satellite images.
• To develop and validate a deep learning-based unsupervised domain adaptation approach for accurate land cover mapping.
• To demonstrate the generalizability of the dataset and method across different satellite sensors and geographical regions.

Main Methods:

• Creation of the Five-Billion-Pixels dataset with 150 Gaofen-2 images, annotated into 24 land cover categories.
• Proposal of an end-to-end Siamese network for unsupervised domain adaptation, featuring dynamic pseudo-label assignment and class balancing.
• Validation using diverse satellite data (PlanetScope, Gaofen-1, Sentinel-2) across multiple cities in China and other Asian countries.

Main Results:

• Successful large-scale land cover mapping over 60,000 km² using entirely unlabeled target domain images.
• Promising results achieved across different sensors (3m, 4m, 8m, 10m resolution) and geographical locations.
• Demonstrated high-quality and detailed land cover classification capabilities applicable nationwide and across Asia.

Conclusions:

• The Five-Billion-Pixels dataset and the proposed unsupervised domain adaptation method effectively address limitations in large-scale, high-resolution land cover mapping.
• The approach enables accurate classification without labeled target data, significantly reducing data collection costs and effort.
• The validated generalizability confirms the utility of the method for diverse remote sensing applications in urban and regional planning.