The first all-season sample set for mapping global land cover with Landsat-8 data

Area of Science:

• Earth Observation
• Remote Sensing
• Geospatial Analysis

Background:

• Existing global land cover datasets primarily used single-date training samples, often from the growing season.
• The limitations of single-date training samples for year-round land cover classification remain largely unquantified.

Purpose of the Study:

• To develop and evaluate the first all-season training and validation sample sets for global land cover classification using Landsat-8 data.
• To assess the impact of seasonal variations on land cover classification accuracy.
• To investigate methods for improving global land cover mapping accuracy.

Main Methods:

• Collected Landsat-8 imagery and corresponding training/validation samples across four distinct seasons.
• Employed the Random Forest algorithm to compare classification performance using seasonal and all-season training sample sets.
• Evaluated classification accuracy at global and regional (10°x60° zones) scales.

Main Results:

• Individual seasonal training samples yielded the highest accuracy when validated within the same season.
• Global classification accuracy using combined best seasonal results was 67.2% for 11 Level-1 classes.
• An all-season training set achieved 67.0% global accuracy, while spatially zoned all-season subsamples reached 70.2% accuracy.

Conclusions:

• All-season training sample sets are crucial for achieving optimal and universally applicable global land cover classification.
• Spatially grouping training data can significantly enhance classification accuracy.
• The developed all-season datasets enable more robust and consistent global land cover mapping throughout the year.