The migration of training samples towards dynamic global land cover mapping

High quality training samples are essential for global land cover mapping. Traditionally, training samples are collected by field work or via manual interpretation based on high-resolution Google Earth images. Due to the difficulty of training sample collection, regular global land cover mapping is still a challenge. In this study, we developed an automatic training sample migration method based on the first all-season sample set in 2015 and all available archived Landsat 5 TM images in the Google Earth Engine cloud-based platform. By measuring the spectral similarity and spectral distance between the reference spectral and image spectral, we detected and identified the change state of training sample pixels in 2010, 2005, 2000, 1995, and 1990. Overall, 170,925 (66%), 118,586 (64%), 112,092 (67%), 154,931 (63%), and 147,267 (60%) respective training sample pixels were found with no changes over each five-year period. The detection (user's) accuracies of migrated training sample pixels as no change for the first four time periods were 99.25%, 97.65%, 95.03%, and 92.98%, respectively, by comparing with CCI-LC (Climate Change Initiative Land Cover) maps. Classification experiment showed that the migrated training samples can obtain a similar classification accuracy of 71.42% in 2010, when compared to the classification result in 2015 using the same number of training samples. Our study provides a potential solution to resolve the problem of lack of training samples for dynamic global land cover mapping efforts.