An all-season sample database for improving land-cover mapping of Africa with two classification schemes

Stable classification with limited sample: transferring a 30-m resolution sample set collected in 2015 to mapping 10-m resolution global land cover in 2017

“…CR FR GR SR WE WB TU IA BL SI Total PA (%) Cropland 1864 262 629 205 2 4 0 33 53 0 3052 61.07 Forest 304 7951 628 455 5 9 48 17 25 1 9443 84.20 Grassland 441 502 4378 632 15 15 111 66 625 4 6789 64.49 Shrubland 203 680 1083 2444 8 7 33 10 346 0 4814 50.77 Wetland 25 22 89 13 30 66 13 1 33 4 can be found in urban classification [8], crop field classification [9], and land cover mapping at the continental scale [10] but have never been tested at the global scale. The experiments presented in Fig.…”

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

“…Landsat 5 images were used as an alternative when the cloud cover of each Landsat 8 images with the same path/row exceeds 50%. This research is an expansion of the all-season sample database development for Africa that was completed in 2014 based primarily on Landsat-8 images collected in 2013 and 2014 [7]. For the remaining of the world, sampling was conducted in 2015 and 2016.…”

“…The training and validation samples were then cross-checked and finalized by the senior author who has more than 5 years' experience in global land cover image interpretation and has led the development of both training and validation sample sets for Africa [7]. The sample dataset has a consistent number of records (Table S3) including the location, time of image acquisition, http://engine.scichina.com/doi/10.1016/j.scib.2017.03.011 spectral reflectance for each season, land cover type for all four seasons for both FROM-GLC and GLC2000 land cover classification systems, as well as level of interpretation uncertainty.…”

“…Both training and validation sample locations were taken from Gong et al [1] except for Africa where over 1000 additional locations were included [7]. Training samples were selected for their representativeness and typicality by an image interpreter.…”

“…When interpreters first selected training samples, they chose the most obvious land cover categories in each Landsat scene representing major land cover types of the imaged area. This was assisted by spectral clustering whose results were examined to screen any dominant spectral clusters that were not picked up in the initial round of training sample selection [7]. In this manner, the representativeness and typicality were preserved by the training sample set.…”