Automated (global) land cover mapping Kwame Oppong Hackman Peng Gong, Congcong Li, Le Yu, Jie Wang, Luyan Ji, Huabing Huang, Nicholas Clinton, Yuqi Bai, Greg Biging, Zhiliang Zhu (Tsinghua University, RADI CAS, Google Inc., UC Berkeley, USGS) June 14, 2017, UENR, Sunyani (Ghana).
Aims Encourage use/validation of these products Prevent duplication Collaboration
National, continental, and global EO products Sample selection software
National, continental, and global EO products Global training and test samples Samples spatially sufficient but temporally incomplete Working on making samples temporally sufficient Training samples Test samples Gong et al., 2013, IJRS
National, continental, and global EO products Initial sample summary
National, continental, and global EO products First 30 m resolution global land cover maps (version 1) Gong et al., 2013, IJRS
National, continental, and global EO products Improvements
Download site – 2010 30 m global land cover map http://data.ess.tsinghua.edu.cn (120K downloads from 155 countries)
National, continental, and global EO products Global land cover maps (version 2) Training Validation Dec. 19741 8067 Season 4 84887 35902 Jan. 13596 5724 Feb. 51550 22111 Mar. 7154 3155 Season 1 91034 38325 Apr. 14486 6598 May 69394 28572 Jun. 34589 14299 Season 2 90167 37913 Jul. 20942 8664 Aug. 34636 14950 Sep. 9532 4012 Season 3 77159 32194 Oct. 13933 5908 Nov. 53694 22274 Total 343247 144334 Li CC et al., Science Bulletin, 2017
National, continental, and global EO products Global land cover maps (version 2) Li CC et al., Science Bulletin, 2017
National, continental, and global EO products Global land cover maps (version 2) Classification results – single season vs all year Training season 1 season 2 season 3 season 4 all-season Season_1 67.81% 63.43% 65.17% 64.64% 67.75% Season_2 62.49% 63.97% 61.29% 60.27% 63.77% Season_3 62.88% 60.50% 65.91% 64.14% 65.55% Season_4 67.24% 60.45% 69.68% 70.95% 70.72% All-season 65.02% 62.11% 65.08% 65.03% 67.00% Li CC et al., Science Bulletin, 2017
National, continental, and global EO products Circa 2015 land cover map of Ghana Mapped Classes Ground Truth User’s Acc. (%) W F B O C R S G Water (W) 103 100 Forests (F) 175 27 1 10 82.2 Built-up (B) 87 5 93.5 Orchards (O) 2 233 14 8 88.9 Croplands (C) 16 29 283 19 81.1 Rubber (R) 68 Shrublands (S) 70 289 79.4 Grasslands (G) 37 92.5 Prod. Acc. (%) 98.1 97.2 82.1 80.1 75.5 88.4 Overall Accuracy 85.5% Kappa 0.77 Hackman, Gong, and Wang, 2017
Circa 2014 African land cover maps (Feng et al. 2016) Other products Circa 2014 African land cover maps (Feng et al. 2016) Cropland mapping of Ghana (Yidi Xu – ongoing) Global water products (Luyan Ji – completed & ongoing) Hackman, Gong, and Wang, 2017
Global land cover mapping portal GlobCover GlobalLand30 MOD12Q1 US NLCD FROM-GLC Global Land Cover Mapping Portal Representative land cover data Referenced data GOFC-GOLD Reference Data Portal GLC 2000 database GlobCover 2005 database STEP database VIIRS database GLCNMO 2008 training dataset Urban dataset from the University of Tokyo Crowd sourcing data Land-Cover Geo-Wiki User input User augmentation Uploaded photos FAO Statistics data Web Global Mapping Data Google Earth Bing Maps Satellite map Aerial photos Time-series maps users producers Layman helpers specialists Overall design
Global land cover mapping portal – core functions User Account User registration User profile management Classification scheme Default IGBP FROM-GLC User defined Reference Site Sample creation/loading Sample dataset management Classification task Classification algorithm management Classification task management
Summary for GLC Portal The GLC mapping portal enables an automatic global land cover mapping for anyone to map anywhere. Google Earth image, Landsat data archive and other useful knowledge (ancillary data) for land cover mapping are integrated. Users could customize the classification schemes, build dedicated sample dataset, fulfill unique classifications to meet their specific requirements. Support for multiple classification algorithms has been planned.
Conclusions – land cover mapping Classification is an abstractive exercise, different applications need different classification schemes. Rather than attempting to harmonize global land cover maps with different classification schemes, we suggest that efforts should be spent on harmonizing data. Previous efforts should be better organized so that earlier knowledge can be re-used as much as possible – this leads to the development of a universal sample library for image classification anywhere anytime. Global collaboration will improve global land cover mapping results