A Synthesizing Land-cover Classification Method Based on Google Earth Engine: A Case Study in Nzhelele and Levhuvu Catchments, South Africa

Abstract

This study designed an approach to derive land-cover in the South Africa with insufficient ground samples, and made a case demonstration in Nzhelele and Levhuvu catchments, South Africa. The method was developed based on an integration of Landsat 8, Sentinel-1, and Shuttle Radar Topography Mission (SRTM) Digital Elevation Model (DEM), and the Google Earth Engine (GEE) platform. Random forest classifier with 300 trees is employed as land-cover classification model. In order to overcome the defect of insufficient ground data, the stratified sampling method was used to generate the training and validation samples from the existing land-cover product. Likewise, in order to recognize different land-cover categories, the percentile and monthly median composites were employed to expand input metrics of random forest classifier. Results showed that the overall accuracy of the land-cover of Nzhelele and Levhuvu catchments, South Africa in 2017–2018 reached to 76.43%. Three important results can be drawn from our research. 1) The participation of Sentinel-1 data can slightly improve overall accuracy of land-cover while its contribution on land-cover classification varied with land types. 2) Under-fitting problem was observed in the training of non-dominant land-cover categories using the random sampling, the stratified sampling method is recommended to make sure the classification accuracy of non-dominant classes. 3) When related reflectance bands participated in the training process, individual Normalized Difference Vegetation index (NDVI), Enhanced Vegetation Index (EVI), Soil Adjusted Vegetation Index (SAVI), Normalized Difference Built-up Index (NDBI) have little effect on final land-cover classification result.