COMPARISON AND ANALYSIS OF THE ACCURACY OF GEE PLATFORM PIXEL-BASED SUPERVISED CLASSIFICATION-TAKING SHANDONG PROVINCE AS AN EXAMPLE
Keywords: Landcover, GEE, PBSC, Shandong, Landsat, GlobeLand30
Abstract. Remote sensing is going through a basic transformation, in which a wide array of data-rich applications is gradually taking the place of methods interpreting one or two imageries. These applications have been greatly facilitated by Google Earth Engine (GEE), which provides both imagery access and a platform for advanced analysis techniques. Within the field of land cover classification, GEE provides the ability to create fast new classifications, particularly at global extents. Despite the role of indices and other ancillary data in classification, GEE platform pixel-based supervised classification (GEE-PBSC), as a relatively fast and common classification method in remote sensing, was not directly analysed and assessed about accuracy in current researches. We ask how high the classification accuracy of GEE-PBSC is, and which type of land cover is more suitable to be classified by GEE-PBSC method with a credible accuracy. Here we adopt GEE-PBSC method to classify Landsat 5 TM imageries in Shandong province in 2010, and compare the result with GlobeLand30 product in 2010 from three aspects: type composition, type confusion and spatial consistency to assess the classification accuracy. Before the comparison, multiple cross-validation, which shows that the overall average test accuracy is about 74%, is required to ensure the reliability. The comparison experiment shows that the spatial consistency ratio of artificial surface, cultivated land and water is about 99.30%, 85.78% and 73.02% respectively. The pixel purity of artificial surface and cultivated land is about 90.26% and 81.45% respectively. The overall spatial consistency ratio is about 82.04%. Although the GEE-PBSC method can achieve high test accuracy, the result is still far from GlobeLand30 product in 2010. Because the GEE-PBSC only uses the pixel information of imageries and does not integrate other multi-source data to assist classification. In addition, classification result also shows that using GEE-PBSC to classify artificial surface and cropland has obvious advantages over other land classes, and their classification results is close to GlobeLand30.