Optimal parameters of random forest for land cover classification with suitable data type and dataset on Google Earth Engine

Abstract

Read online

Exact land cover (LC) map is essential information for understanding the development of human societies and studying the impacts of climate and environmental change. To fulfill this requirement, an optimal parameter of Random Forest (RF) for LC classification with suitable data type and dataset on Google Earth Engine (GEE) was investigated. The research objectives were 1) to examine optimum parameters of RF for LC classification at local scale 2) to classify LC data and assess accuracy in model area (Hefei City), 3) to identify a suitable data type and dataset for LC classification and 4) to validate optimum parameters of RF for LC classification with a suitable data type and dataset in test area (Nanjing City). This study suggests that the suitable data types for LC classification were Sentinel-2 data with auxiliary data. Meanwhile, the suitable dataset for LC classification was monthly and seasonal medians of Sentinel-2, elevation, and nighttime light data. The appropriate values of the number of trees, the variable per split, and the bag fraction for RF were 800, 22, and 0.9, respectively. The overall accuracy (OA) and Kappa index of LC in model area (Hefei City) with suitable dataset was 93.17% and 0.9102. In the meantime, the OA and Kappa index of LC in test area (Nanjing City) was 92.38% and 0.8914. Thus, the developed research methodology can be applied to update LC map where LC changes quickly occur.

Keywords

• land cover classification
• sentinel data
• random forest
• Google Earth Engine
• Hefei City
• Nanjing City