Object-based random forest classification for detecting plastic-mulched landcover from Gaofen-2 and Landsat-8 OLI fused data

Abstract

Plastic-mulched landcover (PML) is an important type of agricultural landscape and remote sensing is an effective way for monitoring and mapping PML. Based on Gaofen-2 (GF-2) and Landsat-8 operational land imager (OLI) fused data, this study applied an object-based random forest classification (OBRFC) method, which combines object-based image analysis (OBIA) technology with random forest (RF) model, to map PML. The method consists of the following steps: (1) image segmentation with a multiresolution segmentation (MRS) algorithm; (2) selection of sample objects (or segments) and 50 features of index, texture, and shape based on prior knowledge and relevant references; and (3) determination of two particularly important parameters, the number of decision trees-T and the feature number of split nodes -F, by comparing classification accuracy of a series of experiments. The results from applying the OBRFC method on the study area show: 1) the best overall accuracy (OA) of OBARFC reaches 91.73%; 2) by setting T to 50, OA curve presents a downward trend with the highest value of 91.72% at F =5; and 3) by setting F to 5, OA reaches its best value at T = 50.