Ensemble learning updating classifier for accurate land cover assessment in tropical cloudy areas

Abstract

Land use/cover information is fundamental for the sustainable management of resources. Notwithstanding the advancement of remote sensing, analysts daunt to generate sufficient-quality land use/cover products due to dense-cloud-contaminated and/or technical issues. This study proposes a novel approach (Ensemble Learning Updating Classifier/ELUC), which can be applied with various classification algorithms and data sets to simplistically generate new classifications or renew existing classifications with a remarkable accuracy improvement. Applying miscellaneous features of Landsat-8 images, the ELUC of a random-forest-based algorithm produces sequences of single-time classifications with a mean overall accuracy of 84%. Through the study period, these sequences of individual classifications were then joined to achieve a final classification which reaches an overall accuracy of 94%. Also, the ELUC of the random-forest-based algorithm outperforms that of Kernel-Density-Estimation with a 5% overall accuracy higher. These outcomes confirm the effectiveness of the ELUC for a remarkably consistent land use/cover estimation with a data-rich environment.