Synergy of multi-temporal polarimetric SAR and optical image satellite for mapping of marsh vegetation using object-based random forest algorithm

• Main Authors: Fan, D. (Author), Fu, B. (Author), Gao, E. (Author), He, H. (Author), Huang, L. (Author), Liu, L. (Author), Sun, J. (Author), Xie, S. (Author), Zuo, P. (Author)
• Format: Article
• Language: English
• Published: Elsevier B.V. 2021
• Subjects: algorithm; Algorithm model; Backscattering; Backscattering coefficient; Backscattering coefficients; China; Decision trees; eigenvalue; Eigenvalues and eigenfunctions; forest; Geometrical optics; Heilongjiang; Honghe Nature Reserve; image classification; Image segmentation; Mapping; mapping method; marsh; Marsh vegetation; Marsh vegetation classification; Multi-scale inheritance segmentation; Parameter estimation; Polarimetric decomposition; Polarimetric decomposition parameter; Polarimetric decomposition parameters; Random forest algorithm; Remote sensing; Sentinel; Spectroscopy; synthetic aperture radar; Synthetic aperture radar; Variable selection; Variables selections; vegetation classification; vegetation index; vegetation mapping; wetland management; Wetland vegetation; Wetlands
• Online Access: View Fulltext in Publisher

 The accurate classification of marsh vegetation is an important prerequisite for wetland management and protection. In this study, the Honghe National Nature Reserve was used as the research area. The VV and VH polarized backscattering coefficients of Sentinel-1B, the polarimetric decomposition parameters of Sentinel-1B, and Sentinel-2A multi-spectral images from June and September were selected to construct 18 multi-dimensional data sets. A highly correlated variable elimination algorithm, a recursive feature elimination variable selection algorithm (RFE-RF), and an optimized random forest algorithm (RF) were used to construct a marsh vegetation identification model. In this study, we searched for an RF model to achieve the accurate classification of marsh vegetation and find the best feature for identifying various types of vegetation. Additionally, the applicability of different optimized RF models to the task of the identification of wetland vegetation and the stability of the identification of marsh vegetation using different classification models were quantitatively analyzed. The results show the following: (1) RFE-RF variable selection and RF parameter optimization can reduce the data dimensionality, improve the accuracy and stability of the wetland vegetation classification model, and achieve a training accuracy of up to 85.39%. (2) The RF model integrating multi-spectral data, backscattering coefficients, and polarimetric decomposition parameters for June and September can obtain the highest overall accuracy (91.16%), and the model has the strongest applicability. (3) The importance of multi-spectral variables in wetland vegetation classification is higher than that of backscattering coefficients and polarimetric decomposition parameters. The visible bands and vegetation index are the most important variables, while the cross-polarized backscattering coefficient (Mean_VH), polarimetric decomposition eigenvalue (Mean_l1, Mean_l2), and calculated eigenvalues of the matrix (Mean_lambda) are the backscattering coefficient features and polarimetric decomposition parameters with the highest contributions. (4) The modified normalized difference water index in June (MNDWI_ Jun), blue band in September (Mean_B_Sep), location feature pixel coordinates (Y_Max_Pxl), and ratio vegetation index in September (RVI_Sep) have the highest contribution to the identification and classification of deep-water marsh vegetation, shallow-water marsh vegetation, forest, and shrubs, respectively. (5) The identification of forest is the strongest, and the classification accuracy for shrubs and deep-water marsh vegetation is greatly affected by the combination of time phase and data sources. © 2021