| This study takes the mangrove forests within the Pearl Bay area of the Beilun River Estuary National Nature Reserve in Guangxi as an example. It utilizes Sentinel-2 imagery to extract water indices, vegetation indices, red-edge features, texture features, and spatial neighborhood features, in addition to topographic features derived from ASTER GDEM data. The study employs feature correlation analysis for initial multi-feature combination selection, followed by feature importance ranking for optimized multi-feature combination selection. Finally, an object-oriented Support Vector Machine classification method is utilized to extract mangrove distribution information. The experimental results indicate that the use of original bands from Sentinel-2 imagery yields high producer爷s accuracy for mangrove distribution information extraction, but relatively low user爷s accuracy. However, after incorporating red-edge features, texture features, spatial neighborhood features, and topographic features, there is a significant improvement in user爷s accuracy. By conducting correlation analysis on various types of features, data redundancy can be reduced while simultaneously enhancing the accuracy of mangrove extraction. Furthermore, by ranking the importance of different types of features, we can quantitatively analyze the contribution of each type to mangrove extraction. This allows for the selection of more accurate feature combinations, leading to better mangrove extraction results with reduced data redundancy. The findings of this research are of significant reference value for the use of Sentinel-2 imagery and multiple features in mangrove extraction. |
