Mangrove species classification with combination of WorldView-2 and Zhuhai-1 satellite images

Abstract

Remotely sensed classification of mangrove species is affected by image resolution, spectral information, classification strategy, and image feature selection methods. The present studies of mangrove species classification using remote sensing mostly focus on comparison of classification accuracy, and few of them discuss the spatial pattern of species distribution and the corresponding influencing factors. The combination of high-resolution and hyperspectral satellite images in species classification of mangrove forest has received less attention. With WorldView-2 and Zhuhai-1 images in Gaoqiao mangrove Reserve, this study aims to compare the effects of different feature selection methods (XGBoost, eXtreme gradient boosting; ERT, extremely randomized trees; SPA, successive projections algorithm) and different image resolutions (the WorldView-2 image with a resolution of 0.5 m was resampled to 1, 2, 4, 8, and 10 m) on the classification accuracy of mangrove species based on random forest classification model and to explore the spatial pattern of mangrove distribution and the corresponding influencing factors based on the coupling of WorldView-2 and Zhuhai-1 images. With each spatial resolution of WorldView-2 image, 248 features were extracted, including 52 spectral features (eight spectral bands, 38 vegetation indices, three principal component bands, and three tasseled cap transformation bands) and 196 texture features (seven windows of 3×3, 5×5, 7×7, 9×9, 11×11, and 15×15; for each window, 28 texture features were extracted). With Zhuhai-1 hyperspectral image, 117 spectral features (32 original spectral bands, 32 first derivative bands, 47 vegetation indices, three principal component bands, and three tasseled cap transformation bands) were extracted. Results showed that XGBoost was superior to ERT and SPA, which had great advantage in image feature selection. Among the six types of WorldView-2 image resolution, the 2 m resolution was optimal for species classification, and the red edge band (705—745 nm) played an important role in species classification. The coupling of WorldView-2 and Zhuhai-1 images (resolution: 2 m, overall accuracy: 88.98%, kappa coefficient: 0.846) had better performance than using single WorldView resolution: 2 m, overall accuracy: 83.47%, kappa coefficient: 0.768 and Zhuhai-1 image (resolution: 10 m, overall accuracy: 78.50%, kappa coefficient: 0.703). The classification map based on the coupled image features illustrated that the area of Aegiceras corniculatum accounted for the largest proportion (33.77 % ), followed by Bruguiera gymnorrhiza (30.44%), Avicennia marina (26.96%), Bruguiera gymnorrhiza (6.08%), Sonneratia apetala (2.72%), and Kandelia candel (0.02%). Moreover, to some extent, forest gap, surface elevation, and offshore distance greatly affected the spatial distribution pattern of mangrove species. This study demonstrated that the combination of WorldView-2 and Zhuhai-1 image had great potential in accurate mapping mangrove species at the landscape and regional scales, thereby facilitating biodiversity protection and scientific management of forest ecosystem and providing technical and data support for retrieval of ecosystem parameters (e.g., carbon storage, net primary production, and leaf area index) and health evaluation of mangrove forests. Future research will focus on the fusion of WorldView-2 and Zhuhai-1 image to simultaneously achieve high spatial resolution and hyperspectral bands and the inclusion of canopy height and leaf trait information (e.g., chlorophyll and water content) to the classification model.