Performance of XGBoost Ensemble Learning Algorithm for Mangrove Species Classification with Multisource Spaceborne Remote Sensing Data

Abstract

Accurately and timely monitoring the spatial distribution and composition of mangrove species is critical for assessing mangroves’ health, dynamics, and biodiversity, as well as mangroves’ degradation and restoration. Recent advances in machine learning algorithms, coupled with spaceborne remote sensing technique, offer an unprecedented opportunity to map mangroves at species level with high resolution over large extents. However, a single data source or data type is insufficient to capture the complex features of mangrove species and cannot satisfy the need for fine species classification. Moreover, identifying and selecting effective features derived from integrated multisource data are essential for integrating high-dimensional features for mangrove species discrimination. In this study, we developed a novel framework for mangrove species classification using spectral, texture, and polarization information derived from 3-source spaceborne imagery: WorldView-2 (WV-2), OrbitaHyperSpectral (OHS), and Advanced Land Observing Satellite-2 (ALOS-2). A total of 151 remote sensing features were first extracted, and 18 schemes were designed. Then, a wrapper method by combining extreme gradient boosting with recursive feature elimination (XGBoost-RFE) was conducted to select the sensitive variables and determine the optical subset size of all features. Finally, an ensemble learning algorithm of XGBoost was applied to classify 6 mangrove species in the Zhanjiang Mangrove National Nature Reserve, China. Our results showed that combining multispectral, hyperspectral, and L-band synthetic aperture radar features yielded the best mangrove species classification results, with an overall accuracy of 94.02%, a quantity disagreement of 4.44%, and an allocation disagreement of 1.54%. In addition, this study demonstrated important application potential of the XGBoost classifier. The proposed framework could provide fine-scale data and conduce to mangroves’ conservation and restoration.