Remote Sensing Feature Selection for Alpine Wetland Classification
HUO Xuanlin,NIU Zhenguo,ZHANG Bo,LIU Linsong,LI Xia
DOI: https://doi.org/10.11834/jrs.20222080
2023-01-01
Journal of Remote Sensing
Abstract:Alpine wetlands are an important surface cover type on the Qinghai―Tibet Plateau because they play a key role in water conservation, climate regulation, and biodiversity maintenance. Accurate and timely knowledge of the temporal and spatial distribution of alpine wetlands is necessary for wetland protection and management. The selection of remote sensing classification features is crucial in wetland mapping. Although spectral, texture, and topographic features have been investigated, studies focusing on spectral index features and their mathematical statistical features and feature selection methods are limited. This study aims to classify alpine wetlands from the aspects of mathematical statistical features, alpine wetland types, feature selection methods, and selected feature sets combined with random forest classification algorithm using Sentinel-2 image data and taking the Shouqu Alpine Wetland Reserve as the research site. An in-depth and comprehensive analysis on the spectral index characteristics of alpine wetlands is performed to optimize the classification characteristics of alpine wetlands.The Gansu Shouqu Alpine Wetland Reserve was used as the research area, and classification characteristics（spectrum, vegetation index, red edge index, and water body index） were obtained on the basis of Sentinel-2 data. Filter and wrapper feature selection methods,including Jeffries – Matusita distance, Spectral Angular Distance（SAD）, Euclidean Distance（ED）, RF-RFE algorithm, and Relief-F algorithm are utilized to optimize these features. Meanwhile, Z test is applied for quantitative evaluation.The following conclusions can be drawn from this study.（1） Among the categories of alpine wetlands involved in the classification,rivers and bare land are the easiest to distinguish, followed by grasslands and swamps and then swampy meadows and meadows. MCARI2,NDWI, DVI, EVI, EWI, IRECI, MCARI, TCARI, and UGWI indices can be used to differentiate among adjacent swamps, swampy meadows, meadows, and grasslands.（2） The order of contribution of different index characteristics to wetland information extraction in terms of degree is water body index characteristics ＞ vegetation index characteristics ＞ red edge index characteristics.（3） ED and Relief-F algorithms in the filter method demonstrate excellent performance from the perspective of feature optimization methods.（4） A suitable alpine wetland information extraction method is selected using the indices RDVI, NDVI, MSR, RVI, VIgreen, RNDWI, NDWI, NDWI＿B,MNDWI, EWI, and CIre.（5） The mathematical statistics of different classification features indicated that the median feature obtains the best classification result, followed by the average value feature.We provide detailed results from feature optimization methods, wetland classification optimization index, statistical feature evaluation,and categories involved in alpine wetland classification using multi-dimensional analysis. To the best of our knowledge, this study provides a novel transferable and universal method for the selection of characteristic variables for wetland information extraction.