摘要
高寒草甸是青藏高原生态屏障的核心组成部分,其优势物种识别对生态健康评估至关重要。受高寒草甸异质性的影响,不同特征选择和分类技术的适用性仍不明确。研究选取川西典型高寒草甸6种优势物种,采用手持光谱仪和无人机高光谱匹配校准数据,系统比较了3种特征选择方法(RFS、ERT、MIC)和5种分类算法(RF、SVM、XGBoost、LightGBM、MLP)的性能。结果表明:(1)原始光谱特征参数对分类贡献最高,关键信息集中在可见光和近红外波段。(2)基于树模型特征的筛选方法能够显著减少特征数量的同时保持较高的分类精度。(3)MLP在优势物种识别中表现最佳,影像分类总体精度达76.98%,Kappa系数为0.72,F1值为0.77,但对非优势绿色植被识别能力仍需提升。研究为国家公园草地生态系统的精细监测与退化评估提供了可参考、推广的技术路径,有助于提升青藏高原国家公园群的生态保护成效和科学管理水平。
Alpine meadows are a crucial component of the ecological security barrier of the Qinghai-Tibet Plateau,and accurately identifying and mapping their dominant plant species is vital for monitoring ecosystem health.However,due to the high spatial and spectral heterogeneity of alpine meadows,it remains unclear how effectively different feature selection strategies and classification algorithms can identify dominant species in hyperspectral images.This study focuses on a typical alpine meadow in Western Sichuan as the research area,and selects six dominant plant species that play significant roles in shaping community structure and local habitats as target categories.Ground spectral data were collected using an ASD FieldSpec HandHeld 2 spectrometer,while co-located hyperspectral images were simultaneously acquired via an unmanned aerial vehicle(UAV).After matching and calibration,a consistent spectral reflectance dataset was established.On this basis,the study systematically compared the performance of combinations of three feature selection methods(RFS:random forest-based selection;ERT:extremely randomized tree-based selection;MIC:maximal information coefficient),and five classification algorithms(RF:random forest;SVM:support vector machine;XGBoost:extreme gradient boosting;LightGBM:light gradient-boosting machine;MLP:multi-layer perceptron)in recognizing dominant plant species in hyperspectral data.The results indicate that:(1)The original spectral features contribute most significantly to classification accuracy,with effective information primarily concentrated in the visible and near-infrared bands.(2)Tree model-based feature selection methods(RFS and ERT)substantially reduce the number of required spectral features while maintaining high classification accuracy.(3)Among the classifiers tested,the MLP model performed best,achieving an overall accuracy of 76.98%,a Kappa coefficient of 0.72,and an F1 score of 0.77 in hyperspectral image classification.However,its ability to identify non-dominant green vegetation requires further improvement.This study proposes a transferable methodological framework for refined monitoring and degradation assessment of grassland ecosystems,thus enhancing the ecological protection and scientific management of the national parks on the Qinghai-Tibet Plateau.
作者
刘维鸿
谢璋健
何贞晓
张路
张志明
LIU Weihong;XIE Zhangjian;HE Zhenxiao;ZHANG Lu;ZHANG Zhiming(Institute of International Rivers and Ecology Security,Yunnan University,Kunming 650091,China;School of Ecology and Environmental Science,Yunnan University,Kunming 650091,China;Research Center for Ecology Environmental Sciences,Chinese Academy of Sciences,Beijing 100085,China)
基金
国家自然科学基金(42171285)
第二次青藏高原综合科学考察研究专题项目(2019QZKK0308)。
关键词
高寒草甸
高光谱
优势物种
光谱特征
分类识别
alpine meadow
hyperspectral
dominant species
spectral features
classification and identification
