基于光谱技术的典型互花米草入侵湿地土壤属性预测研究  

Using Spectroscopy to Predict Soil Properties on Coastal Wetlands Invaded by Spartina Alterniflora

作　　者：陈旭 曹思珩 杨仁敏 陈秋宇 李建国 徐璐 CHEN Xu;CAO Si-heng;YANG Ren-min;CHEN Qiu-yu;LI Jian-guo;XU Lu(School of Geography,Geomatics and Planning,Department of Geographic Information Science,Jiangsu Normal University,Xuzhou 221116,China)

机构地区：[1]江苏师范大学地理测绘与城乡规划学院地理信息科学系,江苏徐州221116

出　　处：《光谱学与光谱分析》2025年第1期197-203,共7页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金项目(42371053,42271287,42171054);江苏省海洋科技创新项目(JSZRHYKJ202212);江苏师范大学科研与实践创新计划校级项目(2021XKT0089)资助。

摘　　要：为了高效监测互花米草入侵海岸带湿地生态系统的土壤属性变化,选取江苏省盐城湿地珍禽国家级自然保护区的一处典型互花米草入侵湿地作为研究区,利用随机分层采样方法选取15个样点,在3个深度(0~30、30~60、60~100 cm)共采集45个土壤样品,测定了土壤可见光-近红外反射光谱和10种土壤理化属性,研究了偏最小二乘和随机森林两种方法的预测能力,分析了不同光谱变换形式对预测精度的影响,探讨了入侵年限和土壤深度作为辅助预测变量的潜力。结果表明:(1)可见光-近红外光谱技术可以较好地预测有机碳、无机碳、全氮、含水量、pH、容重、盐分和黏粒等属性;(2)偏最小二乘回归法比随机森林法更适合监测互花米草入侵湿地的关键土壤理化属性,利用偏最小二乘法对土壤属性建立的预测模型精度(R^(2))在0.341~0.979之间,随机森林方法对土壤属性建立的预测模型精度(R^(2))最高为0.722;(3)基于原始光谱可以获得土壤全氮的最优预测模型(R^(2)为0.769,RMSE为0.091 g·kg^(-1)),而其他土壤属性的最优模型多是基于微分变换或倒数变换建立的模型,微分变换和倒数变换可以有效地提高模型预测精度;(4)模型预测精度在加入入侵年限和土壤深度两个变量后总体上有所提高,其中有机碳、全氮、盐分、pH和容重等属性的预测精度对这两个变量更为敏感,土壤有机碳最优模型的精度(R^(2))从0.794提高到0.806,pH最优模型的精度(R^(2))从0.838提升至0.884,盐分最优模型的精度(R^(2))从0.978提升至0.997。综上所述,可见光-近红外光谱技术在互花米草入侵湿地关键土壤理化属性预测方面具有明显的优越性,通过适当的光谱变换、变量筛选、模型选择等方面可以实现互花米草入侵湿地土壤变化的快速监测。This study aimed to effectively monitor the changes in soil properties after Spartina alterniflora invasion on coastal wetland ecosystems.The study area is a typical Spartina alterniflora wetland in the Yancheng Wetland Rare Birds National Nature Reserve of Jiangsu Province.A total of 15 sites were identified by a stratified-random sampling method,and 45 soil samples were collected from three depth intervals(0~30,30~60,and 60~100 cm).The visible-near infrared spectral reflectance and 10 soil physicochemical properties were measured.The performance of partial least squares regression(PLSR)and random forest(RF)was studied,spectral transformation forms'influence on prediction accuracy was analyzed,and the potential of invasion years and soil depth as auxiliary predictors were discussed.The results show that:(1)the visible-near infrared spectral reflectance can be used to predict organic carbon,inorganic carbon,total nitrogen,water content,pH,bulk density,salinity,and clay contents in soils with reasonable accuracy;(2)the method of partial least squares generally outperform random forest algorithm,the R^(2)of prediction models developed using the PLSR method was between 0.341 and 0.979,and the biggest R^(2)of random forest models was 0.722;(3)Differential transformation and reciprocal transformation of spectral reflectance can substantially improve the model performance.The optimal prediction model of full nitrogen can be obtained based on the original spectra(R^(2)is 0.769 and RMSE is 0.091 g·kg^(-1)).In contrast,the optimal models for other soil properties are mostly based on differential or reciprocal transformation of the original spectra.(4)In general,the model performance can be improved by adding variables of invasion years and soil depth,and the prediction accuracy of organic carbon,total nitrogen,salinity,pH and bulk density models are more sensitive to the two variables.The prediction model accuracy(R^(2))for estimating soil organic carbonincreased from 0.794 to 0.806,the accuracy(R^(2))of the pH model inc

关 键 词：互花米草入侵 可见光-近红外光谱 土壤监测 偏最小二乘法 随机森林 光谱变换 

分 类 号：S153.6[农业科学—土壤学]