Jennifer A. Rogers, Kevin M. Robertson, Todd J. Hawbaker, Daniel J. Sousa
{"title":"Classifying Plant Communities in the North American Coastal Plain With PRISMA Spaceborne Hyperspectral Imagery and the Spectral Mixture Residual","authors":"Jennifer A. Rogers, Kevin M. Robertson, Todd J. Hawbaker, Daniel J. Sousa","doi":"10.1029/2024JG008217","DOIUrl":null,"url":null,"abstract":"<p>The effort to map terrestrial biodiversity, in recent years limited mostly to the use of broadband multispectral remote sensing at decameter scales, can be greatly enhanced by harnessing hyperspectral imagery. Interpretation of hyperspectral imagery may be aided by the Mixture Residual (MR) spectral preprocessing transformation. MR integrates the benefits of spectral mixture analysis with the absorption peak-enhancing characteristics of continuum removal. MR characterizes each pixel as a linear combination of generic end-members estimating the spectral continuum, from which the residual of each wavelength is computed and treated as a source of additional information. Using Hyperspectral Precursor of the Application Mission (PRISMA) imagery, we tested the ability of MR-transformed reflectance as compared to untransformed surface reflectance (SR) to map plant associations and land cover using ground truthing and random forest classifications across four landscapes within the North American Coastal Plain. We used a forward stepwise selection algorithm to choose bands for each classification and subsequently compared these between SR and MR. Our MR classifications distinguished land cover with 5% greater balanced accuracy on average than the SR-based classifications across all four landscapes. The MR-based classification that integrated data from all landscapes into a unified model encompassing all 21 land cover types achieved a 76% average balanced accuracy over three iterations. Generally, MR utilized the near-infrared region to a greater degree than SR while deemphasizing the green peak. Based on our results, MR improves the accuracy of mapping terrestrial biodiversity, likely extending to other current and planned satellite hyperspectral missions.</p>","PeriodicalId":16003,"journal":{"name":"Journal of Geophysical Research: Biogeosciences","volume":"129 9","pages":""},"PeriodicalIF":3.7000,"publicationDate":"2024-09-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1029/2024JG008217","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Journal of Geophysical Research: Biogeosciences","FirstCategoryId":"93","ListUrlMain":"https://onlinelibrary.wiley.com/doi/10.1029/2024JG008217","RegionNum":3,"RegionCategory":"环境科学与生态学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q2","JCRName":"ENVIRONMENTAL SCIENCES","Score":null,"Total":0}
引用次数: 0
Abstract
The effort to map terrestrial biodiversity, in recent years limited mostly to the use of broadband multispectral remote sensing at decameter scales, can be greatly enhanced by harnessing hyperspectral imagery. Interpretation of hyperspectral imagery may be aided by the Mixture Residual (MR) spectral preprocessing transformation. MR integrates the benefits of spectral mixture analysis with the absorption peak-enhancing characteristics of continuum removal. MR characterizes each pixel as a linear combination of generic end-members estimating the spectral continuum, from which the residual of each wavelength is computed and treated as a source of additional information. Using Hyperspectral Precursor of the Application Mission (PRISMA) imagery, we tested the ability of MR-transformed reflectance as compared to untransformed surface reflectance (SR) to map plant associations and land cover using ground truthing and random forest classifications across four landscapes within the North American Coastal Plain. We used a forward stepwise selection algorithm to choose bands for each classification and subsequently compared these between SR and MR. Our MR classifications distinguished land cover with 5% greater balanced accuracy on average than the SR-based classifications across all four landscapes. The MR-based classification that integrated data from all landscapes into a unified model encompassing all 21 land cover types achieved a 76% average balanced accuracy over three iterations. Generally, MR utilized the near-infrared region to a greater degree than SR while deemphasizing the green peak. Based on our results, MR improves the accuracy of mapping terrestrial biodiversity, likely extending to other current and planned satellite hyperspectral missions.
期刊介绍:
JGR-Biogeosciences focuses on biogeosciences of the Earth system in the past, present, and future and the extension of this research to planetary studies. The emerging field of biogeosciences spans the intellectual interface between biology and the geosciences and attempts to understand the functions of the Earth system across multiple spatial and temporal scales. Studies in biogeosciences may use multiple lines of evidence drawn from diverse fields to gain a holistic understanding of terrestrial, freshwater, and marine ecosystems and extreme environments. Specific topics within the scope of the section include process-based theoretical, experimental, and field studies of biogeochemistry, biogeophysics, atmosphere-, land-, and ocean-ecosystem interactions, biomineralization, life in extreme environments, astrobiology, microbial processes, geomicrobiology, and evolutionary geobiology