Premium
Benchmarking Component Analysis of Remanent Magnetization Curves With a Synthetic Mixture Series: Insight Into the Reliability of Unmixing Natural Samples
Author(s) -
He Kuang,
Zhao Xiangyu,
Pan Yongxin,
Zhao Xiang,
Qin Huafeng,
Zhang Tongwei
Publication year - 2020
Publication title -
journal of geophysical research: solid earth
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.983
H-Index - 232
eISSN - 2169-9356
pISSN - 2169-9313
DOI - 10.1029/2020jb020105
Subject(s) - endmember , parametric statistics , remanence , mixture model , mathematics , pattern recognition (psychology) , reliability (semiconductor) , component (thermodynamics) , computer science , statistics , artificial intelligence , magnetization , image (mathematics) , physics , power (physics) , quantum mechanics , magnetic field , thermodynamics
Geological samples often contain several magnetic components associated with different geological processes. Component analysis of remanent magnetization curves has been widely applied to decompose convoluted information. However, the reliability of commonly used methods is poorly assessed as independent verification is rarely available. For this purpose, we designed an experiment using a series of mixtures of two endmembers to benchmark unmixing methods for isothermal remanent magnetization (IRM) acquisition curves. First‐order reversal curves (FORC) diagrams were analyzed for comparison. It is demonstrated that the parametric method, which unmixes samples using specific probability distributions, may result in biased estimates. In contrast, an endmember‐based IRM unmixing approach yielded better quantitative results, which are comparable to the results obtained by FORC analysis based on principle component analysis (FORC‐PCA). We demonstrate that endmember‐based methods are in principle more suitable for unmixing a collection of samples with common endmembers; however, the level of decomposition will vary depending on the difference between the true endmembers that are associated with distinctive processes and the empirical endmembers used for unmixing. When it is desired to further decompose endmembers, the parametric unmixing approach remains a valuable means of inferring their underlying components. We illustrate that the results obtained by endmember‐based and parametric methods can be quantitatively combined to provide improved unmixing results at the level of parametric model distributions.