11:30 AM - 11:45 AM
[PPS03-10] Development of identifying µm-sized organic matter in carbonaceous chondrites: A method for classification combined with machine learning
Keywords:organic matter, carbonaceous chondrite, machine learning
Carbonaceous chondrites (CC) contain abundant organic matter (OM), the formation of which has been posited to occur via multiple mechanisms, including on the progenitor planetesimals of asteroid bodies, within the protosolar nebula of the early solar system, and within diffuse and molecular clouds in the interstellar medium. Meteoritic OM would have been delivered to the early Earth, with the potential to facilitate the origin of life. Therefore, OM from meteorite provides a record of early solar system processes and an inventory of compounds that could have been essential for the origin of life on Earth. In prior studies the OM was located by analyzing many 10x10 µm2 regions using a secondary ion mass spectrometer (SIMS)1,2. A few OM were then characterized by transmission electron microscopy3–5.
CC contain numerous µm-sized OM (µOM) and diffuse organic matter (DOM) within their matrix, which are randomly distributed. The µOM was identified as a C-rich object characterized by a dark coloration in BSE image, which is surrounded by matrix components. DOM is characterized by its C-rich composition, brighter coloration (compared with µOM) and is intermixed with the matrix components. The matrix components are predominantly phyllosilicate with minor Fe-sulfide and carbonate minerals. Previous studies have characterized µOM as IOM and DOM as diffuse carbon, occurring as many nanoscale objects3–5.This study utilizes a microscale scanning electron microscopy-energy dispersive spectrometry (SEM-EDS) approach to locate numerous OM on the sample. The generation of an OM map involved scanning the whole area of a sample. The BSE and element maps were obtained from this process. Image processing of the element maps was then performed to obtain the OM map. An OM map is thus represented by a Boolean array of pixels with true connected components being designated as instances. The instances were then manually classified into one of four categories such as µOM, DOM, a hole, or noise.
In this study, a method to classify OM using machine learning technique was developed. The method utilizes mosaic backscatter electron (BSE) images, element maps, and OM maps of the mm-sized CC samples for the reference surface (Orgueil) and a new surface (Murray). The two-dimensional properties of µOM and DOM, such as geometry and chemistry were estimated from the BSE image, element maps, and OM map. The performance of five different classification algorithms were compared, and these were logistic regression, Gaussian-naïve Bayes, k-nearest neighbor, support vector machine and random forest. The algorithms were trained on the instances from the reference surface. Certainty parameters such as accuracy (A0), precision (A1), recall (A2), F1-score (ZF1), and Kappa-score (Zk) were estimated for the algorithms. Among the algorithms, the random forest certainty parameters were the highest in the majority of cases and were A0 (0.9), A1 (0.9), A2 (0.90), ZF1 (0.87), and Zk (0.50). Thus, random forest was selected as the algorithm to classify the instances. As such, the instances in the OM maps of Murray were classified using the random forest algorithm.
The resulting OM maps with their classified instances, enabled many µOM and DOM objects to be identified and located. These OM objects had a size ranging from 1 to 15 µm. The spatial occurrence of these objects exhibit heterogeneity, and their orientation is random, suggesting that unidirectional physical forces did not influence their accretion or formation during aqueous alteration. Therefore the OM maps with their classified OM can be utilized to select regions of interest for further detailed spatial analyses, such as transmission electron microscopy (TEM), secondary ion mass spectrometry (SIMS), Raman spectroscopy, and Fourier transform infrared spectroscopy (FTIR).
CC contain numerous µm-sized OM (µOM) and diffuse organic matter (DOM) within their matrix, which are randomly distributed. The µOM was identified as a C-rich object characterized by a dark coloration in BSE image, which is surrounded by matrix components. DOM is characterized by its C-rich composition, brighter coloration (compared with µOM) and is intermixed with the matrix components. The matrix components are predominantly phyllosilicate with minor Fe-sulfide and carbonate minerals. Previous studies have characterized µOM as IOM and DOM as diffuse carbon, occurring as many nanoscale objects3–5.This study utilizes a microscale scanning electron microscopy-energy dispersive spectrometry (SEM-EDS) approach to locate numerous OM on the sample. The generation of an OM map involved scanning the whole area of a sample. The BSE and element maps were obtained from this process. Image processing of the element maps was then performed to obtain the OM map. An OM map is thus represented by a Boolean array of pixels with true connected components being designated as instances. The instances were then manually classified into one of four categories such as µOM, DOM, a hole, or noise.
In this study, a method to classify OM using machine learning technique was developed. The method utilizes mosaic backscatter electron (BSE) images, element maps, and OM maps of the mm-sized CC samples for the reference surface (Orgueil) and a new surface (Murray). The two-dimensional properties of µOM and DOM, such as geometry and chemistry were estimated from the BSE image, element maps, and OM map. The performance of five different classification algorithms were compared, and these were logistic regression, Gaussian-naïve Bayes, k-nearest neighbor, support vector machine and random forest. The algorithms were trained on the instances from the reference surface. Certainty parameters such as accuracy (A0), precision (A1), recall (A2), F1-score (ZF1), and Kappa-score (Zk) were estimated for the algorithms. Among the algorithms, the random forest certainty parameters were the highest in the majority of cases and were A0 (0.9), A1 (0.9), A2 (0.90), ZF1 (0.87), and Zk (0.50). Thus, random forest was selected as the algorithm to classify the instances. As such, the instances in the OM maps of Murray were classified using the random forest algorithm.
The resulting OM maps with their classified instances, enabled many µOM and DOM objects to be identified and located. These OM objects had a size ranging from 1 to 15 µm. The spatial occurrence of these objects exhibit heterogeneity, and their orientation is random, suggesting that unidirectional physical forces did not influence their accretion or formation during aqueous alteration. Therefore the OM maps with their classified OM can be utilized to select regions of interest for further detailed spatial analyses, such as transmission electron microscopy (TEM), secondary ion mass spectrometry (SIMS), Raman spectroscopy, and Fourier transform infrared spectroscopy (FTIR).