11:00 AM - 1:00 PM
[MZZ49-P01] End-member estimation using photographs of the midden deposit sequence at Catalhoyuk, Central Anatolia
Keywords:Catalhoyuk, Midden deposit, Non-negative matrix factorization, Anatolia
Catalhoyuk in central Anatolia is a residential complex flourished from ~9100 to ~7600 years ago. In the courtyards within the dwellings, there are sediments called midden deposits with a maximum thickness of ~2 m. They are accumulation of garbage from the dwellings and are expected to record the living activities of the people (e.g., Shillito et al. 2011; 2013).
The main objective of our project is to reconstruct a continuous record of people's life in Catalhoyuk by splicing multiple midden deposits. In 2021, slab-shaped samples of ~50 x5 x 3 cm in size were collected from the midden deposit with a total length of ~1.4 m. In this presentation, we report the end-member estimation of the constituent particles of midden based on color analysis of the slab samples using photographs taken by a high-resolution camera. Non-negative matrix factorization (NMF; Lee and Seung 1999), one of the multivariate analysis methods, was applied to estimate the end members.
Since the color of sediments reflects the type and quantity of the constituent particles, the end members of the constituent particles can be estimated from the color data of the sediments. End-member estimation by the NMF method was previously conducted for deep-sea sediments using diffuse reflectance spectrophotometry data (Heslop et al. 2007). If end-member estimation can be conducted based on RGB values extracted from photographs of sediments, it is easier and it is possible to make two-dimensional maps of end-member intensity.
To apply the NMF method, each data must be a linearly combinable parameter proportional to the amount of material. Since the RGB values in image data are nonlinearly converted during encoding, the NMF method cannot be applied directly to RGB values extracted from the photographs. Photographs of a color chart and grayscale of known color were taken at the same time as the samples, and the conversion equations from the RGB values on the photograph to the RGB values as reflectance were derived. The RGB values as reflectance were then converted to absorbance equivalent values (absorption coefficient/scattering coefficient) by the Kubelka-Munk function. From Lambert-Baer's law, absorbance is a linearly combinable quantity proportional to the amount of material. Assuming that the scattering coefficient is constant, the value converted by the Kubelka-Munk function can be regarded as a linearly combinable quantity proportional to the amount of material.
In general, the NMF method does not provide a unique solution. Thus, the constraint that the sum of the intensities of end members is one (Du et al. 2005) was given to obtain a unique solution. In addition, it is necessary to provide initial values for iterative calculations and to specify the number of end members in advance. For the initial values, we used the Fuzzy C-means clustering method (Zadeh 1965; Heslop et al. 2007). The center of each cluster was used as the initial value of the end members, and the membership of individual points to each cluster was used as the initial value of the intensities. The NMF method was first conducted with three end members and repeatedly conducted by increasing the number of end members by one. If the end members added were confirmed as corresponding to grains on the photographs, they were accepted as end members.
As a result of the above analysis, five end members were estimated: 1) black gravels and fine particles, 2) white to yellowish-white gravels and their fragments, 3) dark brown fine particles, 4) reddish-brown to yellowish-brown fine particles, and 5) gray to dark gray fine particles. In comparison with previous studies on thin-section observations of midden deposits (Shillito et al. 2011; 2013), these end members may correspond to charcoal, plaster fragments, brick fragments, excrement and rotten fruit, and ash, respectively. The distribution of the intensity of each end member and the relation with the results of semiquantitative XRF analysis will be discussed by Tada, R. et al. (this session)
The main objective of our project is to reconstruct a continuous record of people's life in Catalhoyuk by splicing multiple midden deposits. In 2021, slab-shaped samples of ~50 x5 x 3 cm in size were collected from the midden deposit with a total length of ~1.4 m. In this presentation, we report the end-member estimation of the constituent particles of midden based on color analysis of the slab samples using photographs taken by a high-resolution camera. Non-negative matrix factorization (NMF; Lee and Seung 1999), one of the multivariate analysis methods, was applied to estimate the end members.
Since the color of sediments reflects the type and quantity of the constituent particles, the end members of the constituent particles can be estimated from the color data of the sediments. End-member estimation by the NMF method was previously conducted for deep-sea sediments using diffuse reflectance spectrophotometry data (Heslop et al. 2007). If end-member estimation can be conducted based on RGB values extracted from photographs of sediments, it is easier and it is possible to make two-dimensional maps of end-member intensity.
To apply the NMF method, each data must be a linearly combinable parameter proportional to the amount of material. Since the RGB values in image data are nonlinearly converted during encoding, the NMF method cannot be applied directly to RGB values extracted from the photographs. Photographs of a color chart and grayscale of known color were taken at the same time as the samples, and the conversion equations from the RGB values on the photograph to the RGB values as reflectance were derived. The RGB values as reflectance were then converted to absorbance equivalent values (absorption coefficient/scattering coefficient) by the Kubelka-Munk function. From Lambert-Baer's law, absorbance is a linearly combinable quantity proportional to the amount of material. Assuming that the scattering coefficient is constant, the value converted by the Kubelka-Munk function can be regarded as a linearly combinable quantity proportional to the amount of material.
In general, the NMF method does not provide a unique solution. Thus, the constraint that the sum of the intensities of end members is one (Du et al. 2005) was given to obtain a unique solution. In addition, it is necessary to provide initial values for iterative calculations and to specify the number of end members in advance. For the initial values, we used the Fuzzy C-means clustering method (Zadeh 1965; Heslop et al. 2007). The center of each cluster was used as the initial value of the end members, and the membership of individual points to each cluster was used as the initial value of the intensities. The NMF method was first conducted with three end members and repeatedly conducted by increasing the number of end members by one. If the end members added were confirmed as corresponding to grains on the photographs, they were accepted as end members.
As a result of the above analysis, five end members were estimated: 1) black gravels and fine particles, 2) white to yellowish-white gravels and their fragments, 3) dark brown fine particles, 4) reddish-brown to yellowish-brown fine particles, and 5) gray to dark gray fine particles. In comparison with previous studies on thin-section observations of midden deposits (Shillito et al. 2011; 2013), these end members may correspond to charcoal, plaster fragments, brick fragments, excrement and rotten fruit, and ash, respectively. The distribution of the intensity of each end member and the relation with the results of semiquantitative XRF analysis will be discussed by Tada, R. et al. (this session)