Abstract |
The excavation of mass graves and sites of accidents or natural disasters, which reveal
numbers of hard tissue remains originating from multiple individuals, is usual in
archaeology and forensic science. The discrimination of the individuals and the
classification of their remains is useful for revealing the identity, as well as information
about life and death of those individuals. However, poor preservation makes the task
of discrimination/classification extremely difficult and time consuming, using
conventional methods based on morphological characteristics or DNA analysis. Thus,
the need for simple, direct and cost-effective analysis of hard tissue remains, with
minimal damage to the artifacts, has emerged.
In the current work, Laser Induce Breakdown Spectroscopy (LIBS), combined with
Machine Learning algorithms and a simple Artificial Neural Network, were employed
for the discrimination and classification of hard tissue remains. Several bone
fragments and teeth were studied, using a LIBS microscopy setup (micro-LIBS) for data
collection, while Machine Learning algorithms and a Neural Network were used for
data analysis.
Micro-LIBS is a micro-destructive, fast and transferable method, with high spatial
resolution (around 50μm/spot) that enables analysis of the surface or the crosssection of samples, with little or no sample preparation, providing massive amounts
of data in little time. Thus, it is a suitable technique to be combined with machine
learning algorithms for the analysis of the collected data.
Hydroxyapatite (Ca5(PO4)3OH) is the main component of both bones and teeth, while
proteinaceous materials(mainly collagen) and water, in different ratios, complete the
hard tissue matrix. Magnesium (Mg), Strontium (Sr) and Barium (Ba) can replace
Calcium (Ca) in metabolic processes and thus can replace it in hydroxyapatite’s crystal.
Spectral emission lines from biogenic elements in the remains are observed across the
spectral range used (200 - 660 nm). Hence, the data collected provide significant
information to the algorithms employed.
Machine Learning and Neural Networks enable computers to learn from experience
following a similar process with several living organisms. This process is based on
pattern recognition on given data, improving future decisions and giving computers
the ability to learn without being explicitly programmed. This pattern recognition on
LIBS data is the aim of this work. The comparison of four different methods (k-Nearest
Neighbors, Random Forest, Support Vector Machine, Artificial Neural Network) with
gradual complexity, after parameter tuning and feature selection, provided the best
behaved model to achieve the requested task.
Artificial Neural Network had significantly better results compared to the rest of the
models used, while the selection of specific spectral areas corresponding in spectral
lines from biogenic elements increased the resulting classification accuracy. The
achieved classification varied from decent to excellent, giving a good classification
accuracy regardless of the used data.
Concluding, the present work is an attempt for development of a fast, accurate and
easily accessible and applicable methodology for the discrimination and classification
of hard tissue remains, based on the analysis of LIBS data using machine learning
models.
|