A Cluster Analysis Methodology for the Categorization of Soil Samples for Forensic Sciences Based on Elemental Fingerprint. Issue 1 (31st December 2022)

• Record Type: Journal Article
• Authors: Maione, Camila; Luíza da Costa, Nattane; Barbosa, Fernando; Barbosa, Rommel Melgaço
• Is Part Of: Applied artificial intelligence. Volume 36:Issue 1(2022)
• Publication Date: 2022-12-31
• Subjects: Artificial intelligence -- Periodicals; 006.3
• DOI: 10.1080/08839514.2021.2010941
• Languages: English
• ISSNs: 0883-9514

Abstract

ABSTRACT: Soil and dirt fragments are easily transferred from the ground to objects such as clothing, shoes, skin, nails, and tires. The elemental analysis of these samples involved in crimes can be an important source of information for forensic scientists because they can present substantial evidence by creating links between victims, suspects, crime scenes and other relevant actors or places. In this work we present a promising new approach for the study of soil samples, using data mining techniques applied to the elemental fingerprints of soil. We experimented on soil samples obtained from southeast Oregon and northern Nevada, two neighboring states in the United States that have similar geological characteristics while also displaying some specific differences. The chemical composition of soil and sediments samples were determined by the use of inductively coupled plasma-mass spectrometry (ICP-MS). Thirty-three elements were analyzed, and we used their concentrations to conduct the analysis. Cluster analysis was performed employing the K-means clustering algorithm. We found three clusters that showed interesting chemical patterns. In order to investigate the most significant chemical elements that distinguish the clusters, we employed the Correlation-Based Feature Selection (CFS) algorithm. Lastly, we developed a classification model based on support vector machine (SVM), which can predict in which of the found clusters an arbitrary soil sample would fall with a 99% prediction accuracy when all 32 variables were used for training the model, and a 95% prediction accuracy when only the 10 most relevant elements were used for training the model. Following this methodology, forensic scientists and experts would be able to establish profiles of soil samples extracted from the crime scene and nearby regions, and use classification models to predict which of these profiles an arbitrary soil sample found on the subjects involved in the crime would be associated with.