FTIR spectroscopic analysis of soil in forensic science

Soil is a complex mixture of organic matter and inorganic materials of varying sizes including sand, minerals, salts, and clay. Soil may also contain heavy metals based upon the pH and the use of the land including mining, tanning, and other industrial activities. Spectroscopic tools can be used to assess the metal ion complexation and bridging of natural organic matter (NOM) and chemical moieties of the NOM. Attenuated total reflectance (ATR) FTIR spectroscopy can be used to characterize soils toward forensic geolocation. Soils have been found to have unique vibrational spectral fingerprints as NOM structure, makeup, pH, and bound metals and materials can aid in individualizing the soils. The focus of this review is the application of Fourier transform infrared spectroscopy (FTIR) for the forensic analysis of soils. It provides an overview of the sampling process, locations, collection and homogenization, instrumental settings, and data analysis groups have used in their studies. Different soils have been found to have unique vibrational spectral fingerprints and FTIR has been shown to characterize soils for forensic geolocation. The review captures the approaches and findings across the field and will be informative to guide the future direction and methods for ongoing research. A standardized and consensus approach for sampling, preprocessing and data collection would accelerate data comparisons and conclusions that can be made from soil investigations and applications to casework.

1 Introduction

Soil is a complex mixture of organic matter and inorganic materials of varying sizes including sand and clay. Soil may also contain heavy metals based upon the pH and use of the land including mining, tanning, and other industrial activities. Several studies focus on soil type from diverse locations including samples collected near a lake (Kizil et al., 2024), sandy soil (Newland et al., 2023; Newland et al., 2022), subtropical soil (e Silva et al., 2022), and soil collected from flowerbeds, woodland areas, and riverbanks (Baron et al., 2011).

Spectroscopic tools can be used to assess the metal ion complexation and bridging of natural organic matter (NOM) and chemical moieties of the NOM. Unique structures, makeup, and materials can be used to individualize the soils. Vibrational spectroscopy tools including attenuated total reflectance (ATR) Fourier transform infrared (FTIR) spectroscopy and Diffuse reflectance infrared Fourier transform spectroscopy (DRIFTS) have been applied to soil analysis. As FTIR spectroscopy is non-destructive, rapid, and requires little to no sample preparation, it is ideal for the analysis of trace forensic samples.

FTIR spectroscopy is regarded as an emerging analytical method for forensic soil analysis. It is widely used in the analysis of trace evidence including seized drugs, paints, and fibers, and has been applied to body fluid analysis (Elkins, 2011; Elkins, 2019). Compounds such as humic and fulvic acid NOM and clays in soil with a permanent dipole interact with infrared radiation. Organic matter is composed of bonds that vibrate and match infrared frequencies in the micron range. The absorptions or percent transmittance at specific wavenumber values can be used to assign functional groups using correlation charts. The resulting spectra can then be analyzed using various statistical methods, including linear discriminant analysis, first and second derivative analysis, principal component analysis, and partial least squares analysis. Statistical tools can draw out variations that are not detected by visual inspection or comparison. The wide variety of bonds and functional groups in organic molecules lead to specific variations in FTIR spectra.

Apart from FTIR, some studies have used instrumental techniques such as microspectrophotometry (Newland et al., 2023; Newland et al., 2022; Woods et al., 2014), X-ray diffraction (Newland et al., 2023; Newland et al., 2022), X-ray fluorescence (de Caritat et al., 2022; de Caritat et al., 2021; e Silva et al., 2022), inductively coupled plasma optical emission spectroscopy (Kizil et al., 2024), inductively coupled plasma mass spectrometry (de Caritat et al., 2022; de Caritat et al., 2021), gamma-ray spectrometry (e Silva et al., 2022), and laser-induced breakdown spectroscopy (Xu et al., 2020). Other examinations the authors conducted included magnetic susceptibility analysis (de Caritat et al., 2021; e Silva et al., 2022), textural analysis (Kizil et al., 2024), sedimentation test (Kizil et al., 2024), visual color analysis using the Munsell color chart (Kizil et al., 2024; Cox et al., 2000), and quantitative color analysis using an Instrutherm ACR-1023 handheld analyzer (e Silva et al., 2022). These methods can provide further insight into soil analysis.

Soil is one of the earliest trace evidence types to be used in forensic investigations to link suspects to victims and crime scenes. Soil has been used in forensic cases since 1904 when Georg Popp employed it in the Eva Disch murder case. Soil recovered from the suspect’s pant cuffs was compared to soil where she was found in a bean field and a pathway leading from the suspect’s home to the crime scene. The suspect, Karl Laubach, confessed when presented with the soil and other analysis conclusions (Bergslien, 2012).

The purpose of this paper is to review the collection and preparation of soil material and the application of FTIR spectroscopy to characterize soils toward forensic geolocation over larger geographic areas. The locations, sampling procedures, sample preparation, data analysis, and results are the focus of this paper.

2 Selection of articles

The scope of this work is FTIR spectroscopy applied to the analysis of soil for forensic geolocation. Articles were identified using databases including Google Scholar and PubMed. Articles were included if they were original research and included location, sampling, sample preparation, FTIR spectroscopy, data analysis and study conclusions relevant to the scope and including the keywords FTIR spectroscopy, soil, and forensic. Papers were excluded if they lacked a forensic focus. The search was performed from September 5th to 10th, 2024. Fourteen studies were reviewed.

Eleven of the studies used ATR FTIR spectroscopy to analyze soils (Baron et al., 2011; Chauhan et al., 2018; de Caritat et al., 2022; de Caritat et al., 2021; e Silva et al., 2022; Kizil et al., 2024; Koçak et al., 2021; Newland et al., 2023; Newland et al., 2022; Sangwan et al., 2024; Woods et al., 2014; Xu et al., 2020) while three used an FTIR with KBr pellets (Cox et al., 2000; Idrizi et al., 2021; Liu et al., 2015). One study compared ATR FTIR spectroscopy with DRIFTS (Koçak et al., 2021).

3 Locations and sampling procedures

3.1 Locations studied

Soil samples from Australia (de Caritat et al., 2022; de Caritat et al., 2021; Newland et al., 2023; Newland et al., 2022; Woods et al., 2014), India (Chauhan et al., 2018; Sangwan et al., 2024), the United States (Cox et al., 2000; Koçak et al., 2021), the United Kingdom (Baron et al., 2011), China (Liu et al., 2015; Xu et al., 2020), Turkey (Kizil et al., 2024), Brazil (e Silva et al., 2022), and North Macedonia (Idrizi et al., 2021) were investigated. Sample sizes ranged from five to hundreds. Two of the studies focused on the same set of samples collected in North Canberra, Australia, but evaluated different methods of data analysis (de Caritat et al., 2022; de Caritat et al., 2021). As such, when discussing sample collection, preparation, and instrumental analysis, the two de Caritat et al. (2021), de Caritat et al. (2022) are referred to as one study, but when discussing data analysis and results they are referred to as two distinct studies.

Studies focus ranged from the soil type/environment when selecting where to sample, such as from different woodland areas, flowerbeds, and riverbanks (Baron et al., 2011) or from urban parks (Idrizi et al., 2021) to different areas across a region or country, such as different districts in Haryana, India (Sangwan et al., 2024), different sites across Australia (Woods et al., 2014), and different regions across a province in China (Liu et al., 2015).

3.2 Selecting a sampling area

Nine out of the 14 studies with unique sample sets (i.e., counting the two papers from de Caritat et al. as one unique sample set) collected multiple samples from within a marked-off area (i.e. 10 × 10 m area, 2 m radius, etc.) (Chauhan et al., 2018; Cox et al., 2000; de Caritat et al., 2022; de Caritat et al., 2021; Idrizi et al., 2021; Kizil et al., 2024; Koçak et al., 2021; Newland et al., 2022; Sangwan et al., 2024; Woods et al., 2014). Specifically, Chauhan et al. (2018) collected 10 samples (five from the surface and five from the sub-surface) from within a 10 × 10 m area in 18 different cities across five states in India. The surface and sub-surface samples were subsequently combined, resulting in two samples from each location (Chauhan et al., 2018). Cox et al. (2000) collected samples from 25 1.5 × 1.5 m subsections within a 7.5 × 7.5 m area in a field in Ashland, Oregon. Apart from these samples, Cox et al. (2000) also collected 65 samples from within a 32 km radius of the 7.5 × 7.5 m sampling area, and samples at varying depths from a third location several miles away. De Caritat et al. (2021), De Caritat et al. (2022) analyzed samples from 268 locations across North Canberra, Australia, with triplicates being collected at 28 of the locations for a total of 324 samples. At each location, five samples were collected from within a 1 × 1 m area and then subsequently combined (de Caritat et al., 2022; de Caritat et al., 2021). Three blind samples were also collected in order to evaluate the performance of the provenancing methods (de Caritat et al., 2022; de Caritat et al., 2021). Idrizi et al. (2021) collected 16 samples from within a 3 × 3 m area in five different parks in Tetovo, North Macedonia for a total of 80 samples. Three years after the initial sampling, Idrizi et al. (2021) collected twelve additional samples from three out of the five original parks (four samples from each park). Kizil et al. (2024) collected five samples from within a 1 × 1 m area in seven different regions at a distance of 1 m from the shore at Ömerli Dam in Istanbul, Turkey for a total of 35 samples. Koçak et al. (2021) collected five samples from within a 2 m radius in three different areas of New York for a total of 15 samples which were divided into subsamples for analysis. Newland et al. (2022) collected five samples from within a 30 × 30 cm area in nine locations across a Western Australian metropolitan region. Two of the samples from each location were randomly selected for analysis (Newland et al., 2022). Sangwan et al. (2024) collected eight samples (four surface and four sub-surface samples) from within a 10 × 10 m area in 29 different locations in Haryana, India for a total of 232 samples. The surface and sub-surface samples were subsequently combined, resulting in two samples from each location (Sangwan et al., 2024). Woods et al. (2014) collected 18 samples (nine surface and nine sub-surface samples) from a 5 × 5 m area in six sites around Canberra, Australia. The surface and sub-surface samples were subsequently combined, resulting in two samples from each location (Woods et al., 2014).

One study collected samples from four different flowerbeds, four different woodland areas, and four different riverbanks in Lincoln, United Kingdom (Baron et al., 2011). Five samples were collected in each location, spaced evenly apart along a straight line (Baron et al., 2011).

e Silva et al. (2022) collected 232 samples across the Curitiba Metropolitan Region of Brazil, with the addition of around 40 sub-surface samples and 14 samples from simulated evidence. Liu et al. (2015) collected 18 samples from different regions in the Shandong Province of China. Newland et al. (2023) collected five samples from four locations that were chosen for their apparent similarities in the Perth metropolitan region of Western Australia as part of a mock case. Xu et al. (2020) collected 100 samples from five major regions of China.

3.3 Sampling density

Four studies collected one sample from each corner of their sampling area as well as one sample from the center of their sampling area (Chauhan et al., 2018; de Caritat et al., 2022; de Caritat et al., 2021; Kizil et al., 2024; Newland et al., 2022). In one study, samples were collected from each corner of the sampling area (Sangwan et al., 2024). In another study, samples were collected at 50 cm intervals along a line (Baron et al., 2011).

In two studies, samples were collected at a certain sampling density within their study area, although the authors did not specify how the sample locations within the area were chosen (i.e., random, regularly spaced intervals, etc.) and may have been randomized (e Silva et al., 2022; Koçak et al., 2021). e Silva et al. (2022) collected 1.1–6.8 samples per square kilometer within their 100 km² study area while Koçak et al. (2021) collected five samples within a few 2 m radii. e Silva et al. (2022) did state that they avoided sampling in landfills, urban areas, and areas with crops, and tried to sample more in woodland field areas.

In three other studies, a similar approach was used in collecting a certain number of samples within a sample area, but the authors indicated that the collected samples were evenly spaced from each other (Cox et al., 2000; Idrizi et al., 2021; Woods et al., 2014). In one of those studies, 16 samples were collected approximately 1 m apart within a 9 × 9 m grid (Idrizi et al., 2021) while in another nine samples were collected at unspecified “regular” intervals in a 5 × 5 m grid (Woods et al., 2014). In a third study, samples were collected from 25 to 1.5 × 1.5 m subsections within a 7.5 × 7.5 m area (Cox et al., 2000). One of those subsections was divided into 16 equal parts, though the authors did not specify the sampling density or locations of the sample(s) in the other sections of the grid (Cox et al., 2000).

The remaining three papers did not specify sample density for their collected locations (Liu et al., 2015; Newland et al., 2023; Xu et al., 2020).

3.4 Sampling depth and collection method

In four studies, soil samples were collected from the surface and from a lower depth (Chauhan et al., 2018; e Silva et al., 2022; Sangwan et al., 2024; Woods et al., 2014). Two of those four studies did not give a value for the depth they considered to be “surface” or “topsoil” (Chauhan et al., 2018; Sangwan et al., 2024). The other two studies specified that their surface samples were collected at a depth of 0–5 cm (e Silva et al., 2022; Woods et al., 2014). For the lower depth samples, in one study, the samples were collected from a depth of 5–10 cm (Woods et al., 2014), another sampled from a depth of 15 cm (Sangwan et al., 2024), while another sampled from a depth of 1 m (Chauhan et al., 2018), and yet another sampled “road cuttings” of unspecified depth (e Silva et al., 2022).

Seven studies only collected soil samples from one depth (Baron et al., 2011; Cox et al., 2000; de Caritat et al., 2022; de Caritat et al., 2021; Idrizi et al., 2021; Kizil et al., 2024; Koçak et al., 2021; Newland et al., 2022). One study analyzed samples collected at a depth of 5 cm, specifying that the entire sample had the dimensions of 15x15x5 cm (de Caritat et al., 2022; de Caritat et al., 2021). Other authors sampled from varying depths or ranges. One study sampled from a depth of 0–0.5 inches or 1.27 cm (Koçak et al., 2021), another study sampled from a depth of 0–5 cm (Newland et al., 2022) and two studies sampled at a depth of 10 cm (Baron et al., 2011; Idrizi et al., 2021). One study collected samples at a depth of 20 cm (Kizil et al., 2024). One study collected a limited number of their samples from a depth of 0–3.4 m, though the majority of the samples were collected from an unspecified depth (Cox et al., 2000). The remaining three papers did not specify the sampling depth (Liu et al., 2015; Newland et al., 2023; Xu et al., 2020).

Four papers stated that surface debris (leaves, rocks, sticks, etc.) were partially, gently, or completely cleared from marked-off sampling sites (Chauhan et al., 2018; Newland et al., 2022; Sangwan et al., 2024; Woods et al., 2014). One study mentioned the removal of surface vegetation before collection of samples (de Caritat et al., 2022; de Caritat et al., 2021).

Out of the nine studies that make no mention of surface preparation, seven utilized sieving in sample preparation (Baron et al., 2011; Cox et al., 2000; e Silva et al., 2022; Kizil et al., 2024; Koçak et al., 2021; Newland et al., 2023; Xu et al., 2020). The remaining two studies made no mention of surface preparation or sieving (Idrizi et al., 2021; Liu et al., 2015).

The majority of studies examined in this review did not specify how the samples were collected. One set of authors used a shovel (Newland et al., 2022) and others stated that they used a core sampler (Baron et al., 2011). One study mentions the use of sterile plastic scoops for their surface samples but did not specify what they used to collect their lower depth samples (e Silva et al., 2022).

4 Sample preparation

4.1 Sample drying

Water in the soil can impact the collected spectra. Different groups applied different approaches to sample preparation including drying.

The authors of four out of the 14 studies air-dried their samples (Baron et al., 2011; Chauhan et al., 2018; Sangwan et al., 2024; Xu et al., 2020). In two out of these four studies, the samples were air-dried for at least 5 days (Chauhan et al., 2018; Sangwan et al., 2024); in one study, they were air-dried for 3 days (Baron et al., 2011), and the final of the four studies did not specify how long the samples were air-dried (Xu et al., 2020).

Four studies specifically mentioned the use of an oven to dry out samples – in one study it was set at 47 °C for 24–48 h (Woods et al., 2014), in another it was set at 40 °C for 48 h (de Caritat et al., 2022; de Caritat et al., 2021), in yet another, it was set at 100 °C for 48 h (Kizil et al., 2024), and a final study set the samples to dry at 110 °C to a constant weight (Koçak et al., 2021). Two other studies did not explicitly state using an oven but stated that samples were dried at 110 °C for an unspecified length of time (Idrizi et al., 2021) and at 120 °C for 12 h (Liu et al., 2015).

In one study, samples were air-dried after collection for an unspecified amount of time before they were further dried in an oven at 30 °C for 24 h (e Silva et al., 2022). After drying the samples in the oven, samples were separated into three fractions based on particle sizes and dried these fractions at 70 °C (e Silva et al., 2022). The finer fraction samples were heated at 650 °C for 3 h in a muffle furnace to remove the organic material (e Silva et al., 2022).

In one of the studies in which the samples were air-dried, the researchers also heated some of the soil samples at 650 °C for 15 min to examine the effects on the collected IR spectra (Chauhan et al., 2018). In another study, samples were dried in a microwave at 100 °C for 30 min, IR spectra were collected, and then samples were heated in a muffle furnace at 650 °C for 15 min and the IR spectra were collected again (Cox et al., 2000).

Most of the authors did not specify what they collected their samples in or how they stored them prior to sample preparation. Some authors reported storing samples in bags. Plastic bags were used in two studies (e Silva et al., 2022; Sangwan et al., 2024) and aluminum bags were used in another (Idrizi et al., 2021). The authors of another study stored samples in a clean plastic container in a −20 °C freezer before freeze-drying samples (Newland et al., 2022).

4.2 Sample sieving and homogenization

In twelve of the studies, sieves were used in the sample preparation (Baron et al., 2011; Chauhan et al., 2018; Cox et al., 2000; de Caritat et al., 2022; de Caritat et al., 2021; e Silva et al., 2022; Kizil et al., 2024; Koçak et al., 2021; Newland et al., 2023; Newland et al., 2022; Sangwan et al., 2024; Woods et al., 2014; Xu et al., 2020). The sizes of the sieves varied. Sieves of 38 µm (Woods et al., 2014), 50 µm (Cox et al., 2000; Kizil et al., 2024), 75 µm (de Caritat et al., 2022; de Caritat et al., 2021), 106 µm (Koçak et al., 2021), 125 µm (Baron et al., 2011), 2 mm (Baron et al., 2011; Chauhan et al., 2018; Newland et al., 2023; Newland et al., 2022; Sangwan et al., 2024; Xu et al., 2020), and 2.8 mm sizes were employed (Koçak et al., 2021), and one study used a “coarse” sieve of unspecified size (Cox et al., 2000). In one study, soil samples were wet sieved into three different particle size fractions (coarse, sandy, and fine) after chemically dispersing samples using tetrasodium pyrophosphate (e Silva et al., 2022). One sieve size was used in most of the studies, but four of them used a combination of a coarser sieve and a finer sieve (Baron et al., 2011; Cox et al., 2000; e Silva et al., 2022; Koçak et al., 2021).

In eight studies, samples were ground/homogenized in the sample preparation (Baron et al., 2011; Chauhan et al., 2018; Cox et al., 2000; e Silva et al., 2022; Koçak et al., 2021; Liu et al., 2015; Sangwan et al., 2024; Woods et al., 2014). In all of the studies in which samples were ground -- except for Liu et al. (2015) -- at least one sieve was used, although the order of the process varied (i.e., if samples were ground before or after sieving).

In two studies in which samples were collected from Perth, Australia, the authors separated out and analyzed quartz-recovered fine fractions from the soil (Newland et al., 2023; Newland et al., 2022). The authors sieved the soil samples using a 2 mm sieve and then hand-picked quartz grains out of the sieved samples under a microscope (Newland et al., 2023; Newland et al., 2022). The quartz grains were then subjected to a process involving sonication in deionized water, centrifugation of the liquid fraction, removal of the supernatant, and resuspension of the semi-solid layer, before drying the samples into thin films (Newland et al., 2023; Newland et al., 2022).

5 Data collection and analysis

The focus of this review is the use of FTIR spectroscopy for forensic geolocation. Some of the studies applied additional instruments as well. FTIR spectroscopy alone was used to analyze the samples in six studies (Baron et al., 2011; Chauhan et al., 2018; Idrizi et al., 2021; Koçak et al., 2021; Liu et al., 2015; Sangwan et al., 2024), while eight studies reported the use of at least one other type of instrumentation or analysis along with FTIR spectroscopy (Cox et al., 2000; de Caritat et al., 2022; de Caritat et al., 2021; e Silva et al., 2022; Kizil et al., 2024; Newland et al., 2023; Newland et al., 2022; Woods et al., 2014; Xu et al., 2020).

Out of the studies that used FTIR spectroscopy, eleven used an ATR attachment (Baron et al., 2011; Chauhan et al., 2018; de Caritat et al., 2022; de Caritat et al., 2021; e Silva et al., 2022; Kizil et al., 2024; Koçak et al., 2021; Newland et al., 2023; Newland et al., 2022; Sangwan et al., 2024; Woods et al., 2014; Xu et al., 2020). In three of the studies, samples were prepared and analyzed as KBr pellets (Cox et al., 2000; Idrizi et al., 2021; Liu et al., 2015). One study analyzed samples using both ATR and DRIFT techniques (Koçak et al., 2021).

5.1 FTIR settings

Four of the studies used a Perkin Elmer brand FTIR: two studies employed the Spectrum Two model (Chauhan et al., 2018; Sangwan et al., 2024) while one used the System 2000 model (Idrizi et al., 2021), and another used the Spectrum 100 model (Baron et al., 2011). In seven of the studies, Thermo Fisher Scientific brand instrument was used (Cox et al., 2000; de Caritat et al., 2022; de Caritat et al., 2021; e Silva et al., 2022; Koçak et al., 2021; Newland et al., 2023; Newland et al., 2022; Xu et al., 2020). Three studies used the Nicolet iS50 model (de Caritat et al., 2022; de Caritat et al., 2021; Newland et al., 2023; Newland et al., 2022), one used the Nicolet 380 model (e Silva et al., 2022), one used the Nicolet Nexus 670 model (Koçak et al., 2021), one used the Nicolet 5SXC model (Cox et al., 2000), and one used the TruDefender FT model (Xu et al., 2020). One study employed a Bruker AXS VERTEX 70 FTIR (Liu et al., 2015). One study used a Smiths Detection IdentifyIR portable FTIR (Woods et al., 2014). The remaining study did not specify the brand or model of the ATR-FTIR that was used (Kizil et al., 2024).

In the majority of the studies, data was collected from 4,000–400 cm^-1 (Baron et al., 2011; Chauhan et al., 2018; de Caritat et al., 2022; de Caritat et al., 2021; e Silva et al., 2022; Koçak et al., 2021; Liu et al., 2015; Newland et al., 2023; Newland et al., 2022; Sangwan et al., 2024). A few collected data from 4,000–650 cm^-1 (Kizil et al., 2024; Woods et al., 2014; Xu et al., 2020) and one collected data from 1,500–400 cm^-1 (Idrizi et al., 2021). The remaining study did not specify a range but, based on the graphs included in the paper, the collected range was from 4,000 to around 400 cm^-1 (Cox et al., 2000).

A resolution of 4 cm^-1 was used in eight of the studies (Baron et al., 2011; de Caritat et al., 2022; de Caritat et al., 2021; Idrizi et al., 2021; Koçak et al., 2021; Newland et al., 2023; Newland et al., 2022; Sangwan et al., 2024; Woods et al., 2014). One study used a resolution of 16 cm^-1 (Kizil et al., 2024), one used 8 cm^-1 (Chauhan et al., 2018), one used 2.87 cm^-1 (Xu et al., 2020), and the remaining three did not specify the resolution used (Cox et al., 2000; e Silva et al., 2022; Liu et al., 2015).

The number of scans collected included 16 (Chauhan et al., 2018; Sangwan et al., 2024), 32 (e Silva et al., 2022; Idrizi et al., 2021), 64 (de Caritat et al., 2022; de Caritat et al., 2021; Koçak et al., 2021; Newland et al., 2023; Newland et al., 2022; Woods et al., 2014), and 128 (Baron et al., 2011; Kizil et al., 2024). Three of the papers did not specify the number of scans they collected for each sample (Cox et al., 2000; Liu et al., 2015; Xu et al., 2020).

In five of the studies, all samples were run in triplicate on the FTIR (Baron et al., 2011; Newland et al., 2023; Newland et al., 2022; Sangwan et al., 2024; Woods et al., 2014). One study specified that a small portion of samples were run in triplicate (de Caritat et al., 2022; de Caritat et al., 2021) and one study specified that samples were run five times on the FTIR (Chauhan et al., 2018).

5.2 Data analysis

In twelve out of the fifteen papers, a statistical analysis method was used to interpret the data/build models (Baron et al., 2011; Chauhan et al., 2018; de Caritat et al., 2022; de Caritat et al., 2021; e Silva et al., 2022; Idrizi et al., 2021; Koçak et al., 2021; Newland et al., 2023; Newland et al., 2022; Sangwan et al., 2024; Woods et al., 2014; Xu et al., 2020). In the remaining three papers, the authors visually compared infrared spectra and identified peaks corresponding to certain soil components or chemical bonds (Cox et al., 2000; Kizil et al., 2024; Liu et al., 2015). Apart from these three papers, seven of the papers that used statistical analysis methods also identified spectral peaks corresponding to certain soil components or chemical bonds (Table 1) (Baron et al., 2011; e Silva et al., 2022; Koçak et al., 2021; Newland et al., 2022; Sangwan et al., 2024; Woods et al., 2014; Xu et al., 2020).

Table 1

Table 1. Assignment of infrared peaks for soil NOM moieties.

The specific spectral bands cited or reported in the papers are listed in Table 1.

Variations in NOM constitution and proportion as well as metal ion binding capacity and pH all impact the recorded ATR FTIR spectra. An infrared spectra representation with soil organic matter (SOM) contributions at various stretches from published work is shown in Figure 1.

Figure 1

Figure 1. ATR FTIR spectrum (“Pearson correlation between SOM content and FTIR-ATR spectral intensity”, reproduced from Xu et al., 2023, Figure 3, published under a creative commons license https://www.mdpi.com/2072-4292/15/4/1072).

All of the papers that used statistical analysis on their data also indicated employing some form of preprocessing of the data. Eight of the twelve papers normalized their data (Chauhan et al., 2018; e Silva et al., 2022; Idrizi et al., 2021; Newland et al., 2023; Newland et al., 2022; Sangwan et al., 2024; Woods et al., 2014; Xu et al., 2020). At least five of the twelve performed baseline correction on their data (Baron et al., 2011; Koçak et al., 2021; Newland et al., 2022; Newland et al., 2023; Sangwan et al., 2024). Xu et al. (2020) indicated baseline correcting their laser-induced breakdown spectroscopy (LIBS) data, but it was unclear if the same was applied to the FTIR data (Xu et al., 2020). Three papers included data smoothing (Sangwan et al., 2024; Woods et al., 2014; Xu et al., 2020). Six papers removed sections of the FTIR data before chemometric analysis (Baron et al., 2011; de Caritat et al., 2022; de Caritat et al., 2021; Newland et al., 2023; Newland et al., 2022). In both papers, Newland et al. (2022), Newland et al. (2023) removed the section 2350–1950 cm^-1 giving the reason that this area is associated with interference from the ATR crystal. In both papers, de Caritat et al. (2021), de Caritat et al. (2022) removed the section 2,749–1,800 cm^-1 designating it as a non-relevant interval. Baron et al. (2011) conducted chemometric analysis on both the full FTIR spectrum (in this case 4,000–400 cm^-1) and a reduced spectrum which removed the section 2,399–1,851 cm^-1. Chauhan et al. (2018) conducted chemometric analysis on the 1,800–426 cm^-1 range. Woods et al. (2014) calculated the 1st derivative of the collected FTIR spectra and Koçak et al. (2021) calculated the 2nd derivative of their FTIR spectra. Other preprocessing techniques included mathematically reducing the number of variables (Idrizi et al., 2021) and ATR correction (Baron et al., 2011).

The most common type of statistical analysis utilized was principal component analysis (PCA) (Baron et al., 2011; Chauhan et al., 2018; de Caritat et al., 2021; e Silva et al., 2022; Idrizi et al., 2021; Koçak et al., 2021; Newland et al., 2023; Newland et al., 2022; Sangwan et al., 2024; Woods et al., 2014; Xu et al., 2020). Three of these papers specified using the nonlinear iterative partial least squares (NIPALS) algorithm for principal component analysis (Newland et al., 2023; Newland et al., 2022; Baron et al., 2011).

In three studies, PCA was used in combination with linear discriminant analysis (LDA) (Baron et al., 2011; Newland et al., 2023; Sangwan et al., 2024). Two studies used both PCA and LDA, but the authors appeared to use them separately to provide different information rather than using the information obtained from PCA to directly inform the LDA model(s) (Koçak et al., 2021; Chauhan et al., 2018). Baron et al. (2011) and Koçak et al. (2021) also used partial least squares discriminant analysis (PLS-DA) and compared the performance of these models to their PCA-LDA/LDA models.

De Caritat et al. (2022) used a simultaneous multivariate provenancing approach (2021) and a sequential multivariate approach to analyze the same dataset. The sequential approach involved interpolating the properties obtained from the instrumental analyses between the sampled sites using inverse distance weighting (de Caritat et al., 2021). This ultimately produces a map of the area which can be utilized to evaluate which areas are most similar to an unknown sample of interest (de Caritat et al., 2021). Although it does not point to a single location, the tool can help narrow down the areas that should be investigated further (de Caritat et al., 2021). The simultaneous approach produces a similar map product but is made through calculating the degree of geochemical similarity (DOGS) and then interpolation rasters (de Caritat et al., 2022). e Silva et al. (2022) also evaluated two different provenancing approaches – search range intervals and n-dimensional Euclidean distances.

Four studies used the OMNIC software, either for collecting and storing the data or for processing the data in some way (de Caritat et al., 2022; de Caritat et al., 2021; Koçak et al., 2021; Woods et al., 2014). Three studies used QGIS Desktop for spatial analysis/provenancing methods (de Caritat et al., 2022; de Caritat et al., 2021; e Silva et al., 2022). Three studies described the use of Unscrambler X software for PCA or PCA-LDA (Newland et al., 2023; Newland et al., 2022; Sangwan et al., 2024). Sangwan et al. (2024) also used it for data preprocessing. Two studies indicated the use of Origin/OriginPro software (Koçak et al., 2021; Sangwan et al., 2024). Sangwan et al. (2024) utilized it for plotting the spectra for when conducting their visual assessments, while Koçak et al. (2021) included OriginPro in a list of software. Two studies employed a version of IBM SPSS either for statistical hypothesis testing or for conducting PCA/LDA (Chauhan et al., 2018; e Silva et al., 2022). Two studies employed R Studio (de Caritat et al., 2021; Koçak et al., 2021). Koçak et al. (2021) included the use of unspecified R Studio scripts in a list of software and did not specify its particular use, while de Caritat et al. (2021) mentioned the use of the R Studio package “ChemoSpec” for PCA. Two studies employed a version of MATLAB (Idrizi et al., 2021; Xu et al., 2020). Idrizi et al. (2021) used the Self-Organizing Maps Toolbox by Vestanto and the Genetic Algorithm Toolbox from Sheffield University in MATLAB for their statistical analyses while Xu et al. (2020) used MATLAB for preprocessing, statistical analyses, and PCA. Baron et al. (2011) used Perkin Elmer Spectrum software for preprocessing. Other software employed for various statistical analyses include Tanagra (Baron et al., 2011), Excel/spreadsheet applications (de Caritat et al., 2022), PAST (e Silva et al., 2022), ArcGIS Pro (e Silva et al., 2022), Geosoft Oasis Montaj (e Silva et al., 2022), and Minitab16 (Woods et al., 2014).

6 Results

Samples were collected, packaged, and analyzed from locations across the globe using infrared spectroscopy and the results are shown tabulated in Table 2. FTIR and ATR-FTIR spectroscopy were applied to sample analysis and geolocation. Sample collection depth, packaging, storage, sieving, and homogenization varied across the studies reviewed. Infrared spectroscopy data collection settings and parameters and post-processing statistical analysis also varied leading to varied discrimination outcomes.

Table 2

Table 2. Soil samples analyzed using FTIR, locations, and results.

The studies that utilized statistical analyses largely found relatively high levels of discriminating power/classification accuracy in the generated models. Woods et al. (2014) conducted statistical analysis on a composite surface and a composite sub-surface soil sample from six locations that the authors sampled around the Canberra area of Australia, as well as 17 samples from the Australian National Soil Archive collected by the Commonwealth Scientific and Industrial Research Organization (CSIRO). Woods et al. (2014) reported classification of the 29 samples into 28 groups using PCA on the FTIR spectra and 1st derivative spectra, and 99.7% discrimination after a pairwise comparison of all the samples. Idrizi et al. (2021) reported classification accuracies for their supervised self-organizing maps upon cross-validation to be from 96%–100% and found that the twelve samples collected 3 years after the original sampling were correctly grouped with their corresponding park. Newland et al. (2023) set up their study as a mock case with four different site samples – two alibi sites, the crime scene, and another potential site of interest. The authors’ goal was to assess if the sample supposed to have been recovered from a suspect could be matched to the crime scene in a blind analysis using microspectrophotometry, ATR FTIR, and X-ray diffraction along with chemometrics (Newland et al., 2023). Newland et al. (2023) reported 100% classification accuracy of the LDA model of the FTIR dataset when inputting the suspect recovered soil replicates as unknowns (i.e., the suspect recovered sample was correctly matched to the crime scene).

Baron et al. (2011) compared the success of two NIPALS-LDA models and two PLS-DA models when analyzing the full IR spectra and a truncated range. The first NIPALS-LDA and PLS-DA models were built to classify the type of land (i.e., flowerbed, woodland, or riverbank), while the second models were built to attempt to classify the specific site the samples were collected from (Baron et al., 2011). The land type NIPALS-LDA model performed better than the PLS-DA model when analyzing the full spectra (8.3% prediction error versus 10%) and performed equally well when analyzing the truncated spectra (8.3% error) (Baron et al., 2011). Both land type models were most successful in correctly classifying the woodland samples (96.67% accuracy) (Baron et al., 2011). The NIPALS-LDA model developed for site classification also performed better than the PLS-DA model when analyzing the full spectra (21.7% versus 31.7% error) and when analyzing the truncated spectra (23.3% versus 31.7% error) (Baron et al., 2011).

Koçak et al. (2021) compared the success of LDA models and PLS-DA models when analyzing ATR data and DRIFTS data. The LDA models had a 57% classification error (calculated using the leave-one-out method) when analyzing ATR data and a 27% classification error when analyzing DRIFTS data (Koçak et al., 2021). The PLS-DA models had root mean square error of predictions of 0.48 when analyzing ATR data and 0.38 when analyzing DRIFTS data (Koçak et al., 2021). Out of their three sampling locations, they found that location B could be distinguished from locations A and C, but locations A and C could not be statistically differentiated from each other (Koçak et al., 2021).

Two studies compared the discriminating power achieved by visually inspecting the spectra to the classification accuracy of a statistical model (Chauhan et al., 2018; Sangwan et al., 2024). Chauhan et al. (2018) reported the discriminating power of visual analysis to be 99.35% for surface samples and 97.38% for sub-surface samples. They reported the classification accuracies of the LDA models to be 88% for surface samples and 100% for sub-surface samples (Chauhan et al., 2018). The accuracy of the sub-surface model remained the same when applying a leave-one-out cross-validation to the model, but the surface model’s accuracy dropped to 77% (Chauhan et al., 2018). The authors noted that the lower accuracy of the surface model could be due to factors such as higher levels of transfer/change on the surface rather than farther down, though they did not appear to discuss why the visual analysis discriminated surface samples to a much higher degree (Chauhan et al., 2018). Sangwan et al. (2024) reported the discriminating power of visual analysis to be 73.39% for surface samples and 71.42% for sub-surface samples. They reported the classification accuracy of the PCA-LDA models to be 100% for surface samples and 98.85% for sub-surface samples (Sangwan et al., 2024). Five blind samples were used to test each of the models, with the surface model classifying 100% of the blinds correctly and the sub-surface model classifying 80% of the blinds correctly (Sangwan et al., 2024).

When utilizing the simultaneous approach of soil provenancing, de Caritat et al. (2022) found FTIR spectroscopy to be the tool with the highest average accuracy (26.6%) when predicting the possible locations of the blind samples. They found FTIR spectroscopy to have an average precision of 65.4%, somewhat lower than X-Ray fluorescence (de Caritat et al., 2022). When utilizing the sequential approach, de Caritat et al. (2021) did not calculate the accuracy or precision of FTIR alone but instead calculated the accuracy/precision of FTIR in combination with magnetic susceptibility. They found that all analytical methods (with principal components) performed equally in terms of accuracy (61.3%), but that FTIR spectroscopy with magnetic susceptibility (with principal components) was the most precise (80.7%) (de Caritat et al., 2021). Overall, de Caritat et al. (2022) reported that the sequential multivariate approach performed slightly better, but the simultaneous approach is easier to implement.

Newland et al. (2022) and Xu et al. (2020) found that FTIR spectroscopy was useful in distinguishing between soil samples but recommended pairing it with other analytical methods such as X-Ray diffraction (XRD) or LIBS for better results. Cox et al. (2000) found FTIR analysis of the organic portion of soils to be useful in discriminating samples where color analysis is inconclusive. The authors demonstrated this by selecting and comparing four samples with the sample Munsell color values (Cox et al., 2000). Liu et al. (2015) found FTIR to be useful in discriminating the 18 soil samples the authors collected and used the functional groups identified in each sample to create a searchable database for identification of unknowns. The authors did not appear to test or validate the database with unknowns (Liu et al., 2015). Two papers reported that FTIR spectroscopy did not provide sufficient or relevant information for the discrimination and provenancing of the analyzed soil samples (e Silva et al., 2022; Kizil et al., 2024).

7 Evidence-based recommendations

Sampling and analyzing surface or subsurface soil depends on the research question. Surface soil becomes embedded in footwear soles and is tracked between and among locations while sub-surface soil may be detected on soles or on clothing if perpetrator dug a grave to dispose of a body, for example. The standard FTIR recording range was most frequently 4,000–400 cm^-1 unless truncating for a specific reason with a resolution of 4 and 64 scans in triplicate. Groups used various FTIR instrument models manufactured by different companies. Many statistical tools were used with very high reported accuracy (Table 2); more usage cases are needed to determine the best method. Some authors also reported high success with visual discrimination of the FTIR spectra as well. Recommended pairing FTIR with another tool for increased discrimination. The more varied the locations in question (e.g., sand, clay, loam, etc.) including at short and long geographic distances led to better discrimination accuracy. Well-developed databases are expected to improve prediction accuracy.

8 Conclusion

In conclusion, ATR FTIR spectroscopy has been found to be a highly accurate tool that can be used to generate a spectral fingerprint of soil and differentiate soils from a variety of environments for geolocation. We have reviewed numerous papers that focused on geolocation of soil samples including providing a detailed comparison of sampling methods, machine learning techniques, and preprocessing approaches used in the studies reviewed. The authors wrote of the difficult samples that they collected to best evaluate techniques and methods for forensic casework samples. Visual inspection was employed by some groups but could be unreliable as it is based upon the skill/training of the examiner. Statistical analysis removes interpreter variation and was shown to lead to high discriminating power. Higher discrimination was observed when multiple instrumental and analysis tools were employed.

Variations in NOM constitution, clays, salts, and mineral components of the soil impact the recorded ATR FTIR spectral stretches. Various temperatures and lengths of time were reported for drying; high temperatures and long times could alter the constituency of the organic portion. The recorded resolution varied; higher resolution can be important in capturing variations that could be missed with lower resolution. Sample depth will impact what is detected and applications for forensic cases such as victims discarded in shallow graves versus surface trafficking and movement.

Overall, it is evident that a systematic and consensus approach for sampling, preprocessing and data collection would accelerate data comparisons and conclusions that can be made from soil investigations world-wide. An interlaboratory study in which a specified procedure is used by all participants with their samples and compilation in a searchable and mapped database (based on GPS coordinates) would aid in determining the broad conclusions that can be made about a sample set, particularly if surface or subsurface soils can be differentiated in a particular region. In a mock case study, locations were able to be differentiated.

Author contributions

JF: Data curation, Formal Analysis, Investigation, Writing – original draft, Writing – review and editing. KE: Conceptualization, Formal Analysis, Writing – original draft, Writing – review and editing, Project administration.

Funding

The authors declare that no financial support was received for the research and/or publication of this article.

Conflict of interest

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Correction note

This article has been corrected with minor changes. These changes do not impact the scientific content of the article.

Generative AI statement

The authors declare that no Generative AI was used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

References

Baron, M., Gonzalez-Rodriguez, J., Croxton, R., Gonzalez, R., and Jimenez-Perez, R. (2011). Chemometric study on the forensic discrimination of soil types using their infrared spectral characteristics. Appl. Spectrosc. 65 (10), 1151–1161. doi:10.1366/10-06197

PubMed Abstract | CrossRef Full Text | Google Scholar

Bergslien, E. (2012). “A brief history of forensic science and crime scene basics,” in An introduction to forensic geoscience. Editor E. Bergslien (John Wiley and Sons).

Google Scholar

Chauhan, R., Kumar, R., and Sharma, V. (2018). Soil forensics: a spectroscopic examination of trace evidence. Microchem. J. 139, 74–84. doi:10.1016/j.microc.2018.02.020

CrossRef Full Text | Google Scholar

Cox, R. J., Peterson, H. L., Young, J., Cusik, C., and Espinoza, E. O. (2000). The forensic analysis of soil organic by FTIR. Forensic Sci. Int. 108 (2), 107–116. doi:10.1016/s0379-0738(99)00203-0

PubMed Abstract | CrossRef Full Text | Google Scholar

de Caritat, P., Woods, B., Simpson, T., Nichols, C., Hoogenboom, L., Ilheo, A., et al. (2021). Forensic soil provenancing in an urban/suburban setting: a sequential multivariate approach. J. Forensic Sci. 66 (5), 1679–1696. doi:10.1111/1556-4029.14727

PubMed Abstract | CrossRef Full Text | Google Scholar

de Caritat, P., Woods, B., Simpson, T., Nichols, C., Hoogenboom, L., Ilheo, A., et al. (2022). Forensic soil provenancing in an urban/suburban setting: a simultaneous multivariate approach. J. Forensic Sci. 67 (3), 927–935. doi:10.1111/1556-4029.14967

PubMed Abstract | CrossRef Full Text | Google Scholar

Elkins, K. M. (2011). Rapid presumptive “fingerprinting” of body fluids and materials by ATR FT-IR spectroscopy. J. Forensic Sci. 56, 1580–1587. doi:10.1111/j.1556-4029.2011.01870.x

PubMed Abstract | CrossRef Full Text | Google Scholar

Elkins (2019). Introduction to forensic chemistry. Boca Raton, FL: Taylor and Francis/CRC Press.

Google Scholar

e Silva, M. P. N., Guedes, C. C. F., Melo, V. de F., Mascarenhas, R. de O., and Salvador, F. A. S. (2022). Evaluating geostatistical methods along with semi-destructive analysis for forensic provenancing organic-rich soils in humid subtropical climate. Forensic Sci. Int. 341, 111508. doi:10.1016/j.forsciint.2022.111508

PubMed Abstract | CrossRef Full Text | Google Scholar

Idrizi, H., Najdoski, M., and Kuzmanovski, I. (2021). Classification of urban soils for forensic purposes using supervised self-organizing maps. J. Chemom. 35 (4), e3328. doi:10.1002/cem.3328

CrossRef Full Text | Google Scholar

Kizil, S., Boler, I. S., and Atasoy, S. (2024). Forensic comparison of soil samples in omerli dam by FTIR and ICP-OES. Sak. Uni. J. Sci. 28 (4), 794–803. doi:10.16984/saufenbilder.1453097

CrossRef Full Text | Google Scholar

Koçak, A., Wyatt, W., and Comanescu, M. A. (2021). Comparative study of ATR and DRIFT infrared spectroscopy techniques in the analysis of soil samples. Forensic Sci. Int. 328, 111002. doi:10.1016/j.forsciint.2021.111002

PubMed Abstract | CrossRef Full Text | Google Scholar

Liu, Y., Li, Q., Li, Y., Bao, J., Hu, Z., Hao, D., et al. (2015). Detection of nanoscale soil organic matter by infrared spectrum for forensic science. J. Chem. 1, 189421. doi:10.1155/2015/189421

CrossRef Full Text | Google Scholar

Newland, T. G., Pitts, K., and Lewis, S. W. (2022). Multimodal spectroscopy with chemometrics for the forensic analysis of Western Australian sandy soils. Forensic Chem. 28, 100412. doi:10.1016/j.forc.2022.100412

CrossRef Full Text | Google Scholar

Newland, T. G., Pitts, K., and Lewis, S. W. (2023). Multimodal spectroscopy with chemometrics: application to simulated forensic soil casework. Forensic Chem. 33, 100481. doi:10.1016/j.forc.2023.100481

CrossRef Full Text | Google Scholar

Sangwan, P., Nimi, C., Nain, T., Singh, R., and Sharma, N. (2024). Discrimination of soil samples collected from Haryana (India) using non-destructive ATR-FTIR spectroscopy coupled with multivariate statistical analysis. Indian J. Sci. Technol. 17 (11), 1087–1096. doi:10.17485/IJST/v17i11.2930

CrossRef Full Text | Google Scholar

Woods, B., Lennard, C., Kirkbride, P., and Robertson, J. (2014). Soil examination for a forensic trace evidence laboratory-part 1: spectroscopic techniques. Forensic Sci. Int. 245, 187–194. doi:10.1016/j.forsciint.2014.08.009

PubMed Abstract | CrossRef Full Text | Google Scholar

Xu, X., Du, C., Ma, F., Shen, Y., and Zhou, J. (2020). Forensic soil analysis using laser-induced breakdown spectroscopy (LIBS) and Fourier transform infrared total attenuated reflectance spectroscopy (FTIR-ATR): principles and case studies. Forensic Sci. Int. 310, 110222. doi:10.1016/j.forsciint.2020.110222

PubMed Abstract | CrossRef Full Text | Google Scholar

Xu, X., Du, C., Ma, F., Qiu, Z., and Zhou, J. (2023). A framework for high-resolution mapping of Soil Organic Matter (SOM) by the integration of fourier mid-infrared attenuation total reflectance spectroscopy (FTIR-ATR), Sentinel-2 images, and DEM derivatives. Remote Sens. 15 (4), 1072. doi:10.3390/rs15041072

CrossRef Full Text | Google Scholar