J Forensic Sci, March 2014, Vol. 59, No. 2 doi: 10.1111/1556-4029.12319 Available online at: onlinelibrary.wiley.com
TECHNICAL NOTE CRIMINALISTICS; ANTHROPOLOGY Bjørn K. Alsberg,1 Ph.D.; and Jørgen Rosvold,2 Ph.D.
Rapid Localization of Bone Fragments on Surfaces using Back-Projection and Hyperspectral Imaging*
ABSTRACT: Manual localization of bone fragments on the ground or on complex surfaces in relation to accidents or criminal activity may
be time-consuming and challenging. It is here investigated whether combining a near-infrared hyperspectral camera and chemometric modeling with false color back-projection can be used for rapid localization of bone fragments. The approach is noninvasive and highlights the spatial distribution of various compounds/properties to facilitate manual inspection of surfaces. Discriminant partial least squares regression is used to classify between bone and nonbone spectra from the hyperspectral camera. A predictive model (>95% prediction ability) is constructed from raw chicken bones mixed with stone, sand, leaves, moss, and wood. The model uses features in the near-infrared spectrum which may be selective for bones in general and is able to identify a wide variety of bones from different animals and contexts, including aged and weathered bone.
KEYWORDS: forensic science, bones, bone fragments, hyperspectral camera, in situ visualization, chemical imaging, chemical images, projection, chemometrics, near-infrared spectroscopy, PryJector, archeology
The localization of bone fragments is important in different disciplines such as forensics (1), archeology (2–4), food science (5), and medicine (6). The standard method for localizing bone fragments is by manual inspection, which is very time-consuming and inefficient for larger investigations. Depending on the environmental conditions (7), most parts of a dead body decompose relatively quickly when left in natural surroundings. Bones and teeth, however, have the advantage of being composed of materials with very tough properties, in some cases, being able to survive for many thousands of years in the soil (8,9). Most of the bone structure is made up of a composite matrix of mineral (mostly calcium phosphates in the form of apatite) and organic components (mostly collagen fibers), which are highly resistant to physical stress, microbial breakdown, and chemical disintegration (2,10). However, as an effect of different mechanical treatments, such as trampling, animal gnawing, butchering, explosions, etc., bone remains are often fragmented and dispersed. Different types of soil, weathering, and chemical treatments may also alter the color and texture of bones which make them difficult to distinguish from the surroundings (10). Several methods have been developed to ease the recovery of small bones such as dry or wet sieving (11,12) or floatation techniques using water or different chemicals (13). These
1 Department of Chemistry, Norwegian University of Science and Technology (NTNU), N-7491, Trondheim, Norway. 2 Section of Natural History, NTNU University Museum, N-7491, Trondheim, Norway. *Financial support provided by the Department of Chemistry at NTNU. Received 27 Aug. 2012; and in revised form 4 Dec. 2012; accepted 16 Dec. 2012.
474
methods are, however, relatively time-consuming, and in some cases, adding water or chemicals to the bones could hamper the sampling of DNA at a later stage (14). The aim of this study was therefore to determine whether it is possible to create a more time-efficient technique to identify bone fragments on a heterogeneous surface in situ. Previous research has shown that in particular vibrational spectroscopies such as infrared and Raman are useful techniques for detection and characterization of bones (15–20). These results make it reasonable to expect that an improved technology for localization of bone fragments may be based on rapid spectroscopic analyses of surfaces. Thus, it is here investigated whether a recently invented instrument (21) (called the PryJector) that utilizes a combination of spectroscopy, imaging, chemometric modeling, and false color projection can be used for this purpose.
Materials and Methods The PryJector Instrument A PryJector (21) is an instrument for making object properties visible in situ on surfaces which are otherwise invisible or difficult for human eyes to detect. This is accomplished using a combination of hyperspectral imaging, chemometric modeling, and back-projection of spatially distributed chemical information as false colored chemical images using an ordinary computer projector (or with lasers). Hyperspectral cameras record a full spectrum in each pixel of a scene, and multivariate data analytical methods are commonly used to extract the relevant chemical information. Here, a multivariate chemometric method (22) will be investigated whether it is able to discriminate between pixels © 2013 American Academy of Forensic Sciences
ALSBERG AND ROSVOLD
of bone and nonbone origin. Each pixel is thus given a class membership, bone or nonbone, and gives rise to what is referred to as a chemical image. These images contain chemically relevant information in each pixel and are projected, using an ordinary computer projector, back onto the surface of study such that they become available to the user in situ. The current PryJector used for the experiments is a tabletop version with a push-broom hyperspectral camera (Fig. 1) where the camera and light source are mounted on a common translation stage. A HySpex SWIR-320i hyperspectral camera from Norsk Elektro Optikk AS is used which is based on an InGaAs focal plane array (FPA) with a spectral range of 930–1670 nm. The number of across track spatial pixels is 320, and the number of recorded wavelengths is 148. An ordinary computer projector (Hewlett Packard MP3222 [Hewlett-Packard Co., Palo Alto, CA] with XGA, 1024 9 768 and 2000 lumens) is used to back-project the chemical image. The projection of chemical images is performed continuously where color combinations are chosen such that they appear different from each other and the surrounding environment. The high brightness of the projected images means that in practice there are no problems with visually discriminating a point which is highlighted by the projector from one which is not. Here, it is decided to use red color to signify the presence of bone at a pixel. The hyperspectral scan speed is 7.5 cm per second with 1 ms integration time and 100 ms frame time (21). Calibration Samples The calibration samples collected were directly used in the creation of the chemometric classification model. To simplify the task of creating a discriminating model, a single animal bone type was selected for this purpose: untreated chicken (Gallus gallus) bone. Three bags of chicken legs were purchased from different supermarkets in Trondheim, Norway. From each chicken leg, meat, blood, fat, and tendons were mechanically removed to expose the bone. After this treatment, each chicken bone was placed in a food drier for several days to remove as
.
RAPID LOCALISATION OF BONE FRAGMENTS
475
much moisture as possible. A total of 38 chicken leg bones were prepared that measured 5.0 1.5 cm in length. Half of the bones were crushed into smaller fragments measuring from 1–15 mm (Fig. 2). The main reason for using crushed bone fragments was to average out any changes in the spectra due to surface texture and contaminants (such as remnants of blood, meat, tendons, etc.). Detection of bone fragments can be made for a large number of possible surfaces, however, here, we have focused in particular on forest grounds as they are often relevant for forensic applications. Thus, nonbone samples that typically are expected to be present on a forest ground were also collected such as soil, sand, stones, leaves, moss, and bark. Each sample type was recorded under both wet and dry conditions. Wet conditions were created by spraying water droplets on the surface of the samples. The hyperspectral camera was then used to capture images of each of these samples and subsequently analyzed. Test Samples A separate set of samples were collected for test purposes only, that is, hyperspectral images of these objects were not part of the calibration process. The test objects are as follows: (i) a simulated forest ground created by adding chicken bones, soil, sand, stones, leaves, moss, and bark in a cardboard box measuring 24 9 35 9 11 cm (Fig. 3); and (ii) a set of animal bones of different species, contexts, and ages collected from the natural history collections at the NTNU Museum of Natural History and Archaeology (Table 1). The bottom of the cardboard box was covered with black plastic to prevent leakage of water. Soil was first added in the box, followed by a random mix of sand, stones, leaves, wood fragments, moss, and bark. The chicken bone fragments were placed on top in a pattern simple to recognize. As also for the calibration objects, the test samples were recorded under both wet and dry conditions and subjected to hyperspectral image recording and analysis. Data Analytical Methods There are many multivariate classification methods which could have been used for the current study, such as artificial
FIG. 1––Overview of the PryJector system.
FIG. 2––The image shows the box containing the crushed bone fragments which were recorded by a hyperspectral camera to be part of the calibration data set.
476
JOURNAL OF FORENSIC SCIENCES
spectrum belonging to an unknown class is encountered. In such a case, the new spectrum should not be assigned to any of the classes. To rectify this problem, the class prediction from DPLSR is further checked by calculating the Mahalanobis (25) distance from the center of a class j to the new spectrum. If this distance is too large, the membership will be changed into the background class, that is, not shown in the projection. The Mahalanobis cut-off values for each class have here been determined by visual inspection of predicted chemical images where the positions of the class pixels are known. Validation
FIG. 3––This is an image of the box containing the simulated forest ground with leaves, moss, plants, stones, sand, soil and bone fragments which were used to test the optimal classification model.
TABLE 1––List of sampled specimens.
Species Arctic fox (Vulpes lagopus) Brown bear (Ursus arctos) Domestic pig (Sus scrofa) Domestic sheep (Ovis aries) Goose (Anser sp.) Gray seal (Halichoerus grypus) Macaque (Macaca sp.) Norway lemming (Lemmus lemmus)
Skeletal Material
Context
Age (year)
Bone and teeth
Naturally decomposed
>25
Bone
Naturally decomposed
>30
Bone and teeth
Naturally decomposed
>35
Bone
Boiled and dried
5 c. 55
Bone and teeth Bone and teeth
Naturally decomposed Digested, owl pellet
>50 >60
This table summarizes some properties of specimens which were used to test the optimal classification model on bone types different from that used in the calibration.
neural networks (23) or support vector machines (SVM) (24). Here, the discriminant partial least squares (DPLSR) regression (22) is used as it is effective and simple to use. The DPLSR method is based on the PLS2 algorithm using multiple dependent variables (referred to as the Y-variables). Here, a data matrix (the X-matrix) is constructed where each row contains a single spectrum (from one pixel in a hyperspectral image). The wavelength intensity values (one for each column in the Xmatrix) from the spectra represent the independent variables. The Y-variables (the dependent variables) are contained in a matrix which encodes the different class memberships for each object. Here, a binary encoding is used where the Y column index signifies the class index. An object m belonging to class i is stored in row m where column i contains the value 1, and all other columns for this row are zero. The PLSR algorithm computes latent variables which are directed along the maximum X–Y covariance (22). In prediction, the class memberships are determined from which Y column index corresponds to the largest value. By default, DPLSR will force a new spectrum to be a member of one of the k classes. This is undesirable for cases where a
Two methods are used here to validate the classification models generated by DPLSR. The first method is based on using an independent validation data set to determine the optimal number of latent variables in DPLSR. In general, this is regarded the best method for validation of empirical models and is possible here as hyperspectral images contain a large number of spectra. The validation procedure proceeds as follows: • In each hyperspectral image, different and nonoverlapping regions of the same object type are defined as belonging to “calibration” and “validation” data. All spectra within these regions are subsequently collected into the calibration and validation data sets, respectively. • DPLSR classification models created on the basis of the calibration data are then applied to the set of validation spectra for different number of PLS factors (factors are in range a = 1,2,..,20) and used to determine the optimal number of PLSR components needed. The classification error used here is defined to be the percentage of wrong classifications and not the root mean square error of prediction (RMSP) which is common to use in regression problems. The second validation method employed is based on visual inspection of predicted chemical images using the optimal DPLSR model from the steps described above. This method is very effective for evaluating the predictive ability of chemometric models producing chemical images from hyperspectral images because a priori knowledge of the spatial distribution of the target class is often available. This is particularly relevant for the current study as we knew a priori which parts were related to bone and which were not on the calibration and test set objects described above. Set-up, Preprocessing and Analysis All prediction models were constructed using MATLAB 7.1 (The MathWorks, Inc. 3 Apple Hill Drive, Natick MA 017602098, USA) running on both Windows XP and Linux operating systems (Ubuntu 10.04.4 LTS, Canonical Ltd., London, UK). The Windows platform is used for the direct control and operation of the PryJector instrument. All off-line data analyses are performed on Linux workstations using inhouse chemometric software. Spectra from the hyperspectral images are numerically differentiated to first order using a 3rd order polynomial Savitzky– Golay filter with a 15-point window (26) and normalized to unit sum before being subjected to data analysis. For numerical differentiation, the savgol routine in the PLS_Toolbox 6.2 from Eigenvector Research Inc. (3905 West Eaglerock Drive Wenatchee, WA, USA) was used.
ALSBERG AND ROSVOLD
.
RAPID LOCALISATION OF BONE FRAGMENTS
477
Results Near-infrared light is defined to be in the region 730– 2300 nm and has been demonstrated to be useful for quantitative determination of different biological components such as total protein (27), albumin (28), and cholesterol (29). These predictions were performed in the so-called combination region (2000– 2500 nm) of the infrared part of the electromagnetic spectrum. The overtone region (800–1900), however, which is also used for the PryJector instrument, is more problematic due to the presence of strong absorption bands from water which have two dominating peaks around 1440 and 1930 nm. As water is considered a major interferent for this wavelength region, it was initially attempted to carry out the classification of bone fragments from nonbone fragments in dry conditions only. Each bone fragment was carefully dried in a food dryer, and samples of nonbone fragments were also kept in dry conditions. However, the resulting DPLSR models (not shown) from these experiments indicated that water was most likely still being used as the selective factor in discriminating between bone and nonbone fragments. For instance, we observed that pixels in the hyperspectral image of pure water droplets were wrongly predicted to belong to the bone class. These results suggest that the drying of the bone fragments was not sufficient, and thus, water directly interfered with the initial chemometric modeling. A significant amount of water is usually bound to the matrix of fresh bone where some of this is relatively freely bound and will evaporate at fairly low temperatures (below 105°C), while the rest is tightly bound structural water (30). Exposed bones will naturally dehydrate by weathering, but the more tightly bound water only evaporates by prolonged heating at higher temperatures (31). It is therefore likely that bound water was originally being used in the first model. On basis of this result, it was decided to design experiments such that water was being excluded from being used to discriminate between bone and nonbone samples. The approach taken was to record the bone and nonbone samples under both wet and dry conditions. Wet conditions were ensured by spraying water on the objects. Any successful DPLSR model from such a calibration data set would thus most likely not be using free water contribution to enable discrimination. To test whether such an approach would work, a DPLSR analysis was performed on a single hyperspectral image (size 320 9 400 9 148 elements) of a set of bones, stones, and wood fragments under both dry and wet conditions. Pixels in the hyperspectral image belonging to bone and nonbone were localized manually and divided into calibration and validation data matrices. The calibration and validation data matrices contained 10964 and 10963 spectra, respectively, and thus constitute
TECHNICAL NOTE CRIMINALISTICS; ANTHROPOLOGY Bjørn K. Alsberg,1 Ph.D.; and Jørgen Rosvold,2 Ph.D.
Rapid Localization of Bone Fragments on Surfaces using Back-Projection and Hyperspectral Imaging*
ABSTRACT: Manual localization of bone fragments on the ground or on complex surfaces in relation to accidents or criminal activity may
be time-consuming and challenging. It is here investigated whether combining a near-infrared hyperspectral camera and chemometric modeling with false color back-projection can be used for rapid localization of bone fragments. The approach is noninvasive and highlights the spatial distribution of various compounds/properties to facilitate manual inspection of surfaces. Discriminant partial least squares regression is used to classify between bone and nonbone spectra from the hyperspectral camera. A predictive model (>95% prediction ability) is constructed from raw chicken bones mixed with stone, sand, leaves, moss, and wood. The model uses features in the near-infrared spectrum which may be selective for bones in general and is able to identify a wide variety of bones from different animals and contexts, including aged and weathered bone.
KEYWORDS: forensic science, bones, bone fragments, hyperspectral camera, in situ visualization, chemical imaging, chemical images, projection, chemometrics, near-infrared spectroscopy, PryJector, archeology
The localization of bone fragments is important in different disciplines such as forensics (1), archeology (2–4), food science (5), and medicine (6). The standard method for localizing bone fragments is by manual inspection, which is very time-consuming and inefficient for larger investigations. Depending on the environmental conditions (7), most parts of a dead body decompose relatively quickly when left in natural surroundings. Bones and teeth, however, have the advantage of being composed of materials with very tough properties, in some cases, being able to survive for many thousands of years in the soil (8,9). Most of the bone structure is made up of a composite matrix of mineral (mostly calcium phosphates in the form of apatite) and organic components (mostly collagen fibers), which are highly resistant to physical stress, microbial breakdown, and chemical disintegration (2,10). However, as an effect of different mechanical treatments, such as trampling, animal gnawing, butchering, explosions, etc., bone remains are often fragmented and dispersed. Different types of soil, weathering, and chemical treatments may also alter the color and texture of bones which make them difficult to distinguish from the surroundings (10). Several methods have been developed to ease the recovery of small bones such as dry or wet sieving (11,12) or floatation techniques using water or different chemicals (13). These
1 Department of Chemistry, Norwegian University of Science and Technology (NTNU), N-7491, Trondheim, Norway. 2 Section of Natural History, NTNU University Museum, N-7491, Trondheim, Norway. *Financial support provided by the Department of Chemistry at NTNU. Received 27 Aug. 2012; and in revised form 4 Dec. 2012; accepted 16 Dec. 2012.
474
methods are, however, relatively time-consuming, and in some cases, adding water or chemicals to the bones could hamper the sampling of DNA at a later stage (14). The aim of this study was therefore to determine whether it is possible to create a more time-efficient technique to identify bone fragments on a heterogeneous surface in situ. Previous research has shown that in particular vibrational spectroscopies such as infrared and Raman are useful techniques for detection and characterization of bones (15–20). These results make it reasonable to expect that an improved technology for localization of bone fragments may be based on rapid spectroscopic analyses of surfaces. Thus, it is here investigated whether a recently invented instrument (21) (called the PryJector) that utilizes a combination of spectroscopy, imaging, chemometric modeling, and false color projection can be used for this purpose.
Materials and Methods The PryJector Instrument A PryJector (21) is an instrument for making object properties visible in situ on surfaces which are otherwise invisible or difficult for human eyes to detect. This is accomplished using a combination of hyperspectral imaging, chemometric modeling, and back-projection of spatially distributed chemical information as false colored chemical images using an ordinary computer projector (or with lasers). Hyperspectral cameras record a full spectrum in each pixel of a scene, and multivariate data analytical methods are commonly used to extract the relevant chemical information. Here, a multivariate chemometric method (22) will be investigated whether it is able to discriminate between pixels © 2013 American Academy of Forensic Sciences
ALSBERG AND ROSVOLD
of bone and nonbone origin. Each pixel is thus given a class membership, bone or nonbone, and gives rise to what is referred to as a chemical image. These images contain chemically relevant information in each pixel and are projected, using an ordinary computer projector, back onto the surface of study such that they become available to the user in situ. The current PryJector used for the experiments is a tabletop version with a push-broom hyperspectral camera (Fig. 1) where the camera and light source are mounted on a common translation stage. A HySpex SWIR-320i hyperspectral camera from Norsk Elektro Optikk AS is used which is based on an InGaAs focal plane array (FPA) with a spectral range of 930–1670 nm. The number of across track spatial pixels is 320, and the number of recorded wavelengths is 148. An ordinary computer projector (Hewlett Packard MP3222 [Hewlett-Packard Co., Palo Alto, CA] with XGA, 1024 9 768 and 2000 lumens) is used to back-project the chemical image. The projection of chemical images is performed continuously where color combinations are chosen such that they appear different from each other and the surrounding environment. The high brightness of the projected images means that in practice there are no problems with visually discriminating a point which is highlighted by the projector from one which is not. Here, it is decided to use red color to signify the presence of bone at a pixel. The hyperspectral scan speed is 7.5 cm per second with 1 ms integration time and 100 ms frame time (21). Calibration Samples The calibration samples collected were directly used in the creation of the chemometric classification model. To simplify the task of creating a discriminating model, a single animal bone type was selected for this purpose: untreated chicken (Gallus gallus) bone. Three bags of chicken legs were purchased from different supermarkets in Trondheim, Norway. From each chicken leg, meat, blood, fat, and tendons were mechanically removed to expose the bone. After this treatment, each chicken bone was placed in a food drier for several days to remove as
.
RAPID LOCALISATION OF BONE FRAGMENTS
475
much moisture as possible. A total of 38 chicken leg bones were prepared that measured 5.0 1.5 cm in length. Half of the bones were crushed into smaller fragments measuring from 1–15 mm (Fig. 2). The main reason for using crushed bone fragments was to average out any changes in the spectra due to surface texture and contaminants (such as remnants of blood, meat, tendons, etc.). Detection of bone fragments can be made for a large number of possible surfaces, however, here, we have focused in particular on forest grounds as they are often relevant for forensic applications. Thus, nonbone samples that typically are expected to be present on a forest ground were also collected such as soil, sand, stones, leaves, moss, and bark. Each sample type was recorded under both wet and dry conditions. Wet conditions were created by spraying water droplets on the surface of the samples. The hyperspectral camera was then used to capture images of each of these samples and subsequently analyzed. Test Samples A separate set of samples were collected for test purposes only, that is, hyperspectral images of these objects were not part of the calibration process. The test objects are as follows: (i) a simulated forest ground created by adding chicken bones, soil, sand, stones, leaves, moss, and bark in a cardboard box measuring 24 9 35 9 11 cm (Fig. 3); and (ii) a set of animal bones of different species, contexts, and ages collected from the natural history collections at the NTNU Museum of Natural History and Archaeology (Table 1). The bottom of the cardboard box was covered with black plastic to prevent leakage of water. Soil was first added in the box, followed by a random mix of sand, stones, leaves, wood fragments, moss, and bark. The chicken bone fragments were placed on top in a pattern simple to recognize. As also for the calibration objects, the test samples were recorded under both wet and dry conditions and subjected to hyperspectral image recording and analysis. Data Analytical Methods There are many multivariate classification methods which could have been used for the current study, such as artificial
FIG. 1––Overview of the PryJector system.
FIG. 2––The image shows the box containing the crushed bone fragments which were recorded by a hyperspectral camera to be part of the calibration data set.
476
JOURNAL OF FORENSIC SCIENCES
spectrum belonging to an unknown class is encountered. In such a case, the new spectrum should not be assigned to any of the classes. To rectify this problem, the class prediction from DPLSR is further checked by calculating the Mahalanobis (25) distance from the center of a class j to the new spectrum. If this distance is too large, the membership will be changed into the background class, that is, not shown in the projection. The Mahalanobis cut-off values for each class have here been determined by visual inspection of predicted chemical images where the positions of the class pixels are known. Validation
FIG. 3––This is an image of the box containing the simulated forest ground with leaves, moss, plants, stones, sand, soil and bone fragments which were used to test the optimal classification model.
TABLE 1––List of sampled specimens.
Species Arctic fox (Vulpes lagopus) Brown bear (Ursus arctos) Domestic pig (Sus scrofa) Domestic sheep (Ovis aries) Goose (Anser sp.) Gray seal (Halichoerus grypus) Macaque (Macaca sp.) Norway lemming (Lemmus lemmus)
Skeletal Material
Context
Age (year)
Bone and teeth
Naturally decomposed
>25
Bone
Naturally decomposed
>30
Bone and teeth
Naturally decomposed
>35
Bone
Boiled and dried
5 c. 55
Bone and teeth Bone and teeth
Naturally decomposed Digested, owl pellet
>50 >60
This table summarizes some properties of specimens which were used to test the optimal classification model on bone types different from that used in the calibration.
neural networks (23) or support vector machines (SVM) (24). Here, the discriminant partial least squares (DPLSR) regression (22) is used as it is effective and simple to use. The DPLSR method is based on the PLS2 algorithm using multiple dependent variables (referred to as the Y-variables). Here, a data matrix (the X-matrix) is constructed where each row contains a single spectrum (from one pixel in a hyperspectral image). The wavelength intensity values (one for each column in the Xmatrix) from the spectra represent the independent variables. The Y-variables (the dependent variables) are contained in a matrix which encodes the different class memberships for each object. Here, a binary encoding is used where the Y column index signifies the class index. An object m belonging to class i is stored in row m where column i contains the value 1, and all other columns for this row are zero. The PLSR algorithm computes latent variables which are directed along the maximum X–Y covariance (22). In prediction, the class memberships are determined from which Y column index corresponds to the largest value. By default, DPLSR will force a new spectrum to be a member of one of the k classes. This is undesirable for cases where a
Two methods are used here to validate the classification models generated by DPLSR. The first method is based on using an independent validation data set to determine the optimal number of latent variables in DPLSR. In general, this is regarded the best method for validation of empirical models and is possible here as hyperspectral images contain a large number of spectra. The validation procedure proceeds as follows: • In each hyperspectral image, different and nonoverlapping regions of the same object type are defined as belonging to “calibration” and “validation” data. All spectra within these regions are subsequently collected into the calibration and validation data sets, respectively. • DPLSR classification models created on the basis of the calibration data are then applied to the set of validation spectra for different number of PLS factors (factors are in range a = 1,2,..,20) and used to determine the optimal number of PLSR components needed. The classification error used here is defined to be the percentage of wrong classifications and not the root mean square error of prediction (RMSP) which is common to use in regression problems. The second validation method employed is based on visual inspection of predicted chemical images using the optimal DPLSR model from the steps described above. This method is very effective for evaluating the predictive ability of chemometric models producing chemical images from hyperspectral images because a priori knowledge of the spatial distribution of the target class is often available. This is particularly relevant for the current study as we knew a priori which parts were related to bone and which were not on the calibration and test set objects described above. Set-up, Preprocessing and Analysis All prediction models were constructed using MATLAB 7.1 (The MathWorks, Inc. 3 Apple Hill Drive, Natick MA 017602098, USA) running on both Windows XP and Linux operating systems (Ubuntu 10.04.4 LTS, Canonical Ltd., London, UK). The Windows platform is used for the direct control and operation of the PryJector instrument. All off-line data analyses are performed on Linux workstations using inhouse chemometric software. Spectra from the hyperspectral images are numerically differentiated to first order using a 3rd order polynomial Savitzky– Golay filter with a 15-point window (26) and normalized to unit sum before being subjected to data analysis. For numerical differentiation, the savgol routine in the PLS_Toolbox 6.2 from Eigenvector Research Inc. (3905 West Eaglerock Drive Wenatchee, WA, USA) was used.
ALSBERG AND ROSVOLD
.
RAPID LOCALISATION OF BONE FRAGMENTS
477
Results Near-infrared light is defined to be in the region 730– 2300 nm and has been demonstrated to be useful for quantitative determination of different biological components such as total protein (27), albumin (28), and cholesterol (29). These predictions were performed in the so-called combination region (2000– 2500 nm) of the infrared part of the electromagnetic spectrum. The overtone region (800–1900), however, which is also used for the PryJector instrument, is more problematic due to the presence of strong absorption bands from water which have two dominating peaks around 1440 and 1930 nm. As water is considered a major interferent for this wavelength region, it was initially attempted to carry out the classification of bone fragments from nonbone fragments in dry conditions only. Each bone fragment was carefully dried in a food dryer, and samples of nonbone fragments were also kept in dry conditions. However, the resulting DPLSR models (not shown) from these experiments indicated that water was most likely still being used as the selective factor in discriminating between bone and nonbone fragments. For instance, we observed that pixels in the hyperspectral image of pure water droplets were wrongly predicted to belong to the bone class. These results suggest that the drying of the bone fragments was not sufficient, and thus, water directly interfered with the initial chemometric modeling. A significant amount of water is usually bound to the matrix of fresh bone where some of this is relatively freely bound and will evaporate at fairly low temperatures (below 105°C), while the rest is tightly bound structural water (30). Exposed bones will naturally dehydrate by weathering, but the more tightly bound water only evaporates by prolonged heating at higher temperatures (31). It is therefore likely that bound water was originally being used in the first model. On basis of this result, it was decided to design experiments such that water was being excluded from being used to discriminate between bone and nonbone samples. The approach taken was to record the bone and nonbone samples under both wet and dry conditions. Wet conditions were ensured by spraying water on the objects. Any successful DPLSR model from such a calibration data set would thus most likely not be using free water contribution to enable discrimination. To test whether such an approach would work, a DPLSR analysis was performed on a single hyperspectral image (size 320 9 400 9 148 elements) of a set of bones, stones, and wood fragments under both dry and wet conditions. Pixels in the hyperspectral image belonging to bone and nonbone were localized manually and divided into calibration and validation data matrices. The calibration and validation data matrices contained 10964 and 10963 spectra, respectively, and thus constitute
