Spectra interpretation
In order to reveal the particular spectral characteristics of each wood specimen studied as well as the reliability of the applied identification techniques, it is important to record the LIBS spectra under the same experimental conditions. Thus, the pulse energy of the 30fs Ti:Sapphire laser beam (λ = 800nm) was kept constant (0.99mJ/pulse) for all the recorded LIBS spectra, the laser operating frequency was 10Hz, while the gate width of the CCD camera was 120ns. Ten LIBS spectra were recorded from each irradiated spot of the samples. The spectra presented in Fig. 1 are the sum of the recordings from 100 different points.
The resolution of the spectra depicted in Fig. 1 is relatively low (0.6nm), thus, the assignment of the observed spectral features is based on a narrower wavelength range spectra of higher resolution (0.04nm) recorded utilizing a different diffraction grating. The NIST database for LIBS spectral analysis [39] was used to identify all the spectral peaks corresponding to atomic transitions. Spectral features that correspond to Carbon (C) atomic transition at ~ 248 nm (and its second-order diffraction at ~ 496nm) have been recorded, while spectral lines that could be attributed to atomic iron (Fe) transitions have not been observed, despite being previously reported [40]. In Table 2 the assignment of the observed transitions is presented in detail.
Table 2
Assignment of the recorded spectral features
Element/diatomic molecule | Wavelength (nm) | Transition |
C | 247.86 | 1P°→1S (ΔJ = 1) |
Mg+ | 279.55 279.80 | 2P°→2S (ΔJ = 1) 2D→2P°(ΔJ = 0,1) |
Mg | 285.21 | 1P°→1S (ΔJ = 1) |
Ca+ | 393.37 396.85 | 2P°→2S (ΔJ = 1) 2P°→2S (ΔJ = 0) |
Ca | 422.81 430.25 443.50, 443.57 445.48, 445.59, 445.66 526.42, 526.56 | 1P°→1S (ΔJ = 1) 3P →3P° (ΔJ = 0) 3D →3P° (ΔJ = 0,1) 3D →3P° (ΔJ = 0, ± 1) 3D →3P° (ΔJ = 0, -1) |
K | 404.41 | 2P°→2S (ΔJ = 1) |
Sr+ | 407.77 421.55 | 2P°→2S (ΔJ = 1) 2P°→2S (ΔJ = 1) |
Ba+ | 455.40 | 2P°→2S (ΔJ = 1) |
Na | 589.00, 589.59 | 2P°→2S (ΔJ = 1,0) |
NH | 336.3 | A3πi→ Χ3Σ− (Δv = 0) |
CN | 358.4–359.0 385.03-388.32 415.2-421.6 | B2Σ+ → Χ2Σ+ (Δv = + 1) B2Σ+ → Χ2Σ+ (Δv = 0) B2Σ+ → Χ2Σ+ (Δv=-1) |
C2 | 467.82-473.66 512.87, 516.52 544.61-563.48 | d3Πg→ a3Πu (Δv = + 1) d3Πg→ a3Πu (Δv = 0) d3Πg→ a3Πu (Δv=-1) |
The spectral bands attributed to transitions of diatomic molecular species (C2, CN, NH) are worth noting because their formation has been a subject of different interpretation scenarios [41]–[43]. It is characteristic that CN transitions have also been observed in the LIBS spectra of substances that do not contain nitrogen, so their formation is attributed to chemical reactions of the atmospheric nitrogen with C atoms and/or C2 fragments ejected from the ablated sample [41]. Furthermore, a detailed analysis of these molecular bands can be exploited for studies of the isotope ratio 12C/ 13C determination [44] and the temperature of the specific diatomic molecules, thus deriving information about as well as the temperature of the specific diatomic molecules, thus deriving information about both the material under study and the plasma produced.
In the present work, they were utilized to study the plasma temperature and its dynamic evolution. Recently, the non-trivial dependence of the fs-LIBS spectra characteristics on the laser focus point with respect to the sample (Aluminum oxide, AlO) surface has been reported [45]. Hence, the laser beam focusing conditions on the wood specimens’ surface were kept scrupulously constant. The temperature was determined using the Boltzmann Equilibrium Spectrum Program [46]. By simulating the corresponding spectral peaks (Fig. 2) the temperature of the CN and C2 molecules were estimated and found to be as high as ∼8400K and ∼6400K, respectively. Our findings align with previously reported measurements [47].
For the investigation of the temperature decay the temporal width of the CCD gate was reduced to 30ns and the measurements was repeated by varying stepwise the time delay between the laser beam and the CCD gate.
The temperature decays of C2 and CN, for the case of wild plum, are presented in Fig. 3. Prima facie the temperature decay of C2 seems to be faster than that of CN and this observation valid for all the wood specimens studied. This differentiation in the temperature decay time could be thought of as an indication that some processes are taking place within the plume.
Therefore, the emission dynamics of all species participating in plume have been investigated. The CCD gate width was reduced to 15 ns, and LIBS spectra have been recorded for a delay time of up to 500 ns. The emission intensity is reduced as the delay time increases (Fig. 4). The spectral intensity reduction rate of the species contributed to LIBS spectra has been evaluated by fitting the data with the exponential function:
$$y=A\times \text{exp}\left(-\frac{x}{t}\right)+y\text{0}$$
(1)
The determined values lie in the range of ~ 25 up to ~ 110ns.
From the analysis, it is concluded that the emission intensity reduction rate of CN (103 ± 16 ns) and Na (85 ± 13 ns) are the slowest. On the contrary, the evaluated rate for the rest of the atoms (tc=24 ± 4 ns, tCa=68 ± 10 ns, tCa+=43 ± 6 ns, tMg=26 ± 4 ns, tMg+=23 ± 3 ns) and the C2 (31 ± 5 ns) fragment is much faster which means that the composition of plasma changes as time proceeds.
In order to explore the reason for this variation, LIBS spectra have been recorded in an argon (Ar) atmosphere (Fig. 5).
From Fig. 5, it is evident for the spectra recorded in an Ar atmosphere, that the continuum background increases, along with the intensity of the spectral lines corresponding to the transitions of all atoms except those of Na. At the same time, the intensity of the bands assigned to CN transitions is reduced. The increase in the intensity of the carbon (C), calcium (Ca and Ca+) and magnesium (Mg) peaks is conceivable because in the experiments performed in an Ar atmosphere, the oxidation processes (e.g. 2C + O2→ 2CO) leading to the formation of CO, CaO and MgO[48], [49] are practically eliminated and therefore an intensity reduction of the corresponding atomic spectral peaks is avoided.
In addition, we notice that the increased signal of C is accompanied by a relative increase in C2. This could imply that the C2 population increases via the gas phase C + C→ C2 reaction [42]. Nevertheless, it is more likely that C2 is produced primarily by the irradiated sample rather than in the plume because in all the wood samples organic polymers with aromatic rings i.e., strong C = C double bonds, exist. This approach is also compatible with the fact that the intensity of the C2 bands gains in amplitude in the Ar atmosphere, while those assigned to CN exhibit an apparent decrease. The latter is understandable by taking in account that the reaction leading to CN production in expense of C2 (C2 + N2→ 2CN) can be efficient only in ambient atmosphere. Consequently, the activation of this reaction in the plume also accounts for the determined faster decay of C2 as compared to that of CN.
Another reason which could cause a reduction in the CN signal is the reactions involving atomic carbon (C + N2→ CN + N and C + N→ CN) which are unproductive in an Ar atmosphere experiments [50].
Thus, the reduced presence of CN in LIBS spectra in an Ar atmosphere is reasonable. However, even in an Ar atmosphere, CN as well as NH are clearly observed in the fs-LIBS spectra. This shows that these species can also be molecular fragments released by the dissociation of wood components and that they are not only reaction products in the gas phase. After all, it is known that in the wood composition the percentage participation of N ranges up to ~ 0.5%, while that of H ~ 6.5%.
As noted above, the only spectral atomic lines whose intensity decreases (like those corresponding to CN with which they show a common decay) are those corresponding to the characteristic double-line system of sodium (Na). The decrease in the intensity of the Na lines is understood as through a series of reactions, which are only possible in ambient atmosphere. The population of the 2P1/2, 3/2 states of Na is increased, and deexcitation to the ground state is accompanied by the emission of photons with wavelength equal to 589.0/589.6nm [51]. This process is described by the following reactions:
2Ο2 + hf → O3 + O
Na + O3 → NaO + O2
NaO + O → Na (2P1/2, 3/2) + O2
Na (2P1/2, 3/2) → Na(2S1/2) + hf.
Finally, regarding the increase of the spectral background, which is clearly apparent in the Swan transition of C2 (Fig. 6), it should be noted that these spectra have been recorded by keeping the same CCD gate delay and the temporal width of the gate for both the experiments in ambient and Ar atmosphere.
This observation has been attributed to the fact that for the same temperature gradient, the energy flow is lower in an Ar environment since [49]:
\({\rm E}=-\kappa \frac{dT}{dz}\) [2]
where κ is equal to 0.0163 J/K.m.s and 0.0241 J/K.m.s for Argon and ambient atmosphere, κair = respectively [52].
Equivalently, the presence of a stronger background in the LIBS spectra in an Ar atmosphere, recorded for a specific CCD gate delay, could be attributed to the reduced scattering cross section as compared to that in the ambient atmosphere which is consisted of molecules (N2, O2, etc). This affects the processes leading to the emission of bremsstrahlung radiation and contributes to the differentiation of the time evolution of the background in LIBS spectra.
Furthermore, by repeating the emission dynamics measurements in an Ar atmosphere (Fig. 7), it is found that the reduction rate of the spectral features intensity is similar (within the experimental error) for all the species participating in the plume. This supports the above-presented approach about the processes taking place in ambient atmosphere experiments. In other words, indicates that the composition of the plasma is preserved for longer in the experiments performed in an Ar atmosphere. A similar study on nitroimidazoles has shown that the decay time for C and CN is relatively larger in Ar atmosphere experiments, while for C2 there is no clear trend [53].
The agreement of the conclusions based on the relative variation of the intensity of the spectral features and those obtained from the analysis of the plasma dynamics is apparent. Thus, having analyzed the fs-LIBS data, the next step is to exploit the ability for accurate, sensitive, and fast wood identification by employing data and machine learning techniques.
Data analysis
An extensive analysis was conducted to classify the LIBS data observed from experiments in the ambient atmosphere. The dataset consisted of 1000 collected spectra, with 100 spectra per tree. Eight features (C, Mg+, Mg, CN, Ca+, Ca, C2, and Na) were carefully extracted from each spectrum (each spectrum corresponds to a vector). These features were selected based on their demonstrated relevance in characterizing the elemental composition of trees. Due to the high dimensionality of the dataset (1000 spectra X 8 feature/spectrum), PCA was firstly used as a well-known dimensionality reduction technique. The implementation of PCA in this study proved contributory in unraveling the underlying structure of the LIBS dataset, facilitating the exploration of key factors contributing to the classification of tree species.
Explained variance ratio index (EVR) was analyzed to understand the contribution of different numbers of principal components [54]. EVR represents the proportion of the total variance in the dataset explained by each principal component. After evaluation, it was found that the highest EVR was achieved by keeping three primary components, which captured approximately 99% of the overall information. Specifically, the first principal component explained about 72% of the variance, the second 20%, and the third 7%, respectively.
The contribution of each feature to the formation of each one of the three principal components is shown in Fig. 8. The x-axis represents the main feature produced after analysis, while the y-axis represents the % contribution values. Each feature is depicted as a bar on the x-axis, and the height of each bar corresponds to the magnitude of the feature's contribution to the principal component.
Based on the analysis of the contribution of features to the principal components, it is evident that CN, Ca+, and Ca play a significant role in the formation of the first and second principal components, while Na is crucial for the formation of the third one. On the other hand, elements such as C and C2 have minor contributions to the principal components. That mean that elements like C and C2 occur in similar quantities across all types of wood and do not contribute significantly to the separation and differentiation between the wood species.
A three-dimensional scatter plot was created to visualize the interrelationships among the data points obtained through PCA (Fig. 9a). This visualization provided a comprehensive understanding of the correlations within the dataset and to identify distinct clusters. Arrows are incorporated in the plot to illustrate the directions of the principal components, aiding in the visualization of how the variables or features of the data contribute to the overall variance and their relative relationships.
A detailed examination of the scatter plots revealed the presence of well-defined clusters within the dataset. This observation carries important implications for understanding the underlying structure and patterns in the data. Specifically, certain wood species such as beech, pine, and rosewood exhibit distinct clusters that are clearly separated from the other species. These clusters are non-overlapping, indicating a strong differentiation between all wood species. On the other hand, some wood species, such as acacia, wild plum, and chestnut, exhibit overlapping clusters. This overlap suggests a level of confusion or difficulty in differentiating these specific wood species based on their chemical compositions. The lack of clear separation in these cases highlights the complexity and potential challenges involved in accurately classifying these wood species using the available features.
Recognizing and analyzing these patterns makes it apparent that certain wood species exhibit more distinguishable characteristics, while others may share similarities, making their classification more challenging.
After PCA analysis for dimensionality reduction and visualization purposes of the data, the weighted k-Nearest Neighbors (k-NN) algorithm was applied for wood species classification.
In this experiment, the dataset was repeatedly split into two equal parts: 50% for training and 50% for testing, utilizing a random approach. As previously mentioned, the training set is indeed larger than the validation set. Nevertheless, our experimental results, obtained through various partitioning approaches, demonstrate that this division also produces remarkable results. Subsequently, the k-NN classifier algorithm with weighted distances was applied, with k taking values ranging from 2 to 10. It was observed that the algorithm yielded similar results across the range of k values examined. Consequently, the accuracy of the method was evaluated using k = 5 nearest neighbors.
Figure 10 displays the confusion matrix obtained from the application of the weighted k-NN algorithm for wood species classification. By examining Fig. 10, one can gain a comprehensive understanding of the algorithm's effectiveness in accurately classifying wood species. The diagonal elements of the matrix represent the correctly classified instances, while the off-diagonal elements indicate misclassifications. This visual representation allows for the assessment of the classification model's overall accuracy both the overall accuracy of the classification model and the specific classes that may pose challenges for accurate classification.
The results of the three-evaluation metrics (sensitivity, specificity, and accuracy) for each wood species are presented in Fig. 11. By computing these metrics, we can gain a more detailed understanding of the classification model's performance, including its ability to correctly classify positive instances (sensitivity), its ability to correctly classify negative instances (specificity), and the overall accuracy of the model in classifying wood species based on the provided features.
Sensitivity, representing the proportion of correctly identified positive instances, varied across the classes, with values ranging from 0.66 to 1.0. The k-NN classifier demonstrated high sensitivity for class 3 (Beech) and class 8 (Rose), correctly identifying all positive instances. Specificity, indicating the proportion of correctly identified negative instances, ranged from 0.96 to 1.0, suggesting a high accuracy in identifying negative instances in all classes. The overall accuracy of the classifier was 0.95, indicating that most of the predictions were correct.