Open Access
Add to BibSonomy
Abstract
Multispectral LiDAR enables joint observations of the 3D geometry and material properties of natural targets by combining ToF-based distance measurements with remote spectroscopy. Established multispectral LiDAR solutions provide mm-level range resolution and reflectance estimates of the target material over some tens of spectral channels. We propose a novel multispectral LiDAR approach based on an ultra-broadband frequency comb that enables enhanced remote spectroscopy by resolving relative delays in addition to reflectance. The spectrally-resolved delay and power measurements are transformed into distance and reflectance spectra by differential observations to a common reference object and adequate system calibration. These distance and reflectance spectra encode material information related to the surface and sub-surface composition and small-scale geometry. We develop the proposed comb-based multispectral LiDAR on an implementation covering the spectral range between 580 nm and 900 nm on 2 different spectral configurations with 7 and 33 channels of different spectral width. The performance assessment of the implemented system demonstrates a distance measurement precision better than 0.1 mm on most channels. Table-top probing results on five material specimens show that both the distance and the reflectance spectra alone enable discrimination of material specimens, while the novel distance signature particularly complements reflectance and increases classification accuracy when the material surface exhibits significant reflectance inhomogeneity. Material classification results using a support vector machine with radial basis function kernel demonstrate the potential of this approach for enhanced material classification by combining both signature dimensions.
© 2022 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement
1. Introduction
The increasing interest in perception and digitization of the environment relies on suitable methods for remote data acquisition. Multispectral LiDAR [1] is a promising solution to simultaneously access geometrical and material information by combining LiDAR and active remote spectroscopy. By probing the surface of natural targets with different wavelengths, it provides not only geometrical information but also spectrally-resolved reflectance estimates of the target material in the form of reflectance spectra. Such spectroscopic reflectance data has been successfully applied to assist segmentation of 3D point clouds [2,3], extract material parameters [4,5], and enable LiDAR-based material classification and object identification [6,7].
Approaches to multispectral LiDAR implement spectroscopic measurements based on different source configurations. Using multiple lasers with different optical wavelengths [8] is the most straightforward solution in terms of source simplicity and cost, but largely prevents application scalability by fixing the available spectral channels by design while introducing alignment complexity. More flexible spectral selection can be achieved by using tunable [9] or supercontinuum lasers [10,11]. While tunable lasers best exploit optical power, supercontinuum sources can typically cover larger spectral ranges and are not restricted to time-multiplexing spectroscopy, thus enabling simultaneous measurements at multiple wavelengths. Most supercontinuum lasers applied to multispectral LiDAR are derived from incoherently broadened pulsed lasers, which restricts the ranging principle to direct delay-based ToF approaches [10,12,13] with little room for pulse optimization, limiting the achievable distance measurement precision to the mm- or cm-level.
We have previously proposed a novel approach for multispectral distance measurement that uses a coherent supercontinuum derived from a mode-locked femtosecond (fs) laser, preserving its original frequency-comb structure [14]. Potential for sub-mm level distance precision in five spectral channels was demonstrated and exploited to derive spectrally-resolved optical delays [15]. Since optical power reflects not only on the target surface but also backscatters from sub-surface layers after slight penetration into the body of the target [16], both the spectrally-resolved reflectances and optical delays encode material information related to the outside surface condition and the size and concentration of scatters in the sub-surface layers. As a result, these reflectance and distance spectra convey the spectral signatures of the target material and can be used as unique features for its identification and classification. The goal of this research is to explore the capability of this technology to provide sufficiently precise distance and reflectance spectra for enhanced material classification. In this paper, we show a significantly improved design of the comb-based multispectral LiDAR prototype in terms of measurement accuracy, repeatability, and spectral resolution. Also, for the first time, we demonstrate its capability in extracting material-dependent distance and reflectance spectra with five material specimens (wood board, plastic, foam, plaster, and cardboard), and analyze the target identification potential of the extracted spectra toward delay-augmented material classification.
The paper is structured as follows: The measurement principle used to obtain distance and reflectance spectra, the design of the comb-based multispectral LiDAR, and the data processing workflow are introduced in Section 2. In Section 3, we analyze the performance of the comb-based multispectral LiDAR when acquiring distance and reflectance spectra over short (2 min) and long (11 h) terms. Section 4. shows and discusses the distance and reflectance spectra of the five material specimens (wood board, plastic, foam, plaster, and cardboard), along with material classification results using different spectral signature dimensions. Finally, the main conclusions and outlook of the work are summarized and commented on in Section 5.
2. Methods
2.1 Measurement principle
The proposed comb-based multispectral LiDAR deploys an ultra-broadband optical frequency comb. By feeding a mode-locked fs-laser with spectral bandwidth of up to a few tens of nm into a photonic crystal fiber, the comb spectrum coherently broadens to a supercontinuum that can span several hundred nm while preserving the line structure and the spacing defined by the repetition rate $f_r$ of the seed laser. As shown in Fig. 1, spectral channels ($C_1$ to $C_n$) are selected from the supercontinuum. This selection can be realized by either optical filtering or using a spectroscopic approach. Illuminating a photodetector (PD) directly with the backscattered light from targets in spectral channel $C_i$ centered at wavelength $\lambda _i$ generates a radio frequency comb at the electrical output of the PD due to intermode beating [17]. The radio frequency comb, whose span is determined by the PD bandwidth, preserves the constant frequency spacing $f_r$.
Fig. 1. Schematic diagram of measurement principle. Direct detection at a spectral channel $C_i$ with a center wavelength $\lambda _i$ of an ultra-broadband optical frequency comb with repetition rate $f_r$ produces a corresponding radio frequency comb on the photodetector (PD) output due to intermode beating.
The phase $\phi (mf_r,\lambda _i)$ of the electrical beat note at an integer multiple $m$ of $f_r$ can be converted into a distance estimate $d(\lambda _i)$ in the selected spectral channel $C_i$ as
2.2 Design of the prototype
We use a supercontinuum optical frequency comb spanning 550 nm to 1000 nm with 100 MHz repetition rate, produced from coherent broadening of a mode-locked fs-laser centered at 780 nm (C-fiber 780, MenloSystems) with 20 nm spectral bandwidth. As shown in Fig. 2, the supercontinuum output passes through one of the band-pass filters mounted in the filter wheels. A beam sampler splits the filtered beam into a reference and a probing path. In the probing path, the beam passes through a hole in the center of an off-axis parabolic mirror and illuminates the target surface with approximately perpendicular incidence. A target holder is fixed around 0.5 m away from the parabolic mirror, where the samples under test are mounted. The light backscattered from the target surface is subsequently collected and focused onto the probing detector APDpro by the parabolic mirror. In the reference path, the beam is directly focused onto the reference detector APDref by an achromatic lens. APDpro and APDref are both avalanche photodiodes (APD210, MenloSystems) with 1 GHz bandwidth. Forty band-pass filters mounted in eight filter wheels implement two independent spectral configurations (Fig. 3). Spectral configuration 1 consists of seven spectral channels with 40 nm bandwidth and central wavelengths between 600 nm and 900 nm. Spectral configuration 2 consists of thirty-three spectral channels with 10 nm bandwidth and central wavelengths spanning 580 nm to 900 nm. A reflectance standard (SG 3070, SphereOptics) with 60% reflectance, constant across the relevant spectral range, flips in and out of the probing path to provide radiometric referencing for the reflectance measurement and inline self-calibration for the distance measurement. The spectral channels are multiplexed sequentially for each complete measurement, while the backscattered light from the reflectance standard and the target are acquired successively for each spectral channel.
Fig. 2. Optical setup (SC-OFC: supercontinuum optical frequency comb, FWs: filter wheels, BS: beam sampler, M: mirror, FL: focusing lens, APDref: reference avalanche photodiode, PM: parabolic mirror, S: flip-in reflectance standard, T: target, APDpro: probing avalanche photodiode).
Fig. 3. Spectral configurations. The black line shows the optical spectral density of the supercontinuum comb. Spectral configurations 1 and 2 consist of channels with 40 nm and 10 nm spectral bandwidths (BWs), respectively. $C_{a,b}$ represents the channel with central wavelength at $a$ nm and $b$ nm spectral bandwidth.
The relatively short 0.5 m ranging distance and the time-multiplexed spectral selection using filter wheels are selected for this work prioritizing implementation simplicity considering the proof-of-concept focus of the research. Higher spectral resolution and spatial-multiplexed spectroscopy for simultaneous measurements on different spectral channels are however fundamentally feasible with the proposed approach, and can be achieved by optimizing spectroscopic devices and using parallel photodetection. Practically useful ranging distances without significant degradation of the precision can be achieved by (1) optimizing the spectral power distribution of the continuum laser source to allocate the optical power in all spectral channels more evenly or tuned to the spectrum of interest of a given application, (2) increasing the integration time, or (3) using a broader spectral bandwidth per channel considering the trade-off between measurement rate, spectral resolution, and ranging distance.
2.3 Data processing
The 900 MHz beat notes ($m$ = 9) of the radio frequency combs generated in APDpro and APDref are individually band-pass filtered and downmixed with a 1.001 GHz tone from a local oscillator. The lower frequency component at 101 MHz resulting from the mixing is simultaneously digitized on both probing and reference channels with a 2.5 GS/s sampling rate and 8-bit resolution. The phases and amplitudes of the acquired beat notes are then extracted from the digitized signals by synchronous I/Q demodulation, enabled by a common time-base for the laser and all electronics derived from a rubidium frequency standard.
In the distance measurement, the fixed reference path serves as phase reference but also helps reducing common-mode noise, e.g., the phase instability of the laser source. Differential drifts affecting both detectors independently, such as those caused by pointing variations or thermally-induced changes on the frequency response of the detectors and electronics, can thus introduce significant errors on the phase estimations. A self-reference compensation (SRC) approach is implemented to reduce these by subtracting the estimated distance of the reflectance standard from the estimated distance of the target. The estimated distance of the target at center wavelength $\lambda _i$ after self-reference compensation is
where $d_T$ and $d_S$ indicate the distance measurements according to Eq. (1) for target and reflectance standard, respectively.To obtain the target reflectance, the prototype exploits the same reflectance standard ($R_S = 60\%$). The estimated reflectance of the target at central wavelength $\lambda _i$ is computed as
Although the measurement geometries of the target and reflectance standard are fixed in this investigation, the ratio of coupling efficiencies $\eta _{\mathit {ratio}}$ is wavelength-dependent due to slightly different footprint sizes and positions of the incident beams for different spectral channels. Conversely, the functional relationship $G_S(A)$ can be safely considered wavelength-independent given that the spectral responsivity of APDpro is cancelled out in the ratio computation in Eq. (3). When operating in the linear region of the APD, $G_S(A)$ is linear and the estimated reflectance of the target can be calculated as $R_T(\lambda _i) = A_{\mathit {pro},T}(\lambda _i)\eta _{\mathit {ratio}}(\lambda _i)R_S/A_{\mathit {pro},S}(\lambda _i)$. However, given the uneven distribution of the optical spectral density of the supercontinuum across the investigated spectral range (see Fig. 3), the spectral channels with central wavelengths between 750 nm and 800 nm may observe much higher optical intensity than the rest and thus induce operation in the non-linear region of the APD. To calibrate the non-linear response of APDpro and the different ratios of coupling efficiencies for different spectral channels, $G_S(A)$ and $\eta _{\mathit {ratio}}(\lambda )$ were obtained experimentally (see Section 3.2) and are applied to all the reflectance estimations in the following material probing results in Section 4.
3. Performance analysis
Prior to investigating material-related information, the system is calibrated and the base precision of the distance and reflectance measurements in each spectral channel is analyzed. In this separate experiment for calibration and performance analysis, a reflectance standard (SG 3074, SphereOptics) with constant 25% reflectance over the investigated spectral range is used as the target instead of the five material specimens described in the following Section 4. The short- and long-term precision of the distance measurements and the long-term precision of the reflectance measurements for each spectral channel are assessed. The improvement in repeatability obtained by applying the self-reference compensation mechanism explained above, as well as the quality of the reflectance calibration, is also analyzed.
3.1 Distance measurement
The short-term precision of the distance measurements for each spectral channel is represented by the standard deviation of 100 distance measurements acquired within 2 min. Each individual measurement is derived by integrating observations over 1 ms. Since the optical intensity distribution of the supercontinuum is not stable over time, the short-term precision of distance measurement is assessed in 10 measurement cycles carried out within 13 hours, to account for the variability of the collected optical power, and thus signal-to-noise ratios (SNR), for each channel. All the spectral channels are selected consecutively on each measurement cycle, and 100 distance measurements are obtained in the selected spectral channel. Figure 4 shows the standard deviation obtained for each spectral channel on each measurement cycle. As shown in the figure, the short-term precision of the distance measurements is better than 20 $\mathrm {\mu m}$ for all channels in spectral configuration 1, and better than 100 $\mathrm {\mu m}$ for 27 of the 33 channels in spectral configuration 2.
Fig. 4. Short-term precision of the distance measurements for each spectral channel in spectral configurations 1 and 2. Each dot represents the standard deviation of a given channel computed from 100 distance measurements over 2 min using 1 ms integration time per measurement. The standard deviations for each of the 10 measurement cycles are illustrated in different colors.
The additive noise in the photodetection and amplification stages dominates the short-term precision of distance measurements, defined by the SNR which in turn depends on the collected optical power. The spectral channels with lower optical intensities in spectral configuration 2 (see Fig. 3), such as $C_{630,10}$ and $C_{640,10}$, therefore exhibit significantly worse short-term precision and introduce higher instrumental noise on the distance spectrum of the target material. This can be improved by increasing the optical intensity in these spectral channels or the integration time per individual measurement. To illustrate and quantify the latter, Fig. 5 shows the dependence of the short-term precision on integration time for two channels in spectral configuration 2. $C_{640,10}$ has lower optical intensity than $C_{760,10}$, which is reflected in the associated precision. The standard deviation in $C_{640,10}$ continuously drops as integration time is increased from 1 ms to 100 ms, the slope indicating a dominant background of white noise within the assessed time scale. The improvement in the standard deviation for $C_{760,10}$ flattens for integration times above 1 ms, likely indicating that the precision for longer integration times is no longer limited by noise but by more correlated processes arising from e.g. slow fluctuations of the laser power, refractivity, or mechanical changes. Achieving sub-10 $\mathrm {\mu m}$ short-term precision in $C_{640,10}$ requires integrating over 70.2 ms, while the same precision is obtained in $C_{760,10}$ in less than 1 ms due to the higher optical intensity of this channel. Increasing integration time is key to achieve adequate short-term precision on the distance measurements, especially for the spectral channels with lower optical intensity resulting from unevenly distributed optical power of the laser source, high-resolution spectral configurations, and longer ranging distances.
Fig. 5. Dependence of short-term precision with integration time in channels $C_{640,10}$ and $C_{760,10}$ (see Fig. 3) with different optical intensity.
The worse long-term repeatability of the distance measurement compared to short-term precision of the distance measurement is caused by slowly-varying drifts likely dominated by pointing variations and changes in the frequency response of the APDs and analog electronic chain, whose impact is mitigated by the self-reference differential measurements on the multiplexed reflectance standard. 30 distance measurements are carried out for each spectral channel over 11 h, and the standard deviation of which indicates the long-term distance measurement precision of the system. The results before and after applying the self-reference compensation are shown in Fig. 6 as black and red lines, respectively. The self-reference compensation brings the long-term precision to a similar level as the short-term precision, indicating that the impact of the slowly-varying drifts is successfully reduced to values smaller than the noise floor.
Fig. 6. Long-term precision of the distance measurements for each spectral channel in spectral configurations 1 and 2. Black and red lines represent the long-term precision before and after self-reference compensation, respectively. The gray shaded area depicts the worst case of the short-term precision as shown in Fig. 4 for easier comparison.
Although increasing integration time can further improve the precision on most channels, it reduces the achievable measurement rates for real-time distance and reflectance extraction. This trade off between probing speed and measurement precision should therefore be taken into account depending on the application needs. Nevertheless, using a relatively short integration time of 1 ms, the proposed comb-based multispectral LiDAR achieves sub-mm long-term distance precision for all forty spectral channels. Spectral configuration 1 displays significantly better results with long-term precision better than 60 $\mathrm {\mu m}$ for all 7 spectral channels, while 28 of the 33 channels in spectral configuration 2 yield long-term precision better than 100 $\mathrm {\mu m}$.
3.2 Reflectance measurement
Adequate reflectance estimation requires calibration of the non-linear response of APDpro and the different coupling efficiencies introduced by the different measurement geometries of the reflectance standard and the target. Since both the target and reflectance standard used in this performance analysis have known constant reflectances ($R_T$ = 25% and $R_S$ = 60%) over the investigated spectral range, the total reflected power for both can be obtained given a known incident optical power. The electrical amplitudes generated in APDpro by the light backscattered from the target and the reflectance standard are individually recorded for different incident optical power, where the latter is regulated using a variable achromatic neutral density filter and measured with a calibrated power meter on the target surface. Figure 7(a) depicts the recorded electrical amplitudes and the calculated reflected optical power from the target and the reflectance standard in spectral channel $C_{800,40}$, which contains the highest optical intensity and is thus the most prone to induce non-linear operation of the APD. The corresponding relation functions $P_{r,S} = G_S(A)$ and $P_{r,T} = G_T(A)$ are polynomial fitted to the recorded amplitudes and reflected optical powers. According to Eq. (3), the wavelength-dependent ratio of coupling efficiency $\eta _\mathit {ratio}(\lambda )$ can be experimentally obtained with known $R_T$, $R_S$, and $G_S(A)$.
Fig. 7. Calibration of the reflectance estimation in spectral channel $C_{800,40}$. (a) Relations between the electrical amplitude $A$ extracted from APDpro and the reflected optical power $P_r$ for the measurement geometries of the target ($T$) and the reflectance standard ($S$). (b) Estimated reflectances of the target (nominal $R_T$ = 25%) without and with reflectance calibration at different incident optical powers $P_{in}$.
Figure 7(b) shows reflectance estimations of the target in $C_{800,40}$ at different incident optical powers. The black line represents the estimated reflectance without calibrating the non-linear response of the APD and the different coupling efficiencies between the measurement geometries of the target and the reflectance standard. The red line illustrates the reflectance estimation after calibration with Eq. (3). The different coupling efficiencies introduce a constant overestimation of the target reflectance. The non-linear response of the APD, on the other hand, increases the reflectance overestimation for higher incident optical power. After calibration, the reflectance estimation remains stably close to the 25% nominal value of the target used in the performance analysis measurements, independently of the incident optical power. In addition to enabling accurate reflectance estimates for the channels with higher optical intensity, the non-linearity calibration also reduces the variability on all channels due to power instability of the supercontinuum source, thus improving the long-term precision of the reflectance estimations (see Fig. 8). After the reflectance calibration, the long-term precision of the reflectance measurements is better than 1% for all forty spectral channels over 11 h.
Fig. 8. Long-term precision of the reflectance measurements for each spectral channel in spectral configurations 1 and 2. Black and red lines represent the standard deviation of 30 reflectance measurements over 11 h without and with calibration, respectively.
4. Material probing
To demonstrate the capability of the proposed comb-based multispectral LiDAR in extracting reflectance and distance spectra of different target materials and the potential of the latter to augment material probing, five ordinary material specimens were tested (see Fig. 9). The system sensitivity and the variability induced by inhomogeneity of the material surfaces in both distance and reflectance spectra were assessed from independent measurements on different surface positions. Each measurement is integrated over 1 ms and processed using the self-reference compensation and calibration method explained above for the distance and reflectance estimations, respectively. The measurement geometry is also equivalent to the one used in the performance analysis, with the specimens mounted on the target holder 0.5 m away from the parabolic mirror and approximately normal incidence. Since the repeatability of the specimen position when mounted on the target holder cannot be guaranteed at the sub-mm level, reposition errors cannot be decoupled from actual material signatures in the absolute distance spectra. Misleading information induced by these errors is avoided by using only relative distance spectra, derived by subtracting the mean distance of all spectral channels in each measurement. The potential of exploiting the absolute distance spectra, which may also include relevant material information, will be investigated in future works.
Fig. 9. Five investigated material specimens. Detail pictures of the five material specimens: (a) wood board, (b) plastic, (c) foam, (d) plaster, and (e) cardboard. (f) The reflectance of the five material specimens measured with a spectrometer (CCS175/M, Thorlabs) and a stabilized tungsten-halogen light source (SLS201L, Thorlabs) on normal incidence at a single surface position.
4.1 Material signatures at a single surface position
Figure 10 shows the relative distance spectra of the five material specimens at an arbitrarily selected single surface position. The underlying measurements were carried out 100 times. The solid lines depict the respective average value of the 100 measurements per specimen and spectral channel. The error bars illustrate the maximum and minimum of the individual 100 measurements per specimen and spectral channel. The relative distance spectra of the five materials exhibit clearly distinctive profiles that can be exploited for classification, with specimen-specific trends differing by several hundred $\mathrm {\mu m}$. The dispersion of each spectrum in spectral configuration 1 is significantly smaller than the average distance between the signatures, showing very good separability already upon visual inspection. The lower optical power per channel and thus lower SNR in spectral configuration 2 induce much higher dispersion in some spectral regions, whereas in others the differences between specimen-specific trends are still clearly visible. The results show how better spectral resolution is generally associated with higher noise levels.
Fig. 10. Relative distance spectra of the five material specimens at a single surface position. The solid line and the error bar represent the average and the maximum-minimum values of the relative distances obtained from 100 measurements.
The reflectance spectra of the five material specimens obtained from the same measurements described above are shown in Fig. 11. In both spectral configurations, the reflectance spectra of the five materials are well separated. Due to the high reflectance measurement precision, the dispersion is on the sub-1% level in many spectral channels and thus not visible in the figure.
Fig. 11. Reflectance spectra of the five material specimens at a single surface position. The solid line and the error bar represent the average and the maximum-minimum values of the reflectances obtained from 100 measurements.
4.2 Material signatures at different surface positions
The material signatures shown in Section 4.1 are obtained from an arbitrarily selected single surface position per specimen and thus include only the impact of system sensitivity. To assess the variability induced by spatial inhomogeneity of the specimens, ten randomly selected surface positions were probed for every material by manually repositioning the specimens. One measurement was taken at each surface position. As expected, spatial inhomogeneities affect the spectra. Increased variability of the measured spectra is observed in all cases. Figure 12 shows the relative distance spectra for 10 surface positions on each material, where an excess in dispersion for foam and plaster with respect to the other three materials can be seen. This is likely caused by the considerably larger surface roughness of those specimens (see Fig. 9(c) and (d)). Consequently, the relative distance spectra of plaster and cardboard overlap. The differentiation between them is difficult based only on the relative distance spectra. The separation of the other material specimens, however, is still adequate upon visual inspection despite larger variations. Increased dispersion is also visible in the reflectance spectra illustrated in Fig. 13 but the mean values are still consistent with the reflectances given in Fig. 9(f) measured by a commercial spectrometer. All the five material specimens exhibit good separability in the reflectance spectra, except for the differentiation between wood board and cardboard on more than half of the spectral channels, likely caused by the visible texture on the surface of the wood board (see Fig. 9(a)) inducing larger reflectance inhomogeneity.
Fig. 12. Relative distance spectra of the five material specimens at 10 different surface positions. The solid line and the error bar represent the average and the maximum-minimum values of the relative distances obtained from 10 measurements.
Fig. 13. Reflectance spectra of the five material specimens at 10 different surface positions. The solid line and the error bar represent the average and the maximum-minimum values of the reflectances obtained from 10 measurements.
All the evaluated material specimens herein were probed under the same measurement geometry (ranging distance and incidence angle). The mounting of targets and the alignment of the current system guarantees the deviation of incidence angle smaller than 6$^{\circ }$ from normal incidence, which introduces a maximal 0.55% reflectance deviation on Lambertian targets. However, it is negligible compared to the 1% reflectance precision dominated by the noise floor. Several works have addressed the impact of incidence angle on reflectance estimation for monochromatic and multispectral LiDARs [18–20], although no concrete solution for correcting the reflectance spectrum has been proposed yet to the best knowledge of the authors. We have also initiated research on geometrical compensation for the distance spectrum with promising first results. The continuation of these investigations towards repeatable compensation of the impact of the measurement geometry on the material signatures is essential for the practical applicability of the proposed approach to multispectral laser scanning.
4.3 Material classification of specimens
To quantify the separability potential of the relative distance and reflectance spectra for the investigated material specimens beyond visual inspection, the acquired signatures are applied in a baseline classifier composed of a support vector machine [21,22] with a radial basis function (RBF) kernel. The spectra measured at ten different surface positions for each specimen are randomly split into seven spectra for training and three spectra for testing. The random data split and classification procedure is implemented ten thousand times, and the generated confusion matrices given in Fig. 14 illustrate the averaged classification results when using relative distance and reflectance spectra both separately and jointly. Figure 14(a) and (d) show that cardboard and plaster are partly miss-classified in both spectral configurations when only using relative distance spectra. In Fig. 14(b) and (e), a miss-classification between cardboard and wood board is obtained in both spectral configurations when exploiting only the reflectance spectra. These results support the observations from Fig. 12 and Fig. 13 regarding average distance and dispersion of the different spectra. The complementarity between both types of signatures with respect to miss-classifications suggests that the combination of relative distance and reflectance spectra can yield enhanced material classification performance. This is confirmed in the results shown Fig. 14(c) and (f), where the classification using both signature dimensions yields no miss-classifications.
Fig. 14. Confusion matrices of material classification among the five material specimens using different spectral dimensions (relative distance spectrum, reflectance spectrum, and the combination of relative distance and reflectance spectra) in two spectral configurations.
The material probing and classification analysis in this work refers to only 5 material specimens to provide proof of concept results. An extended measurement campaign to obtain a significantly larger material dataset would be highly beneficial to understand the relationship and sensitivity of each signature dimension with material characteristics, investigate application-driven spectral optimization solutions, and generally assess the practical potential of the approach on different application cases.
5. Conclusion and outlook
We have developed and demonstrated a comb-based multispectral LiDAR prototype providing spectrally-resolved distance and reflectance estimates on the wavelengths between 580 nm and 900 nm. The underlying technology allows measurements at a nearly arbitrarily large number of spectral channels. Driven by signal quality considerations and ease of implementation, we chose a solution with a fixed number of spectral channels realized by using filter wheels and time-multiplexing for the research herein. The performance assessment of the implemented experimental set-up enabled identifying and proposing compensation and calibration mechanisms for relevant instrumental systematics related to long-term phase drifts and photodetection non-linearities. The calibrated set-up yielded distance measurement precision over 11 h better than 60 $\mathrm {\mu m}$ for 7 spectral channels with 40 nm spectral bandwidth and better than 100 $\mathrm {\mu m}$ for 28 spectral channels with 10 nm spectral bandwidth. The high-precision spectrally-resolved distance measurements enable observing material- and surface- dependent relative distance spectra. The precision of the estimated reflectances was better than 1% on all channels, providing high-quality reflectance spectra. These relative distance and reflectance spectra encode repeatable relative optical delays and reflectance induced by the small-scale surface geometry and the material composition of surface and sub-surface layers. Material probing results on five material specimens show how the novel distance spectrum complements the traditional reflectance spectrum for material classification. Preliminary classification results using a baseline implementation also show enhancement potential in material classification by combing the two spectral signature dimensions.
To better understand the potential and material information encoded in the the distance spectral signatures and advance the proposed technology into practical applications, analyzing on the impact and compensation of varying measurement geometries (namely incidence angle and ranging distance) and establishing an extended material database are key follow-up investigations. These suggested future works will support the proposed comb-based multispectral LiDAR approach as a promising basis towards more comprehensive LiDAR-based environment digitization and perception.
Funding
Schweizerischer Nationalfonds zur Förderung der Wissenschaftlichen Forschung (200021_184988).
Acknowledgments
The authors would like to thank Valens Frangez for providing the plaster specimen and Jemil Avers Butt for the scientific discussions related to material classification. Portions of this work were presented at the Conference on Lasers and Electro-Optics in 2022, paper SF2F.5.
Disclosures
The authors declare no conflicts of interest.
Data availability
Data underlying the results presented in this paper are not publicly available at this time but may be obtained from the authors upon reasonable request.
References
1. B. Riveiro and R. Lindenbergh, Laser scanning: an emerging technology in structural engineering, vol. 14 (CRC, 2019).
2. B. Chen, S. Shi, J. Sun, W. Gong, J. Yang, L. Du, K. Guo, B. Wang, and B. Chen, “Hyperspectral lidar point cloud segmentation based on geometric and spectral information,” Opt. Express 27(17), 24043–24059 (2019). [CrossRef]
3. M. A. Powers and C. C. Davis, “Spectral ladar: active range-resolved three-dimensional imaging spectroscopy,” Appl. Opt. 51(10), 1468–1478 (2012). [CrossRef]
4. C. Jiang, Y. Chen, H. Wu, W. Li, H. Zhou, Y. Bo, H. Shao, S. Song, E. Puttonen, and J. Hyyppä, “Study of a high spectral resolution hyperspectral lidar in vegetation red edge parameters extraction,” Remote Sens. 11(17), 2007 (2019). [CrossRef]
5. J. Sun, S. Shi, W. Gong, J. Yang, L. Du, S. Song, B. Chen, and Z. Zhang, “Evaluation of hyperspectral lidar for monitoring rice leaf nitrogen by comparison with multispectral lidar and passive spectrometer,” Sci. Rep. 7(1), 1–9 (2017). [CrossRef]
6. B. Chen, S. Shi, W. Gong, J. Sun, B. Chen, L. Du, J. Yang, K. Guo, and X. Zhao, “True-color three-dimensional imaging and target classification based on hyperspectral lidar,” Remote Sens. 11(13), 1541 (2019). [CrossRef]
7. H. Shao, Y. Chen, Z. Yang, C. Jiang, W. Li, H. Wu, Z. Wen, S. Wang, E. Puttnon, and J. Hyyppä, “A 91-channel hyperspectral lidar for coal/rock classification,” IEEE Geosci. Remote Sensing Lett. 17(6), 1052–1056 (2020). [CrossRef]
8. G. Wei, S. Shalei, Z. Bo, S. Shuo, L. Faquan, and C. Xuewu, “Multi-wavelength canopy lidar for remote sensing of vegetation: Design and system performance,” ISPRS J. Photogramm. Remote. Sens. 69, 1–9 (2012). [CrossRef]
9. I. H. Woodhouse, C. Nichol, P. Sinclair, J. Jack, F. Morsdorf, T. J. Malthus, and G. Patenaude, “A multispectral canopy lidar demonstrator project,” IEEE Geosci. Remote Sensing Lett. 8(5), 839–843 (2011). [CrossRef]
10. T. Malkamäki, S. Kaasalainen, and J. Ilinca, “Portable hyperspectral lidar utilizing 5 GHZ multichannel full waveform digitization,” Opt. Express 27(8), A468–A480 (2019). [CrossRef]
11. Y. Chen, W. Li, J. Hyyppä, N. Wang, C. Jiang, F. Meng, L. Tang, E. Puttonen, and C. Li, “A 10-nm spectral resolution hyperspectral lidar system based on an acousto-optic tunable filter,” Sensors 19(7), 1620 (2019). [CrossRef]
12. Y. Chen, C. Jiang, J. Hyyppä, S. Qiu, Z. Wang, M. Tian, W. Li, E. Puttonen, H. Zhou, Z. Feng, Y. Bo, and Z. Wen, “Feasibility study of ore classification using active hyperspectral lidar,” IEEE Geosci. Remote Sensing Lett. 15(11), 1785–1789 (2018). [CrossRef]
13. B. Wang, S. Song, W. Gong, S. Shi, B. Chen, J. Yang, L. Du, and J. Sun, “High-precision ranging based on multispectral full-waveform lidar,” The Int. Arch. Photogramm. Remote. Sens. Spatial Inf. Sci. 43, 547–551 (2020). [CrossRef]
14. D. Salido-Monzú and A. Wieser, “Simultaneous distance measurement at multiple wavelengths using the intermode beats from a femtosecond laser coherent supercontinuum,” Opt. Eng. 57(4), 044107 (2018). [CrossRef]
15. D. Salido-Monzú and A. Wieser, “An instrumental basis for multispectral lidar with spectrally-resolved distance measurements,” Int. Arch. Photogramm. Remote. Sens. Spatial Inf. Sci. Arch. XLII-2/W13, 1121–1126 (2019). [CrossRef]
16. J. R. Schott, Fundamentals of polarimetric remote sensing, vol. 81 (SPIE, 2009).
17. K. Minoshima and H. Matsumoto, “High-accuracy measurement of 240-m distance in an optical tunnel by use of a compact femtosecond laser,” Appl. Opt. 39(30), 5512–5517 (2000). [CrossRef]
18. S. Kaasalainen, A. Jaakkola, M. Kaasalainen, A. Krooks, and A. Kukko, “Analysis of incidence angle and distance effects on terrestrial laser scanner intensity: Search for correction methods,” Remote Sens. 3(10), 2207–2221 (2011). [CrossRef]
19. J. Yang, Y. Cheng, L. Du, W. Gong, S. Shi, J. Sun, and B. Chen, “Analyzing the effect of the incidence angle on chlorophyll fluorescence intensity based on laser-induced fluorescence lidar,” Opt. Express 27(9), 12541–12550 (2019). [CrossRef]
20. X. Qian, J. Yang, S. Shi, W. Gong, L. Du, B. Chen, and B. Chen, “Analyzing the effect of incident angle on echo intensity acquired by hyperspectral lidar based on the lambert-beckman model,” Opt. Express 29(7), 11055–11069 (2021). [CrossRef]
21. C. Cortes and V. Vapnik, “Support-vector networks,” Mach. Learning 20(3), 273–297 (1995). [CrossRef]
22. F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay, “Scikit-learn: Machine learning in Python,” J. Mach. Learn. Res. 12(8), 2825–2830 (2011).
