Open Access
Add to BibSonomy
Abstract
Hyperspectral LiDAR enables non-contact mapping of the 3D surface geometry of an object along with its spectral reflectance signature and has proved to be effective for automated point cloud segmentation in various remote sensing applications. The established hyperspectral LiDAR methods offer a range precision of a few mm to a few cm for distances exceeding several meters. We propose a novel approach to hyperspectral LiDAR scanning based on a supercontinuum (SC) coherently broadened from a 780 nm frequency comb. It provides high precision distance measurements along with target reflectance over the 570–970 nm range of the SC output. The distance measurements are carried out by monitoring the differential phase delay of intermode beat notes generated by direct photodetection, while the backscattered light spectrum is acquired using a commercial CCD spectrometer with 0.16 nm resolution across the 400 nm bandwidth of the SC output. We demonstrate a measurement precision below 0.1 mm for a stand-off range up to 50 m on a diffuse target with around 89% reflectance. The measured relative accuracy as compared to a reference interferometer is on the order of 10−5 for distances up to 50 m. Initial results also indicate spectrum-based material classification within a 3D point cloud using a linear support vector machine. The results highlight the potential of this approach for joint high-precision laser scanning and automated material classification.
© 2023 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement
1. Introduction
Light detection and ranging (LiDAR) enables accurate three-dimensional (3D) digitization of our surroundings and is widely used for applications like remote sensing, surveying, autonomous driving, and deformation monitoring [1–3]. Conventional laser scanning (also referred to as LiDAR scanning) uses a monochromatic laser source to capture the 3D surface geometry ($x, y, z$) of a target and, after compensating for attenuation due to distance and angle-of-incidence also surface reflectance ($R_\lambda$) at the optical wavelength ($\lambda$). As compared to monochromatic LiDAR, combining the range information with extended spectral reflectance data over several (multispectral) or a continuous distribution of wavelengths (hyperspectral) has proved to be advantageous for finer point cloud segmentation, remote spectroscopy, and material classification [4,5].
Photogrammetric methods can be used to derive 3D point clouds with spectral information ($x, y, z, R_{\lambda _1, . . , \lambda _n}$) from images of conventional cameras and multi- or hyperspectral cameras. However, the co-registration of spectral and range data can be challenging in such imaging methods [6]. Besides, the photogrammetric methods typically rely on identifying radiometric surface features to establish point correspondences between images acquired from different viewpoints. Hyperspectral LiDAR (HSL) can overcome these challenges by simultaneously acquiring geometric and spectral information from the same measurement point and more independently from the surface appearance [7]. This facilitates accurate target identification by capturing the distinct spectral signatures of objects [8,9].
Dual-wavelength solutions using independent laser sources are already used commercially for airborne laser scanning [10]. Some experimental systems on dual or multiwavelength solutions for terrestrial laser scanning have also been presented in the literature. For instance, the Echidna LiDAR uses data from two 5 ns pulsed lasers at 1064 and 1548 nm to distinguish leaves from wood over a range of more than 70 m for characterizing forest canopy structures [11]. A multiwavelength approach using measurements at four different visible and near-infrared wavelengths combined with a separate range finder has also been shown to be advantageous for remote sensing of vegetation [12]. However, these approaches primarily rely on prior knowledge for selecting the optical wavelengths corresponding to a specific classification task e.g. selecting specific wavelength pairs for known vegetation indices. The broader range of available wavelengths provided by HSL allows for greater spectral flexibility to gather further insights into the target material properties.
The first prototype of an active full-waveform HSL was based on a SC incoherently broadened from a pulsed laser and provided spectral data for 16 channels within 450-1050 nm [13]. The authors demonstrated a range precision of 11.5 mm over distances of 20 m using direct time-of-flight (ToF) based ranging. Other techniques based on ToF measurements and advancements over the performance (measurement range, precision) of the initial prototype have also been reported in the literature [14–16]. Nevertheless, these early HSL methods use direct delay-based ToF approaches with limited scope for further pulse optimization and are limited to mm to cm range precision over a few tens of meters.
Recent advances in frequency combs (FC) technology have opened a possibility to achieve high precision for distance measurements even over long range [17]. Key properties of FC laser sources are their very high frequency stability ($\Delta f/f \approx 10^{-10}$ or better) and their extended coherence length (km-range) [18,19]. Since the first demonstration of using FC for distance measurements on cooperative targets (reflectors) [17], various methods have been implemented to extend their application towards non-cooperative targets (natural surfaces). FC-based frequency-modulated continuous wave (FMCW) LiDAR has been demonstrated over distances up to 10.5 m with sub-µm accuracy and measurement precision below 10 µm [20]. Similar precision at 3.5 m has been reported using dual-comb ranging [21]. An intermode beating (self-beating) based non-scanning experimental setup has been also proposed for profile measurements of rough surfaces with a relative accuracy of around 0.4 mm compared to a coordinate measuring machine [22]. While these techniques substantiate the possibility of achieving high precision ranging over several meters, their spectral flexibility is limited to the center wavelength of the FC used in the application.
In this paper, we propose a novel approach for HSL based on a supercontinuum (SC) obtained from coherent broadening of a 780 nm FC. Distance measurements are acquired using the intermode beats derived from direct photodetection of the SC. The spectrum of the backscattered light is captured over the 400 nm (570–970 nm) bandwidth of the SC. We assess the performance of the proposed approach experimentally under controlled conditions over a range of up to 50 m. The assessment includes precision and its dependence on target reflectance, distance, and data integration time, as well as accuracy in comparison to a laser Doppler interferometer (LDI). The achieved results demonstrate the potential of broadband intermode beating-based LiDAR to overcome the ranging performance limits of current HSL alternatives. Furthermore, we present initial results on SC spectrum-based material classification distinguishing different types of wood within the scan of a test object. This provides an outlook on combined high-precision laser scanning and simultaneous material probing.
The paper is organized as follows: the measurement principle, experimental setup, and reflectance properties of the targets used in our work are introduced in Section 2. The experimental results in terms of range precision, relative accuracy, and spectrum-based material classification are presented in Section 3. Section 4 summarizes the primary contribution of this work with directions for further research.
2. Methods
2.1 Experimental setup
We use a SC source spanning the spectral range of 570-970 nm with an available optical power of around 26 mW. An instantaneous spectrum of this SC as observed on a commercial Czerny-Turner CCD spectrometer (Thorlabs CCS 175) is shown in Fig. 1(a). The SC is generated through a photonic crystal fiber (PCF) (Menlo Systems SCG1500) that coherently broadens a 780 nm mode-locked femtosecond laser (Menlo Systems C-fiber 780 SYNC100) with a pulse repetition rate ($f_r$) of 100 MHz locked to a Rubidium (Rb) frequency standard (SRS FS725). Coherent broadening using the PCF ensures retaining the comb line spacing $f_r$ in the SC [23].
Fig. 1. (a) Instantaneous optical spectrum of the supercontinuum, (b) electrical power spectral density (PSD) of the intermode beat notes obtained from direct photodetection of the supercontinuum on a high-speed avalanche photodiode with a bandwidth of 1 GHz
When the SC illuminates a photodetector (PD), these equally separated optical comb lines at $f_r$ beat with the adjacent modes resulting in an electrical comb spectrum $mf_r$ (where $m=1, 2, 3,\ldots$) limited by the PD bandwidth [17]. Figure 1(b) shows the electrical power spectral density of the intermode beats obtained from an avalanche photodiode (APD) with a nominal bandwidth of 1 GHz. The electrical beat notes each separated by 100 MHz can be clearly observed up to the APD bandwidth in Fig. 1(b). The spurious beat notes (prominent beyond 1.2 GHz) are measurement artifacts of temporal aliasing caused by the analog-to-digital converter (ADC) sampling rate. The sampling rate of the ADC was set to 5 GHz.
Compared to established phase-based distance measurement techniques which use amplitude-modulated continuous-wave lasers, the intermode beat notes provide high-quality radio-frequency (RF) modulation signals without requiring an external modulator. The phase of the beat notes is proportional to the propagation delay. The distance measurement $(D)$ is estimated from the differential phase delay observed simultaneously on a local reference and probe path, which can be expressed as
A schematic of the experimental setup is shown in Fig. 2. The fiber-collimated SC output is expanded to a beam diameter of 9 mm using a reflective-type beam expander (RBE) to maintain collimation over a longer range while minimizing chromatic aberration. A beam sampler (BSF) picks off $\sim$10% of the SC output to establish a reference arm which is used to monitor $\phi$ref (see Eq. (1)) using APDref. The probe arm passes through the scanning unit which comprises an achromatic off-axis parabolic (OAP) mirror, the scanning mirror (SM) assembly, the probe photodetector (APDprobe), and a spectrometer. An OAP mirror with a 17.7 mm through-hole and 3" surface diameter is used to optimize the forward beam propagation without clipping and to efficiently collect the back-reflection from diffusely reflecting targets. The scanning mirror assembly consists of a 4" diameter flat deflection mirror placed on a programmable motor-controlled 2-axis gimbal, which enables accurate and repeatable beam steering for laser scanning. The current design of our experimental prototype can measure up to about 180$^\circ$ horizontally and 90$^\circ$ vertically. A detailed explanation of the features and calibration of the scanning unit is provided in Supplement 1.
Fig. 2. Schematic representation of the experimental setup (PCF: photonic crystal fiber; RBE: reflective beam expander; BSF: beam sampler; OAP: off-axis parabolic mirror; SM: Scanning mirror; FL: focusing lens; APD: avalanche photodiode; MMF: multi-mode fiber; BPF: band-pass filter; LO: local oscillator; CLK: clock; ADC: analog-to-digital converter; I/Q: inline-quadrature)
The back-reflection collected by the OAP mirror is focused onto the APDprobe for the distance measurements. The electrical intermode beats generated at the APDs (see Fig. 1(b)) are fed to an analog band-pass filter (BPF) to isolate the 1 GHz beat note (denoted as $f_M$). The 1 GHz signals from the reference and probe paths are downconverted to an intermediate frequency $f_{\rm IF} = 400\,\mathrm {kHz}$ upon mixing with the output of a local oscillator (LO) set to $f_M + f_{\rm IF}$. To further improve the signal-to-noise ratio (SNR) and reduce phase uncertainty, a second set of analog BPF centered at 400 kHz with 100 kHz BW is used. The $f_{\rm IF}$ signals are acquired simultaneously using a 2-channel ADC at a sampling rate of around 78 MS/s. The acquired time series is processed using a digital inline-quadrature (I/Q) phasemeter to estimate the phase difference corresponding to the propagation distance. Accurate timing references to the LO and ADC are also obtained from the Rb-frequency standard used for the mode-locked laser.
Due to practical space constraints in our initial design, the focused light from the OAP could not be coupled simultaneously to both the spectrometer and the $\mathrm {APD_{probe}}$ placed on the scanning module. For the experimental investigation herein, we thus replaced the $\mathrm {APD_{probe}}$ with the spectrometer to record the reflection spectrum of the target in a consecutive set of measurements. The target and the scanning module are not moved within these sets to ensure point-to-point correspondence between the range and spectral information (limited by the position accuracy of the scanning mirror assembly). This experimental choice is not a fundamental limitation of the approach; simultaneous acquisition is possible with a modified setup at the expense of optical power by splitting the backscattered light to the photodetector and spectrometer.
All the experiments are conducted under stable lab conditions at 20 $^\circ$C and 50% relative humidity. The barometric pressure cannot be controlled in the lab but remained within the range of 955 to 958 hPa over the entire duration of the experiments. Thus, the distance error caused by variations of the meteorological conditions and corresponding changes of the refractive index of air as well as potential temperature-induced deformations of the experimental setup can be neglected over the measurement range of 50 m.
2.2 Target properties
A picture of the two targets used in our experiments is shown in Fig. 3. The planar target (Fig. 3(a)) with visibly white, grey, and black cardboard is used to assess the ranging precision and accuracy performance of our experimental setup and to investigate the effect of surface reflectance on the measurement precision. The cardboard pieces are pasted on a 60 $\times$ 60 cm$^2$ aluminum plate for stability and are fixed to a computer-controlled motorized trolley that can move along the 50 m length of a comparator bench. The type of cardboard was chosen such as to be smooth and reflect diffusely over the entire SC spectrum. To assess the relative ranging accuracy, the measurements obtained from our experiments are compared with reference measurements from a 633 nm He-Ne laser Doppler interferometer (Agilent 5529a). The interferometer uses a corner-cube reflector fixed on the trolley to deliver high-accuracy reference measurements and provide position feedback to the motor-controlling software. The experimental results presented in Sec. 3.1 and Sec. 3.2 are obtained using this planar cardboard target.
Fig. 3. (a) Planar target with visibly white, grey, and black cardboard having different reflectances placed on a computer-controlled trolley for 5–50 m displacement tests on the comparator bench; a corner cube reflector is fitted to trolley for laser Doppler interferometer (LDI) measurements, (b) target made of three wood varieties: beech, balsa, and laminated low-density fiberboard (LDF)
Figure 3(b) shows the second target made of three varieties of wood that are used to demonstrate hyperspectral laser scanning and spectral signature-based material classification on a 3D-point cloud since the different materials exhibit different spectral signatures. A 40 cm $\times$ 55 cm white lacquer-coated flat low-density fiberboard (LDF) made from mixed wood is used as a background. Two rectangular pieces of 10 mm thick balsa are placed on the top and bottom part of the LDF, whereas the ETH letters are constructed using 25 mm thick beechwood pieces. Experimental results showing the acquired 3D point cloud of the wooden target along with spectrum-based material classification are presented in Sec. 3.3.
To estimate the reflectance ($R_\lambda$) of the cardboard and wooden specimens, an arbitrarily selected region on each of the samples was illuminated with a stabilized tungsten-halogen broadband light source (SLS201L, Thorlabs) at normal incidence through a contact probe and the reflected spectrum normalized to a diffusive 80% reflectance standard (Labsphere, SRT-80-050 Spectralon) was recorded using the spectrometer. Using a second set of reference measurements acquired on a 5% reflectance standard (Labsphere, SRT-05-050 Spectralon), the sample reflectance was estimated using two-point calibration. Figure 4(a, b) shows the reference reflectances of the cardboard and wooden target captured through the contact probe using the broadband light source. The average estimated $\overline {R}$ for the white, grey, and black cardboard is approximately 89%, 42%, and 7% respectively over the spectral range of our SC. $\overline {R}$ of the wood varieties is around 82%, 63%, and 40% for the LDF, balsa, and beech samples respectively.
Fig. 4. Reference reflectance signatures of the different targets: (a) visibly white, grey, and black cardboard surfaces, (b) laminated low-density fiberboard (LDF), balsa, and beech wood samples; (c) measured spectrum of the SC diffusely reflected from the wood specimens placed at a distance of 5 m and normalized to the LDF spectrum.
The spectrometer used in our experiments allows an integration time of 10 µs to 60 s. For the experiments herein, we chose an integration time of 90 ms as a trade-off between the SNR of the collected light on the spectrometer and the scanning speed. The scanning speed of HSL using a spectrometer for acquiring the spectral signatures is primarily limited by the necessary integration time for an adequate SNR. However, the overall scanning speed of our prototypical lab setup is currently constrained by the real-time data processing of the range measurements, i.e. around 0.8 s per point, as discussed later in Sec. 3.1.
Figure 4(c) shows the SC spectrum reflected from an arbitrarily selected region on the wooden samples and normalized to the spectrum of the background LDF. These measurements are collected using the scanning module for the target placed at a stand-off range of 5 m. The raw spectrum (shown in a lighter color tone) shows a higher measurement noise due to a lower SNR for measurements acquired over a longer distance. The peaks observed toward both extremes of the spectral distribution originate from the very low output power of the SC at these wavelengths and consequently higher deviations after normalizing with the background spectrum. The raw observations are processed with a moving-average window length of 10 nm to indicate the relative spectral differences between the different varieties of wood (shown in a darker color tone). These spectral variations exhibit similar trends as the reference spectrum in Fig. 4(b) which is exploited for their classification. The differences in the observed baselines may occur from a partly non-Lambertian reflection and from temporal variations of the SC power spectral distribution between the individual measurements [26]. These spectrum measurements using the SC do not reflect the absolute reflectance of the target specimens and would require extensive calibration as a function of distance, angle-of-incidence, surface roughness, etc. [27]. Nevertheless, preliminary investigations presented in Sec. 3.3 show spectrum-based classification of the wooden target using the relative differences in their acquired spectral signatures.
3. Experimental results
3.1 Range precision
We conducted the initial experiments on the planar target shown in Fig. 3(a) on different surface reflectances to assess the dependence of measurement precision on both reflectance ($R$) and distance ($D$) to the target. In a simplified approximation, the received signal power $P_s$ can be expressed as
where $\eta$ is the joint transmission and collection efficiency, $\rho _R$ is the target reflectance ($=0.01 R$) at normal incidence, $P_0$ is the output laser power, $\theta$ is the incidence angle of the laser beam on the target, and $A_s$ is the effective area of the scanning mirror [28]. This results in $P_s \propto \rho _R$, whereas $P_s \propto 1/D^2$. On the other hand, optical direct detection systems are typically dominated by the thermal noise of the PD under low optical power, being the standard mode of operation for long-range measurements on diffusive targets. Due to this constant noise contribution, the SNR of the acquired beat notes is expected to be directly proportional to the square of the optical power given the power-to-current conversion at the PD. The measurement uncertainty for I/Q-based phase measurements can be quantified by the standard deviation $\sigma _D \propto 1/\sqrt {SNR}$ [29] or as $\sigma _D \propto 1/P_s$. So, we expect the standard deviation of the distance measurements to be directly proportional to $D^2$ and indirectly proportional to the reflectance $\rho _R$, i.e., where $k(\tau _w)$ is a scale factor depending on the data integration time $\tau _w$ and $\sigma _0$ is the signal-independent measurement noise floor (including mechanical sources of uncertainty such as vibrations).Figure 5(a) shows the range precision empirically assessed using 100 measurements on an arbitrarily chosen fixed beam position on each of the three surfaces. The incidence angle of the laser beam is approximately normal to the white cardboard and $< 2.5^\circ$ on the grey and black cardboard surface. Each data point is obtained by integrating the I- and Q- samples over a quasi-randomly chosen time of $\tau _w'=13\,{\rm ms}$. The processing time for each measurement point is around 0.8 s on the standard PC used for processing the experimental data, herein. This is much longer than the acquisition duration. However, this is not a fundamental limitation for further development, because the processing time can be reduced using a faster computational unit e.g. using a GPU or FPGA. A measurement precision of around 3, 6 and 22 µm is achieved on the white, grey, and black surfaces, respectively, for a stand-off range of 5 m (see Fig. 5(a)). As expected the measurement precision is inversely proportional to $R$. However, the absolute precision values at a close-range (i.e. 5 m) are likely dominated by the underlying noise floor $\sigma _0$. $\sigma _0$ is empirically estimated to be around 2 µm.
Fig. 5. (a) Distance deviation of individual observations averaged over 13 ms on a static plate with three different reflectances, placed at 5 m; (b) precision as a function of the measurement distance ($D$) from 5 m to 50 m for the areas of different reflectance. The dotted lines represent the $\overline {k}D^2/\rho _R$ model shown in Eq. (3)
We carried out further experiments to evaluate the effect of $R$ and $D$ on the range precision. The cardboard plate was displaced from 5 m to 50 m at equidistant steps of 5 m. A continuous time series of distance variations are recorded within 56 s at each of these distances and processed offline. Each measurement series is again integrated over $\tau _w'$ (i.e., 13 ms), which corresponds to around 4300 measurement points in the time series. The data points in Fig. 5(b) show the empirical standard deviation of these measurements at different distances. The theoretical model in the form of Eq. (3) (shown as dashed lines) is also computed for each case. The average scale factor $\overline {k}(\tau _w')$ for the three surfaces is estimated to be around 0.027 µm/m$^2$. A maximum relative error of $6{\%}$ from this average model is observed for the black surface. The experimental results agree with the prior expectations with respect to reflectance and distance. A high measurement precision of 72 µm, 150 µm, and 985 µm at 50 m is demonstrated on the white, grey, and black surfaces respectively for an integration time of 13 ms.
We also analyzed the dependence of precision on the integration time. Using the acquired continuous time series of 56 s at 5 and 50 m, the standard deviations of the time series were processed within non-overlapping consecutive windows of duration $\tau _w (s)$. Within a certain range of window lengths, i.e. integration times, the measurement noise is dominated by white noise, as shown by the slope of $-(1/2)$ in the logarithmic plot of the empirical standard deviation versus $\tau _w$ (see Fig. 6). The slower precision improvement prominent after 20 ms on the white and grey surfaces visible only at 5 m suggest the presence of very slow signal drifts at Hz and sub-Hz rates that appear to be unrelated to the received signal noise distribution, and limit the reachable precision of the current implementation to the µm level. These may be introduced by mechanical vibrations of the measurement setup or electromagnetic interference at the high-frequency cables that are more apparent with increased signal strength. The slower precision improvement after 0.1 s for the black surface at 50 m appears to be dominated by quantization noise due to low SNR inherent to the challenging measurement conditions. Nevertheless, for longer averaging windows the measurement precision can improve significantly, e.g. to about 30, 60, and 400 µm for the white, grey, and black surfaces respectively for 50 m distance and integration time of 0.1 s.
Fig. 6. (a) Empirical standard deviation of distance as a function of data integration (averaging) time $\tau _w$ for different reflectances at (a) 5 m and (b) 50 m
The precision analysis is extended to include the impact of scanning. Figure 7(a) shows a point cloud of the cardboard surfaces when placed at a stand-off range of 5 m. The corresponding histogram showing the statistical spread of the distance deviation is shown in Fig. 7(b). A total of 1722 points (574 points each on cardboard sample) were acquired on the planar target in a continued raster scanning pattern, where the deviations in range (or depth) are visible mainly along the z-direction because this is roughly the scanning direction, and the target is parallel to the x-y plane. We again used an integration time of $\tau _w'$ per point.
Fig. 7. (a) Observed point cloud of the cardboard surfaces showing the distance deviations (along the z-axis, note different scale) for a stand-off range of 5 m, (b) histogram showing the statistical spread on the three surfaces
A standard deviation of 16, 23, and 29 µm is observed in the z-direction for points on the white, grey, and black surfaces. Similar to the observations on single-point precision, the scanning precision reduces for low-reflectance targets. However, the analysis is now also influenced by the non-planarity and surface roughness of the cardboard samples, as well as the remaining calibration errors of our scanning module. This explains the higher standard deviations compared to the precision of the single-point distance measurements (Fig. 5). However, we note that the apparent degradation of precision is large for the white surface for which the pure range noise was 3 µm before, while the degradation is small for the black surface for which the range noise was already 22 µm. These results suggest that the additional uncertainty caused by the surface and the scanning module is on the order of 10 to 20 µm, and the empirical standard deviations calculated from the point cloud represent the superposition of the pure range noise and these additional contributions.
3.2 Relative range accuracy
To also evaluate the measurement accuracy of the developed HSL prototype further experiments were conducted, where our measurements are compared to those obtained from the reference interferometer. These experiments were carried out in 3 cycles, each comprising a forward and reverse motion of the target from 5 to 6 m and from 49 to 50 m at steps of 1 cm. Measurements were acquired for one point on the white cardboard only for each of these target positions. This yields data of high precision, thus yielding high sensitivity of the analysis with respect to the identification of potential systematic effects (causing the accuracy to be lower than the precision). The results are shown in Fig. 8, where each data point represents an individual measurement averaged over $\tau _w'$. Distance residuals were calculated as the differences between the measurements ($D$) of the SC-based distance meter and those of the reference interferometer ($D_\mathrm {HeNe}$).
Fig. 8. Residuals between the measured and reference (interferometer) distances for (a) 5-6 m, (b) 49-50 m, before and after calibration with a sinusoidal fit
Figure 8 shows residuals within 100 µm for 5–6 m and 5 mm for 49–50 m. However, these large residuals are dominated by clearly systematic, sinusoidal error patterns with a period of around 15 cm. This agrees with the range ambiguity corresponding to the 1 GHz beat note used for our experiments. Such errors are well known as ‘cyclic deviations’ from conventional phase-based electro-optic distance measurement [30]. They primarily arise from electrical cross-talk from the RF oscillator or the mixing of spurious optical reflections of the measurement beam and represent one of the main sources of correctable systematics in such phase-based measurement systems. We apply a simple calibration-based correction by fitting a single sinusoidal function to the residuals. The residuals remaining after subtraction of this calibration function indicate a measurement accuracy (root mean square error) of 12 µm for 5 to 6 m and 0.4 mm for 49 to 50 m, with potential for further correction as suggested by the remaining systematics [31]. Nevertheless, a post-calibration accuracy on the order of $10^{-5}$ is achieved already with this first prototypical measurement system for distances up to 50 m.
3.3 Hyperspectral laser scanning
In this part, we present the geometric point cloud and spectral information obtained from scanning the wooden target specimen shown in Fig. 3(b). We also show the preliminary results of a purely spectrum-based material classification within the point cloud. The target specimen was scanned from a distance of 5 m. About 3200 data points were acquired in a continued raster scanning pattern, with 1.7 mrad vertical and horizontal angular resolution. We chose this configuration as a compromise between high spatial resolution for resolving the geometrical features and low resolution for quick data acquisition and low memory needs. The 3D point cloud displaying the target geometry is shown in Fig. 9(a), where the z-coordinate represents the distance deviations measured from the instrument and the x-y plane is approximately parallel to the background surface. The geometrical features are accurately captured in our scan showing the 25 mm thickness of the ETH letters made of beech and the 10 mm thickness of the balsa bars. The points not conforming to any of the wood samples (mixed pixels) are typical artifacts of laser scanning that occur due to the partial overlap of the beam with regions of the target specimen at different depths [1].
Fig. 9. 3D point cloud of the wooden target showing (a) geometrical (range) profile, and (b) material classification using spectral information. z-axis represents the range information for the target placed in the x-y plane
As mentioned earlier, we use a CCD-based spectrometer to capture the hyperspectral reflection spectrum from the target surface within the initial investigation, herein. The reflection spectrum was acquired with 0.16 nm resolution over the entire spectral range of the SC. Although the intensity of the beat note observations is independent of any changes in the ambient illumination conditions, the measurements on the spectrometer are affected by background illumination. The background spectrum observed on the spectrometer was corrected before scanning and ambient light conditions were kept stable during the scanning process. Background correction can be performed regularly for fluctuating environmental illumination if necessary. An active HSL that is unaffected by ambient light fluctuations may be achieved by replacing the CCD-spectrometer with an array of high-speed photodetectors corresponding to individual spectral bands or by sequentially probing over the entire spectral range using tunable optical filters at the cost of scanning speed [9].
We used the entire set of wavelengths (features) between 570–970 nm. The mixed pixels have been temporarily filtered out from the data for training the classifier model. Since the spectral properties are likely to be correlated for the neighboring areas on a single piece of wood, for a more independent evaluation we chose a test/train subset that corresponds to spatially different areas of the point cloud instead of using randomly split subsets. The data points on the lower balsa and left half (where $X<0$) of the beech and LDF are used as a training set while the remaining data is used for testing. The class weights are balanced based on the relative frequencies of the classes in the training data and are then fed to a linear support vector machine (SVM)-based classifier [32]. A 5-fold cross-validation (CV) is used to tune the hyperparameters and evaluate the general performance of the classifier. The classification accuracy is around 96% (CV accuracy is 0.96 with a standard deviation of 0.01). Figure 9(b) illustrates the classified point cloud showing the different material classes, i.e. beech, LDF, and balsa, on all the acquired points (including mixed pixels) demonstrating accurate classification of the wood varieties using their backscattered reflection spectrum. Figure 10 shows the workflow of the classification approach.
A comprehensive evaluation of the classification accuracy would require a large number of independent and diverse training datasets to test and evaluate on fully independent data. The results presented herein still highlight the potential feasibility of extending our high-precision supercontinuum-based hyperspectral LiDAR approach to scanning with point-cloud classification based on spectral signatures of the materials. The enhanced range precision together with hyperspectral information is also beneficial for accurately estimating the surface geometry and orientation of complex geometrical features, and for radiometric calibration of the acquired intensities [33].
4. Conclusion and outlook
We have developed and demonstrated a supercontinuum-based hyperspectral LiDAR prototype capable of high precision ranging over 50 m using phase delay observations from 1 GHz intermode beat notes while capturing spectral information of the diffusely reflecting surfaces between 590–970 nm. For measurements averaged over 13 ms, we obtained standard deviations of the distance measurements of around 3 µm at 5 m, and 72 µm at 50 m on cardboard surface with 89% average reflectance. These standard deviations were found to be directly proportional to distance squared and indirectly proportional to reflectance for distances larger than about 10 m, and dominated by a noise floor of a few µm for strongly reflecting targets at shorter distances. Additionally, we confirmed that the precision is approximately proportional to the square root of the integration time ($\tau _w$) over a large range of $\tau _w$, and can be improved by averaging the measurements over a longer duration. Compared to measurements from a reference interferometer, we achieved a relative accuracy on the order of $10^{-5}$ for stand-off distances up to 50 m after mitigating cyclic errors using a simple sinusoidal model. Preliminary results also show the potential of acquiring hyperspectral 3D point clouds of a target and using its back-reflection spectrum for material classification. Further work is underway to assess the long-term stability of the prototype, analyze the impact of the angle of incidence on targets having a more complex geometry, and investigate the benefit of high-precision measurements for efficient surface reconstruction. Improving the portability of the current prototype will lead to a promising technological basis for outdoor remote sensing and 3D digitization applications.
Funding
Schweizerischer Nationalfonds zur Förderung der Wissenschaftlichen Forschung (200021_184988).
Acknowledgments
The authors thank Robert Presl for preparing the target specimens used in this work. Tomislav Medić, Nando Metzger, and Jemil Avers Butt are acknowledged for scientific discussions on material classification.
Disclosures
The authors declare no conflicts of interest.
Data availability
Data underlying the results presented in this paper are publicly available through the ETH Zurich Research Collection.
Supplemental document
See Supplement 1 for supporting content.
References
1. B. Riveiro and R. Lindenbergh, Laser scanning: An emerging technology in structural engineering (1st ed.) (CRC, 2019).
2. M. Itzler, M. Entwistle, S. Wilton, I. Kudryashov, J. Kotelnikov, X. Jiang, B. Piccione, M. Owens, and S. Rangwala, “Geiger-mode lidar: From airborne platforms to driverless cars,” Imaging Appl. Opt. (2017).
3. J. Eitel, B. Höfle, L. Vierling, A. Abellán, G. Asner, J. Deems, C. Glennie, P. Joerg, A. LeWinter, T. Magney, G. Mandlburger, D. Morton, J. Müller, and K. Vierling, “Beyond 3-d: The new spectrum of lidar applications for earth and ecological sciences,” Remote. Sens. Environ. 186, 372–392 (2016). [CrossRef]
4. E. Puttonen, J. Soumalainen, T. Hakala, E. Räikkönen, H. Kaartinen, S. Kaasalainen, and P. Litkey, “Trees species classification from fused active hyperspectral reflectance and lidar measurements,” For. Ecol. Manage. 260(10), 1843–1852 (2010). [CrossRef]
5. J. Soumalainen, T. Hakala, H. Kaartinen, E. Räikkönen, and S. Kaasalainen, “Demonstration of a virtual active hyperspectral lidar in automated point cloud classification,” ISPRS J Photogramm. Remote. Sens. 66(5), 637–641 (2011). [CrossRef]
6. J. Mundt, D. Streuker, and N. Glenn, “Mapping sagebrush distribution using fusion of hyperspectral and lidar classifications,” Photogramm. Engg. and Remote Sensing 72(1), 47–54 (2006). [CrossRef]
7. S. Kaasalainen, T. Lindroos, and J. Hyyppä, “Toward hyperspectral lidar: Measurement of spectral backscatter intensity with a supercontinuum laser source,” IEEE Geosci. Remote Sensing Lett 4(2), 211–215 (2007). [CrossRef]
8. 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]
9. Y. Han, D. Salido-Monzú, and A. Wieser, “Comb-based multispectral lidar providing reflectance and distance spectra,” Opt. Express 30(23), 42362–42375 (2022). [CrossRef]
10. J. C. Fernandez-Diaz, W. E. Carter, C. Glennie, R. L. Shrestha, Z. Pan, N. Ekhtari, A. Singhania, D. Hauser, and M. Sartori, “Capability assessment and performance metrics for the titan multispectral mapping lidar,” Remote Sens. 8(11), 936 (2016). [CrossRef]
11. E. Douglas, J. Martel, Z. Li, G. Howe, K. Hewawasam, R. Marshall, C. Schaaf, T. Cook, G. Newnham, A. Strahler, and S. Chakrabarti, “Finding leaves in the forest: The dual-wavelength echidna lidar,” IEEE Geosci. Remote Sensing Lett. 12(4), 776–780 (2015). [CrossRef]
12. 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]
13. T. Hakala, J. Suomalainen, S. Kaasalainen, and Y. Chen, “Full waveform hyperspectral lidar for terrestrial laser scanning,” Opt. Express 20(7), 7119–7127 (2012). [CrossRef]
14. M. A. Powers and C. C. Davis, “Spectral ladar: active range-resolved three dimensional imaging spectroscopy,” Appl. Opt. 51(10), 1468–14678 (2012). [CrossRef]
15. 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]
16. 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]
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. A. Diddams, K. Vahala, and T. Udem, “Optical frequency combs: coherently uniting the electromagnetic spectrum,” Science 369(6501), aay3676 (2020). [CrossRef]
19. S. T. Cundiff and J. Ye, “Colloquium: Femtosecond optical frequency combs,” Rev. Mod. Phys. 75(1), 325–342 (2003). [CrossRef]
20. E. Baumann, F. Giorgetta, J. Deschênes, W. Swann, I. Coddington, and N. Newbury, “Comb-calibrated laser ranging for three-dimensional surface profiling with micrometer-level precision at a distance,” Opt. Express 22(21), 24914–24928 (2014). [CrossRef]
21. D. Hu, Z. Wu, H. Cao, Y. Shi, R. Li, H. Tian, Y. Song, and M. Hu, “Dual-comb absolute distance measurement of non-cooperative targets with a single free-running mode-locked fiber laser,” Opt. Commun. 482, 126566 (2021). [CrossRef]
22. T. Onoe, S. Takahashi, K. Takamasu, and H. Matsumoto, “Non-contact precision profile measurement to rough-surface objects with optical frequency combs,” Meas. Sci. Technol. 27(12), 124002 (2016). [CrossRef]
23. J. M. Dudley, G. Genty, and S. Coen, “Supercontinuum generation in photonic crystal fiber,” Rev. Mod. Phys. 78(4), 1135–1184 (2006). [CrossRef]
24. J. Butt, D. Salido-Monzú, and A. Wieser, “Phase ambiguity resolution and mixed pixel detection in edm with multiple modulation wavelengths,” ISPRS J. Photogramm. Remote. Sens. 198, 255–271 (2023). [CrossRef]
25. 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(04), 044107 (2018). [CrossRef]
26. P. Ray, D. Salido-Monzú, and A. Wieser, “High-precision intermode beating electro-optic distance measurement for mitigation of atmospheric delays,” J. Appl. Geod. 17, 93–101 (2023). [CrossRef]
27. K. Tan, X. Cheng, and X. Cheng, “Modeling hemispherical reflectance for natural surfaces based on terrestrial laser scanning backscattered intensity data,” Opt. Express 24(20), 22971–22988 (2016). [CrossRef]
28. M. Amann, T. Bosch, R. Myllyla, and M. Rioux, “Laser ranging: a critical review of usual techniques for distance measurement,” Opt. Eng. 40(1), 10–19 (2001). [CrossRef]
29. D. Salido-Monzú, F. J. Meca-Meca, E. Martín-Gorostiza, and J. L. Lázaro-Galilea, “Snr degradation in undersampled phase measurement systems,” Sensors 16(10), 1772 (2016). [CrossRef]
30. J. M. Rüeger, Electronic Distance Measurement - An Introduction, 3rd Ed (Springer, 1990).
31. J. Qiao and J. Butt, “Self-calibration of terrestrial laser scanner using a m3c2-based planar patch algorithm,” ISPRS J. Photogramm. Remote. Sens. 197, 335–345 (2023). [CrossRef]
32. 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. Learning Res. 12, 2825–2830 (2011).
33. P. Ray, D. Salido-Monzú, A. Wieser, and T. Medic, “Supercontinuum-based hyperspectral laser scanning: towards enhanced 3d surface reconstruction and its benefits for remote sensing,” Proc. SPIE 12621, 126210M (2023).
