Open Access
Add to BibSonomy
Abstract
The calibration and sampling of the multispectral Light Detection and Ranging (lidar) intensity is still challenging because the data acquisition has to be optimized for simultaneous 3D measurement, and the intensity retrieval methods need to be fast to enable real-time detection. We have studied the spectral measurement of low reflectance targets with an 8-channel hyperspectral lidar with improved waveform sampling and sensitivity, which now allow the detection of spectral differences even at low reflectance values. Our initial analysis resulted in a classification accuracy greater than 80%, which indicates that the multispectral lidar is able to detect the small differences in target spectral properties when reflectance at two or more channels is compared at the same time.
© 2020 Optical Society of America under the terms of the OSA Open Access Publishing Agreement
1. Introduction
Multi-wavelength terrestrial Laser Scanning (TLS) has been studied actively during the last decade. The recent developments build on earlier work on active hyperspectral sensing and radiometric calibration of monochromatic TLS ([1,2] and refs. therein). The development of the first multi-wavelength lidar instruments providing multispectral 3D point clouds started in 2012 [3–6]. They are still among the few research instruments available so far, and there is no commercial availability. The dual- or multi-wavelength laser operation is based on either the use of two laser sources operating at different wavelengths [5,6] or a supercontinuum laser [3,4]. These instruments have shown the added value in terms of target classification in 3D space: for example, to differentiate leaves from woody material in trees for estimating photosynthetic capacity, and retrieving spatial distributions of parameters such as leaf chlorophyll or moisture content [4,7]. Multi- or hyperspectral point clouds also enable mapping of spectral properties in 3D over large targets, such as forest stands [8]. At the same time, new challenges have emerged, such as dealing with vast amounts of data and improving the radiometric calibration, especially in far ranges and for low reflectance targets [8,9]. In the case of multiple lasers, also the alignment and matching of the returns need to be considered [7,8].
There are also a number of active hyperspectral lidar instruments that have so far focused mainly on spectral measurement, and not 3D lidar point clouds. This is because either the range measurement has been inaccurate [10] or pointwise data have only been produced [11,12]. While these experiments have provided important information on the sampling and calibration of the backscattered laser pulse and spectral identification of targets, it is also clear that the potential of 3D hyperspectral laser scanning is not yet fully utilized (cf. [7]).
Most of the multispectral lidar experiments so far have focused on vegetation, which has high reflectance in the near-infrared (NIR) compared to the visible [4–8,11,12]. The strong spectral differences are utilized in vegetation spectral indices, because their measurement is feasible even in the presence of instrument noise or inaccuracy [9]. The applications have included, e.g., the retrieval of nitrogen content in leaves [11] or analyzing waveforms for reference targets with 99% reflectivity [13]. Rock samples with distinguishable intensity differences were studied from multi-wavelength laser backscattering in [14]. The signal to noise ratio has been observed to decrease with a decreasing target reflectance [15] The need to extend the utilization of multispectral TLS from vegetation into built environment (for, e.g., moisture or mold detection in buildings), where low reflectance targets such as rocks or bricks are common, creates an increasing demand for improved sensitivity and signal-to-noise ratio (S/N). This is needed to be able to analyze better the low intensity laser returns and improve calibration, which is still a challenge for multi-wavelength TLS (see also [8]).
This study complements the preliminary results on lidar waveform sampling of low reflectance targets presented in [16]. The added value of this extended paper is in presenting further results and a proof-of-concept level demonstration of the advantage of multi-wavelength lidar detection over traditional monochromatic TLS.
2. Data and methods
2.1 The FGI HSL
The first prototype of the Finnish Geospatial Research Institute Hyperspectral Lidar (FGI-HSL) was published in 2012. More details on the first in instrument are given in [4], but the most important features with respect to the detection of low reflectance targets were the 1 GHz digitizing rate together with 1 ns laser pulses. These features affected strongly the detection capability, and hence targets with mostly high reflectance, such as vegetation, were studied. The 1 GHz sampling rate is widely applied in terrestrial laser scanners together with pulse widths of 1–5 ns (e.g., [8,11,13]). Short pulses are preferred to avoid compromising the spatial resolution, which is important for accurate range retrieval.
The improved and field operable FGI-HSL was introduced in 2018 [9]. Several modifications were made to improve both range and spectral accuracy: a new digitizer was implemented to increase the sampling rate from 1 GHz to 5 GHz, and the Avalanche Photodiode (APD) detector array configuration was readjusted to extend the spectral range towards blue wavelengths (closer to 450 nm). The silicon detector still restricts the spectral range into 450–1000 nm, digitized in eight channels, with a channel width of 50 nm. The optical design was optimized for minimal loss of signal, and the detection electronics were improved to reduce the noise level. A new compact supercontinuum laser (200 mW, 5–20 kHz) was utilized. The optical setup allows to choose different wavelength channels by adjusting the position of the APD with respect to the spectral dispersion. The current configuration of channels is presented in Fig. 1. The spectral calibration was carried out with Oriel Cornerstone 260 monochromator instrument. The differences in laser power between channels were normalized by distance calibration, which is done by means of taking measurements of a 99% Spectralon (Labsphere Inc.) reference panel at a range of distances from 2.5 m to 6 m. These improvements allowed the improved signal-to-noise ratio and hence improved the measurement of low reflectance target. More detailed description of the improved features and technical details of the FGI-HSL are presented in [9]. Figure 2 illustrates the construction of the new lidar and the challenging measurement conditions in a rock mine tunnel.
Fig. 1. The spectral channels of the FGI-HSL. Spectral channels overlap with each other by a varying amount due to wavelength dependent sensitivity of the APD and consequently varying signal-to-noise ratio across the measuring range. This also impacts the channel width, as low S/N, especially at the blue end, makes low intensity signal detection challenging.
Fig. 2. Sample caption (Ref. [4], Fig. 2). Left: the structure of the FGI-HSL: The supercontinuum laser is seen in the center of the image, and the mirror and rotators on top of it. The spectrograph is placed on the front edge of the box, with the measurement PC on the right. Right: the FGI-HSL in a mine tunnel, with a 99% Spectralon reference plate in the tunnel wall.
2.2 Low reflectance targets
The aim to improve the FGI-HSL for low reflectance targets was to develop methods to distinguish ore (i.e., the valuable mineral) from gangue (the worthless rock material mixed with the ore) in mineral rocks and underground mine tunnels. This turned out to be a challenge, as the FGI-HSL was mainly used for measuring vegetation and other bright targets with clear spectral differences and indices utilizing this feature [17]. The spectral differences for rock samples turned out to be small and difficult to sample in the first place, which called for improved sampling and resolution.
Although this study focuses on spectral information, it is important to study the lidar intensity retrieval from the operational point of view. Therefore, the data analyzed in this paper was retrieved from scanned rock point clouds instead of a pointwise measurement. Furthermore, a pointwise measurement may not provide a conclusive result, as the differences between ore and gangue only become visible when clustering a bulk of lidar points and observing the differences between these clusters. This is quite common in lidar-based classification studies, such as those for tree species classification [18].
The targets studied in this paper were ferrochrome ore and serpentine and metaperidotite gangue samples in Outokumpu Kemi Mine, Finland. Different rock piles known to contain ferrochrome ore (mixed with gangue) or gangue only, respectively, were scanned with the FGI-HSL with the scanner at about 5-meter distance, where the size of the laser footprint was 3 mm on average. A 99% Spectralon reference panel was used as a calibration standard, and scanned at the same distance as the target. We measured three different sample rock piles located in the mine tunnel, with different type of rock in each: one of these contained ore containing rock, while two of them contained mainly gangue. A typical surface structure of the rock material can be seen in Fig. 2, but the rock piles also contained smaller grains and sand in between different size rocks. Point clouds for the scanned piles were calibrated and processed using Matlab software. The scanning area was adjusted to span over the entire target in most cases, so that there was mostly no need to cut samples from the point cloud.
The intensity and range calibration procedures have been described in more detail in [9]. The backscattered reflectance is retrieved from the measured waveforms by fitting a Gaussian curve in the echo pulse. The measurement error (standard deviation) was about 2%, calculated as an average of 100 samples for a Spectralon 99% panel. The time-of-flight (ToF) distance estimation was carried out using median value of the eight channels, and a millimeter-scale accuracy was achieved [9].
The demonstration presented here is based on clustering the intensities of the lidar points in two and three dimensional plots. The spectral intensities were calibrated similarly to our previous FGI-HSL experiments using the 99% Spectralon reference panel (cf. [4,9]). As the rock piles do not exhibit any distinguishable structural differences between ore and gangue, we used lidar intensities in different channels instead. The range measurement, however, is crucial to minimize the differences caused by distance, and for future lidar-based ore maps, which are possible to produce simultaneously with lidar-based ore-gangue classification.
2.3 Classification accuracy
In addition to the 2D and 3D plots, we also carried out a Support Vector Machine (SVM) based classification for selected rock point clouds to find out if the ore/gangue classification from the point clusters is reliable. We tested binary-classification for each binary combination of the three rock types (ore/gangue and gangue/gangue). The two point clouds of the tested rocks were combined and divided between training and validation by randomly sampling every tenth point and using these points solely for validation. We also implemented multiclass classification, in which we used all three rock types by combining all the point clouds together. Similarly, every tenth point was used solely for validation of the multiclass model. For the SVM training, we utilized libSVM with radial basis function (RBF) kernels. Although further statistical methods should be developed to automatize and improve the classification, this initial analysis was carried out to demonstrate the potential of the FGI-HSL in producing the differences between rock types.
3. Results and discussion
Figure 3 shows a point cloud of an ore containing rock pile, together with a 99% Spectralon target placed on the pile for calibration purposes.
Fig. 3. HSL point cloud of a rock pile scanned in the mine, showing the large difference between rock and Spectralon 99% (green circle) intensities.
3.1 Intensity comparison for ore and gangue samples
Scatterplots for ore and a gangue samples are presented in Figs. 4–5. Backscattered reflectance values are clustered and compared at two different wavelengths in each plot to bring out the differences in reflectance at different wavelengths. Mostly, points of ore samples are clustered around higher intensities than gangue ones, even in cases where the overall trends in intensity appear similar to both targets. In some wavelengths, the ore sample is showing slightly different trends in intensity (e.g., serpentine ore at 612 nm vs. 713 nm in Fig. 5). This trend is even more pronounced in Fig. 6, where three dimensional plots at three different wavelengths are compared. The point clusters begin to separate even close to the origin, i.e., the zero reflectance values. Using a single wavelength, only one feature, i.e., the intensity, would be available, which would not lead into conclusive results. This can be seen in, e.g., Fig. 4 for 713 and 822 nm, where we observe similar trends, even though the points are clustered at different intensities. The trends become more distinct by choosing another wavelength.
Fig. 4. Scatterplots (backscattered reflectance) of metaperidotite gangue (blue) and ferrochrome ore (orange) at different HSL channels (713 and 822 nm). The intensity values for the entire pile are included. The plots on the left show a larger intensity range and also include the points for the Spectralon measured separately for ore (orange points near 1.0 intensity) and gangue (blue points near 1.0). The right-side plots show the rock intensities only near 5m range to minimize the possible error sources caused by inaccuracy in range calibration.
Fig. 5. Scatterplots (backscattered reflectance) of serpentine gangue (blue) and ferrochrome ore (orange). Intensities are plotted at about 5-cm range segment near 5m to minimize possible errors caused by inaccuracy in distance calibration.
Fig. 6. Scatterplots (backscattered reflectance) of serpentine gangue (blue) and ferrochrome ore (orange) compared at three different HSL channels representing different wavelengths. The data range is the same 5-cm segment as in earlier figures.
We have also observed in our previous studies on rocks and sands that there is a slight decreasing trend (5% – 6% in visible and 6% – 8% in NIR from 0° and 40°) in intensity with increasing angle of incidence to the target [17]. When the surface roughness of the sample is smaller than the laser footprint, the incidence angle effect is likely to average over the target, especially as our current sampling included the entire rock pile, where the orientation of the rocks was random. Hence the distribution of incidence angles is likely to smooth out most effects on intensity. In any case, the incidence angle distribution is similar for both ore and gangue piles. Similar averaging can be considered for tree canopies and deciduous shoots [19].
Typically, the backscattered pulse from a low reflectance target can have an intensity close to the noise level. Accurate retrieval of pulse characteristics of such pulses is challenging, especially in the case of rock samples, for which the spectral differences between different rock types can be very small. The low reflectance pulses from low signal returns have been previously observed close to noise level, especially from far ranges [8]. The differences visible in the low intensity plots shown in Figs. 4–5 indicate that it is possible to compare intensities lower than 10% retrieved with the improved FGI-HSL. This shows the potential of the FGI-HSL in detecting a presence of ore mixed in gangue, at least when a quantitative ore content is not necessary.
Figures 4–6 also point out that he differences between rock types are visible from a large bulk of points. As the point clusters have significant overlap, it is obvious that the rock identification is not trivial from a single, pointwise measurement. Extended statistical analysis is called for to establish a reliable detection method. This requires tests with all different combinations of wavelengths to find a combination that produces the most reliable results. A similar point clustering based approach was employed in [18] for tree species classification, where tree structural attributes retrieved from terrestrial laser scanner point cloud data were used as classification features in different combinations.
There are still a number of challenges, such as the large dynamic measurement range, especially when there is large variation in distances and objects located both in the near and far fields. This may complicate the radiometric and distance calibration. To minimize the effect of these uncertainties in this study, we included points within about 5 cm distance (range segment from 5.23m to 5.28m) into the comparison. Differences in intensity trends were observed even at this small range difference. For further optimization for low reflectance targets, we are also improving and optimizing the pulse sampling algorithms with the new 5 GHz sampling rate.
3.2 Study of classification accuracy
The resulting prediction accuracy obtained with the SVM classification is depicted in Table 1. For the binary classification, we achieved a prediction accuracy of about 80% for single point measurement. For multiclass classification, prediction accuracy drops to about 68%. This decrease was expected as the differences between the three rock types was very small, and due to measurement noise, there is overlap in large portion of the combined point clouds. It is likely that further optimizations (such as outlier detection) to the classification algorithms could provide significant further improvements to the accuracy. The role of the distance calibration accuracy of the instrument increases when classifiers are trained with data collected at different measuring range than the evaluation data. Even small errors in distance calibration can introduce erroneous positive identification results. Furthermore, while rock types could be identified by their overall low reflectance across all channels, this may lead to faulty classification in scenarios different than that at the Outokumpu mine. Training of classifiers should be guided to detect spectral differences, rather than overall reflectance values for classification. While we cannot rule out that the presented SVM classification does not rely on the overall reflectance, we have evaluated the intensity values across all channels for different rock types and observed channel-wise intensity difference, rather than overall intensity variations. As the channels of the FGI-HSL are adjustable, it will be possible to optimize the channels at which the differences are mostly visible. The first results suggest that the wavelengths between 500–700 nm provide different trends in intensity for these rocks, but this must be established with further studies.
Table 1. Support vector machine classification for metaperidotite rock (MPRD), rock with ferrochrome ore, and serpentine rock (SP).
4. Conclusion
We have studied the detection of spectral differences in low reflectance targets measured with the FGI hyperspectral lidar (FGI-HSL) with an improved pulse sampling and extended dynamic range sufficient for low intensity measurement. We show that these differences are visible in comparisons of intensity point clusters at some 80% binary classification accuracy.
This study confirms the first results in [16] which suggested that the FGI-HSL could produce spectral differences for low reflectance targets that are sufficient for identifying rock types or to distinguish ore from gangue in mineral samples. The results also point out the added value from multiple lidar channels: the small differences become visible only when observed at multiple wavelengths simultaneously. The fact that the intensity trends were similar at some wavelengths indicates that the reliable detection of these differences would not be possible using one or two wavelengths only, but should rather be based on point clustering in multiple different wavelengths. The inaccuracy of a simple threshold-based classification has also been suggested in previous studies [e.g., 8]. Further study is needed to develop these initial findings into a classification method, for example, to find the wavelength combinations and statistical classification methods providing the best classification accuracy for each mineral type.
Along with the identification algorithms, there is still much to be done to improve the calibration and reflectance measurement of hyperspectral TLS instruments, for example, in terms of laser stability and effects from measurement geometry. The role of range calibration errors will also have to be studied in more detail.
Funding
Tekes (1515/31/2016, 2175/31/2019).
Acknowledgments
Portions of this work were presented at the Optical Sensors and Sensing Congress in 2019, Multispectral terrestrial lidar: improving active spectral sensing of low reflectance targets [15]. The authors want to thank Outokumpu Kemi Mine for providing the targets and facilities for the experiment, as well as Aki Ruotsalainen at Cybercube for consultation.
Disclosures
The authors declare no conflicts of interest.
References
1. A. Manninen, T. Kääriäinen, T. Parviainen, S. Buchter, M. Heiliö, and T. Laurila, “Long distance active hyperspectral sensing using high-power near-infrared supercontinuum light source,” Opt. Express 22(6), 7172 (2014). [CrossRef]
2. 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 (2016). [CrossRef]
3. M. A. Powers and C. C. Davis, “Spectral LADAR: active range-resolved three-dimensional imaging spectroscopy,” Appl. Opt. 51(10), 1468 (2012). [CrossRef]
4. T. Hakala, J. Suomalainen, S. Kaasalainen, and Y. Chen, “Full waveform hyperspectral LiDAR for terrestrial laser scanning,” Opt. Express 20(7), 7119 (2012). [CrossRef]
5. E. S. Douglas, A. Strahler, J. Martel, T. Cook, C. Mendillo, R. Marshall, S. Chakrabarti, C. Schaaf, C. Woodcock, Z. Li, X. Yang, D. Culvenor, D. Jupp, G. Newnham, and J. Lovell, “DWEL: A Dual-Wavelength Echidna Lidar for ground-based forest scanning,” in 2012 IEEE International Geoscience and Remote Sensing Symposium (IEEE, 2012).
6. R. Gaulton, F. M. Danson, F. A. Ramirez, and O. Gunawan, “The potential of dual-wavelength laser scanning for estimating vegetation moisture content,” Remote Sens. Environ. 132, 32–39 (2013). [CrossRef]
7. Z. Li, M. Schaefer, A. Strahler, C. Schaaf, and D. Jupp, “On the utilization of novel spectral laser scanning for three-dimensional classification of vegetation elements,” Interface Focus 8(2), 20170039 (2018). [CrossRef]
8. F. M. Danson, F. Sasse, and L. A. Schofield, “Spectral and spatial information from a novel dual-wavelength full-waveform terrestrial laser scanner for forest ecology,” Interface Focus 8(2), 20170049 (2018). [CrossRef]
9. T. Malkamäki, S. Kaasalainen, and J. Ilinca, “Portable hyperspectral lidar utilizing 5GHz multichannel full waveform digitization,” Opt. Express 27(8), A468–A480 (2019). [CrossRef]
10. 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]
11. L. Du, W. Gong, S. Shi, J. Yang, J. Sun, B. Zhu, and S. Song, “Estimation of rice leaf nitrogen contents based on hyperspectral LIDAR,” ITC J. 44, 136–143 (2016). [CrossRef]
12. S. Hancock, R. Gaulton, and F. M. Danson, “Angular reflectance of leaves with a dual-wavelength terrestrial lidar and its implications for leaf-bark separation and leaf moisture estimation,” IEEE Trans. Geosci. Electron. 55(6), 3084–3090 (2017). [CrossRef]
13. S. Song, B. Wang, W. Gong, Z. Chen, X. Lin, J. Sun, and S. Shi, “A new waveform decomposition method for multispectral LiDAR,” ISPRS J. Photogramm. Remote Sens. 149, 40–49 (2019). [CrossRef]
14. Y. Chen, C. Jiang, J. Hyyppa, 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]
15. G. A. Howe, K. Hewawasam, E. S. Douglas, J. Martel, Z. Li, A. Strahler, C. Schaaf, T. A. Cook, and S. Chakrabarti, “Capabilities and performance of dual-wavelength Echidna®lidar,” J. Appl. Remote Sens 9(1), 095979 (2015). [CrossRef]
16. S. Kaasalainen and T. Malkamäki, “Multispectral terrestrial LiDAR: improving active spectral sensing of low reflectance targets,” in Optical Sensors and Sensing Congress (ES, FTS, HISE, Sensors), OSA Technical Digest (Optical Society of America, 2019), paper HW6B.2.
17. S. Kaasalainen, O. Nevalainen, T. Hakala, and K. Anttila, “Incidence Angle Dependency of Leaf Vegetation Indices from Hyperspectral Lidar Measurements,” Photogramm Fernerkun. 2016(2), 75–84 (2016). [CrossRef]
18. M. Åkerblom, P. Raumonen, R. Mäkipää, and M. Kaasalainen, “Automatic tree species recognition with quantitative structure models,” Remote Sens. Environ. 191, 1–12 (2017). [CrossRef]
19. S. Kaasalainen, M. Åkerblom, O. Nevalainen, T. Hakala, and M. Kaasalainen, “Uncertainty in multispectral lidar signals caused by incidence angle effects,” Interface Focus 8(2), 20170033 (2018). [CrossRef]
