Open Access
Add to BibSonomyAbstract
Hyperspectral light detection and ranging (HSL) is a newly developed active remote sensing technique. In this study, we firstly presented an improved hyperspectral full-waveform LiDAR system with 32 detection channels. Then, the quality of the data collected from two types of leaves by this system was evaluated using signal to noise ratio. Two different reflective factors that can describe the backscatter capability of detected targets were developed based on the HSL data. Hundreds of vegetation indices (VIs) were calculated through a full search for the possible combination of the reflective factors at near-infrared and visible wavelengths. Finally, the high-dimensional VIs (n = 998) were used to estimate three leaf biochemical contents using principle component regression (PCR) models with cross validation. Results showed that high correlations were found between leaf biochemical contents and the HSL-derived VIs at shorter visible wavelengths. The prediction of biochemical contents obtained satisfactory results with a root mean squared error of 0.45% for nitrogen content (R2 = 0.71), 1.41 mgg−1 for chlorophylla/b content (R2 = 0.83), and 0.38 mgg−1 for carotenoid content (R2 = 0.77), respectively. To conclude, the improved HSL system showed great potential for the remote estimation of vegetation biochemical contents, which will significantly extend the scope of quantitative remote sensing with vegetation.
© 2016 Optical Society of America
1. Introduction
Light detection and ranging (LiDAR) is an active optical sensor that provides direct range measurements of land surfaces, which has been widely used in the estimation of vegetation structure-related parameters [1–3] and land cover mapping [4–6]. Most of the available airborne and terrestrial LiDAR systems only work with one detection channel limited by the laser source [7]. In recent years, concepts and prototypes of multi-wavelength or hyperspectral LiDAR (HSL) have been proposed with the development of hardware [8, 9]. Advances in fibre optics greatly accelerated the development of supercontinuum laser sources, which technically allow the laser detection wavelength be extended into a wide spectral range [10]. The HSL system makes it possible to simultaneously collect spectral and geometric attributes at different spectral channels. This greatly facilitates the layered estimation of vegetation structure and biochemical contents, which is significantly important for quantitative remote sensing of vegetation.
Vegetation index (VI) that defined as the combination of reflectance from near-infrared and visible bands is the most commonly used proxy of remote sensing to estimate vegetation biochemical and biophysical parameters [11–13]. Usually, the sensitivity of one vegetation index to different vegetation biochemical parameters varied significantly. This kind of sensitivity analysis has been widely studied based on the passive multi-spectral and hyperspectral data [14, 15]. Similar to passive remote sensing, light backscatter reflective factors can be derived from LiDAR data [16, 17], which makes it possible to construct LiDAR VIs. However, the study of LiDAR-derived VIs is still in early stage after the presence of multi-spectral LiDAR system. Up to now, a few multi-wavelength LiDAR systems have been proposed and designed for various scientific objectives [18–20]. The VIs that calculated from the backscatter reflective factors of these systems have been successfully applied to distinguish woody and non-woody features [21], to distinguish the vegetation index for Norway spruce [10], to classify different vegetation types [18, 20, 22], and to estimate biochemical contents of vegetation [23–25].
To the best of our knowledge, the number of detection channels for the published systems varied significantly, and not all of the systems are full-waveform systems. It is reasonable to assume that a wider range of spectral detection could greatly increase the possibility of capturing the weak variations of vegetation components. The full-waveform recording approach of LiDAR should be rapidly addressed, because it can provide more structural information by fully recording the interaction procedure between laser pulses and vegetation components with a nanosecond temporal resolution. However, increasing the detection channels usually demands a greater coordination among different hardware components in the system as well as financial cost. In order to detect the 3D distribution of vegetation biophysical and biochemical variations, the authors of this study tried to develop a new full-waveform hyperspectral LiDAR system with 32 detection channels covering the visible and near infrared (NIR) wavelengths. A pioneer prototype of this system with four detection channels has been developed to estimate leaf biochemical contents, and its ranging capability was also tested [25, 26]. Recently, the number of detection wavelengths of this system has been extended to 32, and a new scanner that can be controlled by a computer was also added. However, quality assessment and processing workflow of the full-waveform HLS data still need further investigation. With the increased detection channels, more HSL VIs can be calculated compared with previous studies. Nevertheless, the responses of the HSL VIs to the leaf biochemical variations remain unknown.
Therefore, this study is designed to present and test the improved full-waveform HSL system based on our previous studies [25, 26]. The main focus of this study is to evaluate the data quality of the 32-channel system, and to test its potential on the detection of vegetation biochemical variations. Two different reflective factors will be developed that are expected to describe the backscattering attributes of the detected targets. A full search for the possible combination of the reflective factors at near-infrared and visible wavelengths will be conducted to construct hundreds of VIs. The responses of the high dimensional VIs to the leaf biochemical variations will also be fully tested. A rapid and efficient workflow is expected for the estimation of leaf biochemical contents.
2. Materials
2.1 The hyperspectral LiDAR (HSL) system
The improved optical setup of the HSL system is shown in Fig. 1. The features for the main components in the system are illustrated in Table 1. The system transmits 1-2 ns pulses at repetition rate of 20-40 kHz and peak power of 20 kW produced by a supercontinuum laser (NKT Photonics, SuperK). The spectral range of output laser is 450 nm-2400 nm. The scattered laser pulse from the target is filtered and equally spectrally-resolved by an optical grating system. The broadband white light emitted by the laser is collimated (diverging angle is less than 5 mrad) when it passes through the refracting collimator. The collimated beam is reflected twice by two mirrors (M1, M2) and transmitted into the optical axis of the telescope. The mirror M1 is also used as a beam sampler plate for transmitting a small portion of emitted light through an optical fiber to the avalanche photon diodes (APD1) sensor. The APD1 sensor is used to collect a sample of the emitted laser waveform which triggers a measurement for capturing the emitted and returned waveforms [26]. In addition, a centering device that consists of a Pritchard targeting system is fixed inside the collimator, which is quite different from the previously published systems. The centering device enables the emitted laser be directed to the exact position where the users want to detect, which is important for high precision of detection. The scanner consists of two rotators with x/y double axis, producing an absolute geometry accuracy of ± 0.01°. The rotating angle range and resolution can be controlled by a computer that is connected to the system. The transmitting and returned waveforms were recorded using an oscilloscope. Preliminary experiments on testing the ranging capability of the system with four wavelengths have been reported earlier [25, 26].
Table 1. The main features of the hyperspectal LiDAR system
Compared to the preliminary prototype, the greatest improvement of the system is that its spectral range for the detector was extended in 32 channels using a 32-element APD array module. The spectral response range of the detector was extended to 409 nm ~914 nm. The scattered laser pulse from the target is filtered and equally spectrally-resolved by an optical grating system, making a channel interval of nearly 17 nm. The central wavelength for each detector channel is shown in Table 2 and the full width at half maximum (FWHM) is about 10 nm. The spectrally separated light is converted to analog voltages using the APD array module. There are 19 visible channels at the wavelength before the red edge of vegetation (CH1~CH19) and 13 near-infrared channels (CH20~CH32), which facilitate the construction of various vegetation indices.
Table 2. Central wavelength for the 32-channel detector in the hyperspectral LiDAR system
2.2 Measurement of HSL data
The full-waveform HSL data of two types of tree leaves (n = 30) were measured using the new system. Leaf species includes Lagerstroemia indica (denoted as LI) and Bauhinia blakeana Dunn (BD). The 15 pieces of leaves for each species have different visual appearance in colour and were pasted on a black canvas as background. During the measurements, the distance between the leaves and the HSL system was fixed at 10 m. The laser footprint was strictly controlled to locate leaf surface by the centering device inside the collimator. Five positions on the leaf surface were randomly selected and measured for each wavelength. At each position, five waveforms were collected. Thus, twenty five waveforms were measured in one leaf, which were averaged as the final waveform of this leaf. Meanwhile, a Spectralon® panel was also scanned in a similar way immediately after the measurements of each leaf. The full-waveform data of the panel can be used to calibrate the returned HSL data of leaves.
2.3 Leaf biochemical contents
Leaf biochemical contents were tested in laboratory analysis. A spectrophotometer was used to determine the amount of chlorophyll a and b, and total carotenoid content with 80% acetone using absorption features at different spectral wavelengths. Leaf samples were then oven-dried at 105 °C for 30 minutes, followed by drying at 80 °C until constant weights were reached. The nitrogen contents were determined by Micro-Kjeldahl analysis. The tested biochemical contents for the two leaf species (LI and BD) were as follows: 0.71-3.56% of nitrogen; 0.01-12.71 mg g−1 of chlorophylla/b; 0.06-2.89 mg g−1 of carotenoid [22].
3. Methodology
3.1 Analysis of signal to noise ratio of full-waveform HSL data
Although the scattered laser pulse from the target is filtered and equally spectrally-resolved, the collected full-waveform data at different wavelengths are quite different on pulse amplitude, noise level, and overshooting. Thus, evaluation of data quality was conducted on the raw full-waveform data for each wavelength before the data processing. The signal to noise ratio (SNR: defined as the ratio of useful signal energy to noise energy) was calculated to evaluate the data quality. The mean background noise level was firstly estimated by constructing the histogram of raw waveforms [27]. The histogram of the returned energy in the raw waveforms was fitted by a Gaussian function using the least square algorithm. Then, the energy value corresponding to the peak of the fitted Gaussian curve was identified as the mean background noise. Once the noise level was identified, the returned energy in the raw waveform can be classified to signal and noise. The SNR was calculated using the maximum returned signal and the standard deviation of noise, which can be expressed as [28]:
where is the maximum returned energy of each waveform and is the standard deviation of noise. After the analysis of signal to noise ratio, the mean of background noise was removed from the raw waveform data. In this study, the SNR for each piece of leaf and wavelength was estimated based on the averaged waveforms.3.2 Derivation of HSL reflective factors
In this study, two different backscatter reflective factors (RFs) were derived based on the full-waveform data at each wavelength. Before deriving the HSL reflective factors, the negative overshoot in some waveforms were removed. Then, the maximum echo was indentified and the 20 samples of echoes ranging from −10 to 10 samples around it were selected as the main echoes. The echoes outside the main echo were filled with null values. The selected samples were fitted using a Gaussian model in order to catch the real amplitude peak of the waveform that might be missed due to the instrumental sampling.
where, denotes the modeled LiDAR waveform as a function of time (). , , are the amplitude, position, and width of the Gaussian model, respectively. In our previous study, a simple RF was defined as the ratio of the signal integrals returned from the leaf to that returned from the Spectralon® panel [25]. RF was also calculated in this way for this study.Similar to the principle of microwave radar, the radiative transferring technique for lasers can be described by radar equation [29]. The received power of the LiDAR waveform can be expressed as a function of sensor configuration, scattering attributes of the target, observation geometry and atmospheric attenuation [30]. In this study, a new reflective factor was developed based on the transmitted and received waveform data. According to [30], the reflective factor of LiDAR can also be defined as the ratio of received and transmitted energy:
where, and are the powers of received and transmitted waveform, respectively. As both the received and transmitted waveforms can be fitted by the Gaussian model, the powers () can be approximately calculated as [30]:Then, a normalized reflective factor (NRF) was obtained by normalizing the reflective factor of a leaf () to that of Spectralon® panel (), which is expressed as:As the received and emitted waveforms were simultaneously recorded by the HSL system, the NRF for each leaf and wavelength was directly obtained. Compared with the RF method, the transmitted waveform was used in the derivation of NRF in order to reduce the influence from the slight variations in the HSL system and illumination changes.
3.3 Calculation of HSL vegetation index
A full search for the combination of the reflective factors at visible and near-infrared wavelengths was explored through ratio and normalization operations, respectively. The simple ratio vegetation index () and normalized difference vegetation index () were calculated as:
where and are the reflective factor in visible and near-infrared wavelengths, respectively. Similarly, the vegetation indices were also calculated using the normalized reflective factor NRF developed in this study. Therefore, for both RF and NRF, there were totally 247 combinations for each vegetation index based on the reflective factors from 19 visible bands and 13 near- infrared bands.3.4 Estimation of leaf biochemical contents
The response of the HSL-derived vegetation indices to the leaf biochemical contents were analyzed before the estimation procedure. The correlation between the biochemical contents and vegetation index was quantified using Pearson’s correlation coefficient (R) and demonstrated by a correlation matrix. The vegetation indices that showed the lowest and highest correlation to each kind of biochemical content were indentified. Based on the two types of reflective factors at 32 channels, there were totally 998 predictor variables (247 RF-derived SR, 247 RF-derived NDVI, 247 NRF-derived SR. and 247 NRF-derived NDVI). Nevertheless, many of these variables are highly correlated or even collinear due to similar ratio operations. Therefore, a principal component regression (PCR) model was chosen to estimate each kind of biochemical content based on the high-dimensional predictor variables [31]. The principal component analysis (PCA) was firstly conducted based on the predictor variables. PCA is a commonly used approach in the study of passive hyperspectral remote sensing [32–34], which constructs new predictor variables or principal components (PCs) as linear combinations of the original predictor variables. The percent of variance explained by the predictor variables were estimated for each principle component. To avoid an overly-optimistic estimate of expected errors, the leave one out of cross validation (LOOCV) was used to automatically select the number of PCs during the regression procedure. The estimated mean squared error of prediction (MSEP) was calculated to evaluate the explaining power of a different number of PCs. The number of PCs that significantly minimized the expected errors was finally selected and corresponding PCs were used in the subsequent regression procedure.
The estimation accuracy for the regression models was quantified using coefficient of determination (R2), root mean squared error (RMSE) and relative root mean squared error (rRMSE). RMSE is related to the magnitude of the observed variables, while rRMSE is a relative value that can be used to compare the performances of different regression models. A lower rRMSE often indicates a better regression performance.
where is the measured value, and is the predicted value.4. Results
4.1 HSL data quality and reflective factors
The SNR distributions of the HSL waveform data measured from the two types of leaves are shown in Fig. 2. Both LI and BD leaves showed similar SNR in the same wavelength while the SNR varied significantly among different wavelengths. The detection channels in the middle part (572 nm ~768 nm) showed a higher SNR than those in the two terminal parts. The last eight channels obtained relatively lower SNR with the lowest value of 16.58 dB for LI and 15.68 dB for BD at 898 nm. The SNR at 556 nm obtained the highest standard deviation with a value of 3.20 dB for LI and 1.45 dB for BD leaves. The HSL data in this study generally reached a higher SNR level than the satellite large footprint LiDAR waveform data where the same algorithm was employed in the calculation of SNR [27].
Fig. 2 The signal to noise ratio (SNR, dB) of the full-waveform HSL data measured from the leaves of (a) Lagerstroemia indica (LI) and (b) Bauhinia blakeana Dunn (BD). The error bar represents the standard deviation of SNR.
4.2 Response of leaf biochemical contents to HSL reflective factors
The correlations between leaf biochemical contents and the HSL RF-derived vegetation indices are shown in Fig. 3. Results showed that both SRs and NDVIs calculated from the RFs at longer visible wavelengths (>491 nm) were generally less correlated to the leaf biochemical contents than those at shorter visible wavelengths (≤474 nm). A consistently higher correlation was obtained by the vegetation indices at 458 nm and 474 nm (red and yellow columns in the left part of the correlation matrices). Contrasting with low correlations of nitrogen and carotenoid content, the chlorophyll content also showed higher correlation to the vegetation indices at longer visible wavelengths (523 nm ~670 nm), which can be found in the upper right part in Figs. 3(b) and 3(e). The relationship between the chlorophyll content and RF-derived NDVI605, 751 achieved the highest correlation (R = 0.786). According to the maximum R values, the leaf biochemical contents obtained different feature bands with 458 nm, 784 nm, 914 nm for nitrogen content, 605 nm, 751 nm for chlorophyll content, and 474 nm, 768 nm for carotenoid content (Table 3).
Fig. 3 Correlation matrices showing the relationships between leaf biochemical contents (Nitro, Chl and Caro) and vegetation indices (SRs and NDVIs) that derived from HSL reflective factor (RF).
Table 3. The maximum (Max) and minimum (Min) Person’s correlation coefficient (R) between leaf biochemical contents and HSL-derived vegetation indices together with corresponding band combinations.
The correlation matrices in Fig. 4 show that the leaf biochemical contents were generally more highly correlated to the NRF-derived vegetation indices than the RF-derived ones. Furthermore, the number of vegetation indices with shorter visible wavelengths (< 523 nm) and higher R significantly increased as indicated by the red and yellow bins in the left part of all the matrices. Similarly, the NRF-derived vegetation indices at longer visible wavelength obtained weaker correlations to the leaf biochemical contents. Compared with RF-derived vegetation indices, the highest Pearson’s correlation of NRF-derived vegetation indices generally increased by about 10% in most cases (Table 3). The relationship between the Caro and NRF-derived NDVI507, 703 reached the highest correlation (R = 0.879). The feature bands of the leaf biochemical contents are 491 nm, 784 nm for Nitro, 409 nm, 784 nm for Chl, and 507 nm, 703 nm, 751 nm for Caro, which are slightly different from those obtained from RF. Results showed that the HSL-derived SRs showed a highly consistent relationship to the leaf biochemical contents as the NDVIs did in both case of RF and NRF.
Fig. 4 Correlation matrices showing the relationships between leaf biochemical contents (Nitro, Chl and Caro) and vegetation indices (SRs and NDVIs) that derived from the HSL normalized reflective factor (NRF).
4.3 Estimation of leaf biochemical contents
The principle component analysis showed that the first five principle components (PCs) scored 74.88%, 16.11%, 4.69%, 1.63%, and 1.84% of the variance in the high dimensional HSL-derived vegetation indices, respectively Fig. 5(a). In other words, 99.16% of the variance in the 988 predictor variables can be explained by the first five PCs. The results from cross-validated PCR showed that the estimated mean squared error of prediction (MSEP) for the PCs varied specific to each leaf biochemical content parameter Figs. 5(b)–5(d). The MSEP decreased rapidly at the sixth PC for Nitro, the fifth PC for Chl and Caro, respectively. Thus, the number of PCs that sufficient enough to predict the leaf biochemical parameters was finally settled as 6 for Nitro, 5 for Chl and Caro, respectively.
Fig. 5 (a) The number of principle components and corresponding percent variance explained in the predictor data set; and the estimated mean squared error of prediction obtained from cross-validated principle component analysis for (b) nitrogen content (Nitro, mg g−1); (c) chlorophyll content (Chl, mg g−1); and (d) carotenoid content (Caro, mg g−1).
Then, the regression models were established based on the selected PCs and observed biochemical contents. The cross validation of PCR showed that the estimated leaf biochemical contents were highly correlated with the observed ones Fig. 6(a)–6(c). The estimation of Nitro obtained the highest robustness with the lowest rRMSE of 21.49% although the R2 (0.71) was slightly lower than the PCR models for the other two parameters. The PCR model reached the highest explained variation of the observed Chl (R2 = 0.83) with a RMSE of 1.41 mgg−1. The estimated Caro was also highly comparable to the observed one with an R2 of 0.77 and RMSE of 0.38 mg g−1.
Fig. 6 The observed and PCR-predicted values for (a) nitrogen content (Nitro, mg g−1); (b) chlorophylla/b content (Chl, mg g−1); and (c) carotenoid content (Caro, mg g−1).**p<0.01.
5. Discussion
5.1 HSL-derived reflective factors
In this study, the quality evaluation of the raw full-waveform data showed that different detection channels reached different SNR in the detection of two types of leaves. The relative lower SNRs at shorter visible and longer near-infrared wavelengths were mainly attributed to the decreased sensitivity of the detector at those channels. This directly increased the occurrence of noise-like negative overshoots in the waveform at those channels. Future efforts are still needed to improve the detector sensitivity in those channels. Two different reflective factors were developed based on the HSL waveform data. As both transmitted and received waveforms for the leaves and Spectralon® panel were recorded for each measurement, the ratio operations in the calculation of reflective factors helped to decrease the influence of the system errors and illumination changes. As for the passive reflectance measurements, the HSL reflective factors can give insights into the backscattering capability of targets in the hotspot direction. According to our previous study, the HSL-derived RF was generally higher than the passive reflectance following a consistent varying trend [25]. In spite of this, the HSL RF does not strictly follow the definition of reflectance factor and great care should be taken when accurate absolute values are needed [25]. The NRF developed in this study showed another exploration of the HSL data in the perspective of laser power. Compared with RF, the transmitted waveform was used in the derivation of NRF in order to reduce the influence of slight variations in the observation conditions. This made the NRF likely to better describe the backscattered radiation from the observed leaves. The improved relationship between the NRF-derived VIs and leaf biochemical contents indicated that NRF can play a role in capturing the slight biochemical variations in the two types of leaves measured in this study. Future studies should be conducted to test the sensitivity of these two reflective factors to other types of targets.
Aside from the system errors, the observation geometry can also influence the HSL backscatter signals such as the local incidence angle [35] and the stability of the laser source. In this study, influence from the incidence angle was small because the laser spots were nearly perpendicular to the leaf surface and all the backscattering signals were returned from the leaf surface. However, the effect from sub-footprints should be considered in future derivation of HSL reflective factors, because in most cases the laser transmits through many small objects (e.g. coniferous needles and shoots) with different incident angles and optical path. For the analysis of sub-footprint returns, the derivations of LiDAR backscatter reflective factors are expected to be different [30]. The ratio operation for VIs is likely to reduce the influence of the incidence angle and target area [35, 36]. Limited by the recording component of oscilloscope, 8 pulses for a full 32 channel measurement (4 channels for each transmitted pulse) were needed. The measured 32 channel spectrum might have large variability if each pulse has varying intensity, which can cause errors in the RF calculations where the transmitted waveform was not used. Therefore, the averaging of waveforms at each wavelength helped to reduce this kind of errors. On the contrary, this will not be a problem for NRF, because all the transmitted pulses were involved and normalized in the calculation of NRF.
5.2 HSL-derived VIs and corresponding relation to leaf biochemical contents
High correlations were found between the RF-derived VIs at shorter visible wavelengths (≤ 474 nm) and leaf biochemical contents. This relationship was more evident between the NRF-derived VIs and leaf biochemical contents as the number of VIs at visible channels significantly increased (< 523 nm). This indicated that the NRF-derived VIs were more sensitive to the leaf biochemical contents. Previous studies reported that the reciprocal reflectance in the range 510 to 550 nm was found to be closely related to the total pigment content in leaves [37], and similar results were found in this study since the VIs were calculated based on the ratio operation as functions of reciprocal reflective factors. The feature bands for different biochemical contents were slightly different when different reflective factors were used, which can provide an alternative indicator for future studies on the estimates of vegetation parameters. In this study, only two types of vegetation index were tested but they can fully describe the possible contrast between the spectral absorption in the visible wavelengths and spectral reflection at near-infrared wavelengths. The spectral absorption and reflection at different wavelengths are mainly attributed to the variations in pigments and leaf cellular structure, respectively [11]. Thus, we believe that the HSL-derived SRs and NDVIs are sufficient enough to predict the leaf biochemical contents. However, other types of vegetation index should be considered when the observing experiments are conducted on vegetation canopies due to the significant contributions to the backscatter reflective factors from the soil background and canopy structure.
5.3 Estimation of leaf biochemical contents based on HSL data
Single wavelength LiDAR data have been widely used in the estimation of vegetation structure-related parameters, but few cases can be found in the estimation of biochemical contents due to the lack of sufficient spectral information. In this study, the 32-channel hyperspectral full-waveform data was tested in the estimation of leaf biochemical contents for the first time. Compared with previous studies on multi-spectral LiDAR, the unique features of this study are the large number of detection channels and the full-waveform recording in data collection. The increase of the number of detection channels significantly enriched the spectral data source, making it to detect the geometric and spectral variations of different types of targets. In addition, the simultaneous collections of geometric and spectral attributes successfully overcome the problem of registration between different types of data source (e.g. single-wavelength LiDAR and passive hyperspectral images). The full-waveform data helped to record more details on the interactions between laser and vegetation targets, which are important for the reconstruction of vegetation structure, because the structure of vegetation canopies often varies significantly in the 3D space.
For the estimation of leaf biochemical contents, the PCA approach successfully reduced the dimension of hundreds of VIs and accelerated the estimation procedure. Usually, simply using a large number of components can obtain good results in fitting the current observed data, but it may lead to over fitting and make the model unsuitable for other data sets. The variance explained in the predictor variables helped to select a rough number of top PCs, but no consideration on the biochemical contents was involved. Thus, the LOOCV procedure further helped to choosing the number of components in regression models according to the variation of MSEP. The PCR estimation procedure provides some suggestions on how to take full use of the HSL data in future. However, all the analysis was conducted on the individual leaf level. Future testing on vegetation canopies should be conducted to assess the generalization of PCR models. When the HSL data at canopy and landscape level are collected, a rapid and efficient data processing workflow must be used because the data volume will be greatly increased. Furthermore, the sensitivity of the 32 chosen channels to the biophysical and biochemical variations in vegetation canopies needs further investigation.
6. Conclusion
In this study, an improved prototype of full-waveform LiDAR system with 32 channels was presented and tested in the estimation of leaf biochemical contents for the first time. Two different reflective factors were developed based on the HSL data collected from two types of leaves, which were used to calculate hundreds of VIs. Results showed that high correlations were found between the HSL-derived VIs at shorter visible wavelengths and leaf biochemical contents. The high-dimensional VIs (n = 998) were used to estimate three leaf biochemical contents using principle component regression (PCR) models with cross validation. Results showed that the VIs derived from the newly developed reflective factor (NRF) generally showed higher correlations to the leaf biochemical contents. The prediction of biochemical contents obtained satisfactory results with a RMSE of 0.45% for Nitro, 1.41 mg g−1 for Chl, and 0.38 mg g−1 for Caro. The PCR model successfully explained 71%, 83% and 77% of the variation in the observed Nitro, Chl and Caro, respectively. To conclude, the new HSL system showed great potential for the remote estimation of vegetation biochemical contents, which will significantly extend the scope of quantitative remote sensing with vegetation.
Acknowledgments
This work was supported by China’s Special Funds for Major State Basic Research Project of China under grant 2013CB733405; the National Natural Science Foundation of China under grant 41201345 and 41301389.
References and links
1. S. Luo, C. Wang, X. Xi, and F. Pan, “Estimating FPAR of maize canopy using airborne discrete-return LiDAR data,” Opt. Express 22(5), 5106–5117 (2014). [CrossRef] [PubMed]
2. L. Korhonen, I. Korpela, J. Heiskanen, and M. Maltamo, “Airborne discrete-return LIDAR data in the estimation of vertical canopy cover, angular canopy closure and leaf area index,” Remote Sens. Environ. 115(4), 1065–1080 (2011). [CrossRef]
3. W. Li, Z. Niu, N. Huang, C. Wang, S. Gao, and C. Wu, “Airborne LiDAR technique for estimating biomass components of maize: A case study in Zhangye City, Northwest China,” Ecol. Indic. 57, 486–496 (2015). [CrossRef]
4. Y. Qin, S. Li, T. T. Vu, Z. Niu, and Y. Ban, “Synergistic application of geometric and radiometric features of LiDAR data for urban land cover mapping,” Opt. Express 23(11), 13761–13775 (2015). [CrossRef] [PubMed]
5. W. Li, Z. Niu, B. Yu, and S. Gao, “Research on classification of LiDAR images derived from waveform decomposition over a suburban area,” Optik (Stuttg.) 125(23), 6898–6902 (2014). [CrossRef]
6. W. Wagner, M. Hollaus, C. Briese, and V. Ducic, “3D vegetation mapping using small-footprint full-waveform airborne laser scanners,” Int. J. Remote Sens. 29(5), 1433–1452 (2008). [CrossRef]
7. J. Suomalainen, 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]
8. S. Kaasalainen, T. Lindroos, and J. Hyyppa, “Toward hyperspectral lidar: Measurement of spectral backscatter intensity with a supercontinuum laser source,” IEEE Geosci Remote. 4(2), 211–215 (2007). [CrossRef]
9. A. J. Brown, T. I. Michaels, S. Byrne, W. Sun, T. N. Titus, A. Colaprete, M. J. Wolff, G. Videen, and C. J. Grund, “The case for a modern multiwavelength, polarization-sensitive LIDAR in orbit around Mars,” J. Quant. Spectrosc. Radiat. Transf. 153, 131–143 (2015). [CrossRef]
10. Y. Chen, E. Räikkönen, S. Kaasalainen, J. Suomalainen, T. Hakala, J. Hyyppä, and R. Chen, “Two-channel hyperspectral LiDAR with a supercontinuum laser source,” Sensors (Basel) 10(7), 7057–7066 (2010). [CrossRef] [PubMed]
11. C. Y. Wu, Z. Niu, Q. Tang, and W. J. Huang, “Estimating chlorophyll content from hyperspectral vegetation indices: Modeling and validation,” Agric. For. Meteorol. 148(8-9), 1230–1241 (2008). [CrossRef]
12. C. Adjorlolo, O. Mutanga, and M. A. Cho, “Estimation of Canopy Nitrogen Concentration Across C3 and C4 Grasslands Using WorldView-2 Multispectral Data,” IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 7(11), 4385–4392 (2014). [CrossRef]
13. H. Croft, J. M. Chen, and Y. Zhang, “Temporal disparity in leaf chlorophyll content and leaf area index across a growing season in a temperate deciduous forest,” Int. J. Appl. Earth. Obs. 33, 312–320 (2014). [CrossRef]
14. I. D. A. Sanches, C. R. Souza Filho, and R. F. Kokaly, “Spectroscopic remote sensing of plant stress at leaf and canopy levels using the chlorophyll 680 nm absorption feature with continuum removal,” ISPRS J. Photogramm. Remote Sens. 97, 111–122 (2014). [CrossRef]
15. C. Y. Wu, X. Z. Han, Z. Niu, and J. J. Dong, “An evaluation of EO-1 hyperspectral Hyperion data for chlorophyll content and leaf area index estimation,” Int. J. Remote Sens. 31(4), 1079–1086 (2010). [CrossRef]
16. A. J. Brown and Y. Xie, “Symmetry relations revealed in Mueller matrix hemispherical maps,” J. Quant. Spectrosc. Radiat. Transf. 113(8), 644–651 (2012). [CrossRef]
17. A. J. Brown, “Equivalence relations and symmetries for laboratory, LIDAR, and planetary Müeller matrix scattering geometries,” J. Opt. Soc. Am. A 31(12), 2789–2794 (2014). [CrossRef] [PubMed]
18. 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] [PubMed]
19. A. M. Wallace, A. McCarthy, C. J. Nichol, X. Ren, S. Morak, D. Martinez-Ramirez, I. H. Woodhouse, and G. S. Buller, “Design and Evaluation of Multispectral LiDAR for the Recovery of Arboreal Parameters,” IEEE Trans. Geosci. Remote. 52(8), 4942–4954 (2014). [CrossRef]
20. W. Gong, S. Song, B. Zhu, S. Shi, F. Li, and X. Cheng, “Multi-wavelength canopy LiDAR for remote sensing of vegetation: Design and system performance,” ISPRS J. Photogramm. Remote Sens. 69, 1–9 (2012). [CrossRef]
21. S. Tan and R. M. Narayanan, “Design and performance of a multiwavelength airborne polarimetric lidar for vegetation remote sensing,” Appl. Opt. 43(11), 2360–2368 (2004). [CrossRef] [PubMed]
22. J. Vauhkonen, T. Hakala, J. Suomalainen, S. Kaasalainen, O. Nevalainen, M. Vastaranta, M. Holopainen, and J. Hyyppa, “Classification of Spruce and Pine Trees Using Active Hyperspectral LiDAR,” IEEE Geosci Remote. 10(5), 1138–1141 (2013). [CrossRef]
23. O. Nevalainen, T. Hakala, J. Suomalainen, R. Mäkipää, M. Peltoniemi, A. Krooks, and S. Kaasalainen, “Fast and nondestructive method for leaf level chlorophyll estimation using hyperspectral LiDAR,” Agric. For. Meteorol. 198–199, 250–258 (2014). [CrossRef]
24. 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,” Int. J. Appl. Earth. Obs. 44, 136–143 (2016). [CrossRef]
25. W. Li, G. Sun, Z. Niu, S. Gao, and H. L. Qiao, “Estimation of leaf biochemical content using a novel hyperspectral full-waveform LiDAR system,” Remote Sens. Lett. 5(8), 693–702 (2014). [CrossRef]
26. Z. Niu, Z. Xu, G. Sun, W. Huang, L. Wang, M. Feng, W. Li, W. He, and S. Gao, “Design of a New Multispectral Waveform LiDAR Instrument to Monitor Vegetation,” IEEE Trans. Geosci. Remote. 12, 1–5 (2015).
27. S. Nie, C. Wang, G. Li, F. Pan, X. Xi, and S. Luo, “Signal-to-noise ratio–based quality assessment method for ICESat/GLAS waveform data,” Opt. Eng. 53(10), 103104 (2014). [CrossRef]
28. C.-K. Wang, “Exploring weak and overlapped returns of a lidar waveform with a wavelet-based echo detector,” Int. Arch. Photogramm. Remote Sens. Spat. Inf. Sci 39, B7 (2012).
29. W. Wagner, A. Ullrich, V. Ducic, T. Melzer, and N. Studnicka, “Gaussian decomposition and calibration of a novel small-footprint full-waveform digitising airborne laser scanner,” ISPRS J. Photogramm. Remote Sens. 60(2), 100–112 (2006). [CrossRef]
30. Y. Qin, W. Yao, T. Vu, S. Li, Z. Niu, and Y. Ban, “Characterizing Radiometric Attributes of Point Cloud Using a Normalized Reflective Factor Derived From Small Footprint LiDAR Waveform,” IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 8(2), 740–749 (2015). [CrossRef]
31. R. Marbach and H. M. Heise, “Calibration modeling by partial least-squares and principal component regression and its optimization using an improved leverage correction for prediction testing,” Chemometr. Intell. Lab. 9(1), 45–63 (1990). [CrossRef]
32. A. Gowen, C. O’donnell, M. Taghizadeh, P. Cullen, J. Frias, and G. Downey, “Hyperspectral imaging combined with principal component analysis for bruise damage detection on white mushrooms (Agaricus bisporus),” J. Chemometr. 22(3-4), 259–267 (2008). [CrossRef]
33. J. A. Jablonski, T. J. Bihl, and K. W. Bauer, “Principal Component Reconstruction Error for Hyperspectral Anomaly Detection,” IEEE Trans. Geosci Remote S. 12(8), 1725–1729 (2015). [CrossRef]
34. F. Tao, Y. Peng, and Y. Li, “Feature extraction method of hyperspectral scattering images for prediction of total viable count in pork meat,” Int. J. Agric. Biol. Eng. 8, 95–105 (2015).
35. 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]
36. J. U. H. Eitel, T. S. Magney, L. A. Vierling, T. T. Brown, and D. R. Huggins, “LiDAR based biomass and crop nitrogen estimates for rapid, non-destructive assessment of wheat nitrogen status,” Field Crops Res. 159, 21–32 (2014). [CrossRef]
37. A. A. Gitelson, Y. Zur, O. B. Chivkunova, and M. N. Merzlyak, “Assessing carotenoid content in plant leaves with reflectance spectroscopy,” Photochem. Photobiol. 75(3), 272–281 (2002). [CrossRef] [PubMed]
