1. Introduction
Precision agriculture, also known as precision farming, refers to the utilization of modern information technology for intensive farming. The goal of precision agriculture research for this purpose is to optimize the return on investment while preserving resources. Precision agriculture can improve the production efficiency of field operations and reduce the pollution of agriculture to the environment. Spectral profiles collected using a spectrometer have been widely employed by the scientists in precision agriculture for vegetation growth status indication, vegetation classification, etc. [
1,
2]. With fine spectral resolution and coverage, collected spectral information is sufficient for many vegetation applications in precision agriculture, e.g., spectral indexes extraction, water content estimation, chlorophyll content estimation, and vegetation “red edge” detection [
3,
4,
5]. By its nature, different vegetation has different biological structures that contribute to their unique spectral profiles [
6,
7,
8]. Precise measuring and mapping of these contents are of great significance in precision agriculture, e.g., estimating the vegetation growth status or the three-dimensional distribution of these vegetation contents could guide and guarantee more efficient fertilization and pesticide spraying [
6,
7]. However, the traditional passive spectrometer is sensitive to environmental lighting conditions [
6,
7,
8]. Restricted by this flaw, actively acquiring spectral profiles has been the research hotspot in the community; additionally, it is difficult to extract spatial information directly from the spectral imaging results. It is of great importance to develop active methods for collecting spectral profiles and spatial information [
6,
7,
8].
Hyperspectral LiDAR (HSL) is a cutting-edge instrument to actively collect the spectral information with spatial measurements. HSL combines the two complementary functions (spatial and spectral information collecting) into a single framework and produces point clouds accompanying spectral profiles. HSL emits a laser pulse with a ultrawide spectral range, and the spectral data is extracted from the intensity of the reflected laser pulses while the spatial information is calculated by counting the time of flight between the transmitted laser pulse and the received pulse [
7,
8,
9,
10,
11]. During recent years, researchers have endeavored to develop multispectral laser scanning (MLS) for vegetation remote sensing applications; an instant method is to combine several monochromatic laser sources working at different spectral wavelengths together, and a laser combiner can be employed to guarantee the laser beam combination. With this operation, the laser beams from the MLSs are pointed to the same part of the target at the same incident angle and transmitting distance [
7,
8,
9,
10,
11]. However, it is problematic to combine too many single wavelength laser sources for extending the spectral bands coverage and spectral resolution. Most of the MLSs are specially designed for specific applications, and they work at the selected wavelength. Meanwhile, more laser sources mean more hardware cost, a more complex optic system, higher power consumption, and bulky size. Some efforts have been devoted to exploring the MLS in vegetation remote sensing applications, i.e., quantify the vegetation nutrition contents, tree classification, etc. [
10,
11]. These efforts preliminarily demonstrate that the LiDAR with multiple wavelengths is feasible for a vegetation remote sensing application. However, the MLSs have low spectral resolution and restricted spectral bands. These MLSs are not capable of acquiring abundant spectral profiles comparable to that from the spectrometer. The spectral bands should be optimized in advance and designed for particular vegetation spectral indexes determination, which lacks the scalability and versatility of precision agriculture application. Motivated by this problem, another HSL architecture employing a supercontinuum laser (SC) as the laser source is a more practical solution for developing HSL with wider spectral coverage and finer spectral resolution [
12,
13,
14]. In the SC-based HSL, SC source is usually capable of transmitting a laser beam with the spectral wavelength covering approximately 450~2500 nm. With a properly designed optical system and reflected laser-pulses-detection system, it is feasible to develop HSL with better spectral resolution and wider spectral coverage.
Due to the availability of the commercial SC source, the HSL concept was firstly set up, prototyped, and tested by scientists from the Finnish Geospatial Research Institute (FGI) in 2007 [
14]. Based on this, in 2010, a two-channel LiDAR (600 nm and 800 nm) based on the SC source was constructed utilizing commercial components [
15]. In addition, normalized difference vegetation index (NDVI) parameters for Norway spruce were calculated and presented with the two-channel laser-reflected spectral measurements, and the potential of using NDVI in Norway spruce plant classification was firstly investigated and evaluated. In 2012, a full-waveform hyperspectral LiDAR was assembled by researchers from FGI. The novel instrument was able to generate 3D point clouds accompanying eight-channel spectral backscattered reflectance data [
16]. With this instrument, both geometry and spectral information was obtained concurrently, which had great potential for extending the scope of imaging spectroscopy into spectral 3D sensing [
16]. In 2013, researchers investigated the classification of spruce and pine trees using this eight-channel HSL [
17]. Later, in 2016, a 32-channel hyperspectral full-waveform LiDAR covering 400–1000 nm was first presented by Li. [
18]. With the spectral channels increasing, the HSL was employed for vegetation biochemical contents. However, the spectral channels and bands were selected in advance for some biochemical contents’ estimation. In 2018, researchers from the Academy of Optoelectronic (AOE), China Academy of Sciences extended the spectral band from visible spectrum (VIS) to shortwave infrared spectrum SWIR [
19]. The HSL covering 500~1500 nm with 17 channels was employed in ore classification and obtained satisfying classification results due to more spectral channels deployment.
In 2018, with the aim to enhance the HSL spectral resolution, a liquid crystal tunable filter (LCTF) was employed in the HSL (LCTF-HSL) for selecting and filtering the passing laser beam with 10 nm spectral resolution [
20,
21]. With finer spectral resolution, the LCTF-HSL was utilized to extract vegetation “red edge” (RE) parameters [
21]. However, the LCTF working spectral range limited the spectral coverage of the HSL. Another wavelength filter Acousto-Optic Tunable Filter (AOTF) with quicker response and broader spectral range was employed for filtering the spectrum of the ultra-wideband laser beam instead of the LCTF in the HSL. AOTF-HSL was able to cover the spectral range from 500 nm to 1000 nm with a 10 nm resolution, which met the demand for extracting various spectral parameters [
22,
23,
24]. Apart from the improvement in spectral resolution and the band coverage; recently, a portable HSL was presented and employed in underground mining application [
25]. A pulse digitizing scheme in this design was improved, and size was reduced [
25]. Summary of the SC source-based HSL development in terms of spectral range and resolution is listed in the following
Table 1.
Compared with conventional spectrometer, most the HSLs have limited spectral bands and channels, which is insufficient for precision agriculture application. With the aim to provide a more practical HSL to satisfy the demands of the precision agriculture in acquiring abundant spectral profiles, this paper aims to reveal a feasible HSL design with wide spectral coverage and fine spectral resolution; additionally, the HSL performance was assessed with three different experiments: spectral profiles acquirement, spectral parameters extraction, and vegetation classification. The contributions of this paper could be summarized as:
- (1)
for the demand of the vegetation spectral profiles and spectral parameters acquirement in precision agriculture, we presented a practical HSL design; spectral profiles with 10 nm spectral resolution covering 500~1000 nm were acquired; and the spectral parameters widely used in precision agriculture for vegetation status monitoring were extracted and assessed.
- (2)
for the demand of vegetation classification in precision agriculture, based on spectral parameters results, the vegetation sample leaves classification results using the spectral parameters as the input features of a designed support vector machine (SVM) classifier with multiple labels were presented, and the feasibility was assessed.
- (3)
in view of the HSL laser intensity related to the laser transmitting distance and incident angle, the influence of the laser beam incident angle on the spectral parameters calculation was analyzed; the results showed that the spectral parameters with “(A − B)/(A + B)” and “A/B” were independent with the laser beam incident angle.
The paper is organized as follows:
Section 2 presents a short and brief description of the HSL design and the corresponding equations of the selected vegetation indexes;
Section 3 presents the laser beam radiation model and the analysis of the incident angle on the several forms of the spectral indexes calculation;
Section 4 presents the experimental setup, the results, analysis, and the comparisons; and the next sections include the discussion, conclusions, and reference.
4. Results
In view of the demands of the precision agriculture, three experiments were carried out in this section for assessing the HSL-based spectral profiles acquirement, spectral parameters extraction, and vegetation classification. In the first experiments, two different plants, Dracaena (Dracaena angustifolia) and Aloe (Aloe arborescens Mill) with both green and yellow leaves, and another two different plants, Rubber tree (Ficus elastica Roxb. ex Hornem) and Radermachera (Radermachera hainanensis Merr) with only green leaves, were measured by the AOTF-HSL (plants employed in this experiment are the same as that in our previous paper; the figures of these plants can be found in the literature [
22]). The experiments were carried out indoors under controlled environment for collecting the spectral profiles. The plant was placed approximately 17 m in front of the HSL. Since the spectral pulse intensity was related to the incident angle, the laser pulse was perpendicularly pointed to the surface of the vegetation leaves. Apart from the AOTF-HSL, a passive SVC spectrometer (SVC HR-1024) was employed for acquiring the spectral profiles of the above four different plants. The extracted vegetation spectral indexes from the HSL were compared with that from a passive SVC spectrometer for evaluating the difference, which aimed to assess the feasibility of using HSL to extract the above spectral indexes; then, the results of the green and yellow leaves were compared for validating the differences of the selected vegetation spectral indexes, which could demonstrate the effectiveness of the HSL in vegetation growth status indication and vegetation physiological changes detection.
In addition, in the second experiment, another sixteen different leaves were also measured by the AOTF-HSL. A multi-label SVM was designed and employed to classify the leaves. Based on the assessing results of the vegetation parameters, six different spectral parameters were calculated and utilized as the input features of the SVM classifier. The classification results were presented, analyzed, and discussed for demonstrating the great AOTF-HSL potential in vegetation classification and mapping.
Finally, in the third experiment, Fraxinus velutina, Eucommia ulmoides, and Prunus triloba Lindl green leaves and Fraxinus velutina yellow leaf were measured by the HSL and an ASD FieldSpec 4 spectrometer (ASD© FieldSpec 4). The spectral profiles from the spectrometer were employed as the reference to evaluate the results from the HSL. These leaves were measurement by the HSL at different incident angles (0°, 10°, 20°, 30°, 40°, 50°, 60°, 70°, and 80°). The laser pulse incident angle was set by a compass. Distance information and spectral information were obtained at different incident angles. Normalized difference vegetation index (NDVI) and water index (WI) were employed as the representative parameters of the “(A − B)/(A + B)” and “A/B” styles. NDVI and WI results were calculated using the spectral profiles collected at different incident angles.
4.1. Experiment #01—Spectral Indexed Results Comparison
4.1.1. Green Leaf Results
WI extraction results for the employed four different plants with green leaves are presented in
Table 3. It was notable that all the WI values of the four different green leaves were all below one, and the differences between HSL and SVC spectrometer results were all below 5%. Besides, the difference equation was as the following Equation (11):
where the
was the extracted parameters from the HSL measurements, and the
was the corresponding parameters from the SVC spectrometer spectral profiles.
We could observe that the NDVI from HSL and SVC spectrometer performed quite similarly with the results listed in
Table 3. The Dracaena green leaf difference was below 1%, which was the minimum difference among the four green leaves. By contrast, the Aloe green leaf had the maximum difference, and the difference value was 3.8%. In aspects of the GNDVI results, the Dracaena and Aloe green leaves’ difference were more significant than 5%. The Rubber plant and Radermachera green leaves’ difference between HSL and SVC spectrometer were all below 5%.
By comparing the results difference between NDVI and GNDVI, the GNDVI yielded a slightly bigger difference. Through analyzing the NDVI and GNDVI calculation equations (Equations (2) and (3)) and the spectral profiles plotted in
Figure 4, we observed that the spectral reflectance at 750 nm derived from the HSL on Aloe green leaf of the HSL was larger than that of SVC, while they had similar reflectance at 550 nm. This accounted for the phenomenon that Aloe green leaf yielded a larger difference in NDVI and GNDVI parameter extraction.
The rest of the four parameters, TCI, NDREI, RVI, and FRI, were related to the vegetation biochemical and biological contents. TCI, NDREI, RVI, and FRI from HSL and SVC spectrometer spectral profiles are listed in
Table 3. It could be observed that the HSL and SVC yielded more significant differences compared with previous WI, NDVI, and GNDVI parameters results.
For the Dracaena green leaf, the difference of the TCI, NDERI, RVI, and FRI was larger; only NDERI was below 20%. As listed in
Section 2, the spectral profiles covering 500~700 nm were included in the calculation of the TCI, NDERI, RVI, and FRI values, and the reflectance values from HSL and SVC were slightly different. The spectral profiles comparison is shown in
Figure 4a,b. As presented in these figures, compared with the spectral profiles of 700~1000 nm, the 500~700 nm spectral reflectance was quite different.
In aspects of Aloe green leaf results, the RVI parameters from HSL and SVC show clear difference. As plotted in
Figure 5b, the reflectance values (500~700 nm) from HSL and SVC were different, and HSL spectral values were larger than SVC. Especially different was the spectral profiles’ difference at the wavelength at 660 nm (marked with a pink rectangle in
Figure 5b); this would affect the RVI index determination results.
From the presented Rubber tree’s and the Radermachers’s spectral curves in
Figure 6 and
Figure 7, we could observe that the results were also affected by the difference of the 500~700 nm spectral profiles reflectance. Among them, the Radermachera yielded the best similarity in the 500~700 nm spectral reflectance values between HSL and SVC.
4.1.2. Yellow Leaf Results Analysis
Previously,
Section 3.1 listed the green leaves’ results of these vegetation spectral indexes, and the following
Table 4 shows the results from the Dracaena yellow leaf and the Aloe yellow leaf. Additionally, the green leaves’ results are listed in the
Table 4 for comparison. Among the yellow leaves results, the difference was unsatisfactory apart from the Dracaena yellow leaf WI and RVI index results. The spectral profiles from the HSL and SVC spectrometer are presented in
Figure 8 and
Figure 9. Compared with green leaf results, the curves were relatively distinctive due to the fact that the HSL laser pulse footprint diameter was approximately 1 cm on the target with these settings. The sampled area on the yellow leaves had huge difference between two apparatuses; this might also be the reason that the spectral profiles collected by the HSL were steeper at red edge, as shown in
Figure 8 and
Figure 9. Furthermore, the yellow leaves had uneven distributions of the biochemical and biological contents in the covered area of the footprint.
Through analyzing the difference of the parameters between green and yellow leaves, the indexes demonstrated similar changing trends. For instance, the yellow leaf WI index had a minor increase due to the lower water content in yellow leaves. The NDVI values also performed a dramatic decrease since the yellow leaves had lower chlorophyll. The similar changing trend could be observed from the results of other vegetation spectral indexes. Results from the comparison between green and yellow were able to give positive support for the feasibility study of using the HSL in vegetation spectral indexes estimation, determination, and growth status indicating.
4.2. Experiment #02—Vegetation Leaves Classification
Figure 10 presents an example of scanning the leaf with HSL. The red point was the footprint of the HSL laser pulse. The collected spectral profiles of the sixteen different leaves are presented in
Figure 10; the spectral profiles plotted in
Figure 11a included the Chinese pine (Juniperus rigida), Ginkgo (Ginkgo biloba), Metasequoia (Metasequoia glyptostroboides), hawthorn (Crataegus pinnatifida), Horse chestnut (Aesculus hippocastanum), Toothes oak (Quercus), Platanus occidentalis, and Koelreuleria Paniculata. The spectral profiles of the left eight different leaves in
Figure 11b included Morusalba Tortuosa, Catalpa speciose, Willow tree (Salix), Chinese scholartree (Styphnolobium japonicum), Quercus dentate Thunb, liriodendron chinense (Chinese Tulip Tree), Eucommiaulmoides, and Syringa reticulate var. mandshurica. Limited by the HSL working spectral coverage, the spectral profiles presented in
Figure 11 were sampled ranging from 650 nm to 1100 nm. Another spectral parameter set (NDVI, SR, NDVI705, VOG, WI, and RVI) was extracted and presented. SR, NDVI705, and VOG were the new parameters different from that of the first experiment. The definitions and explanations were listed as the following Equations (23)–(25).
(1) Simple ratio Iidex (SR)
where
and
are the spectral reflectance of the wavelength 800 nm and 680 nm.
(2) Vogelmann index (VOG)
where
and
are the spectral reflectance of the wavelength 800 nm and 680 nm.
(3) Normalized difference vegetation index (NDVI705)
Similarly, and are the spectral reflectance at 750 nm and 705 nm wavelength.
Table 5 presents the calculation results of the spectral indexes for these leaves. In this experiment, each leaf was scanned by the HSL four times. The spectral profiles presented in
Figure 11 were the average values of the four-time measurements. Parameters results listed in
Table 5 were also calculated using the averaged spectral profiles measurements.
Figure 12 plots the parameters of the employed sixteen leaves. Basically, the parameters of these leaves were distinctive, which indicated that these leaves were divisible using these spectral parameters as the feature vector.
As presented in
Figure 12, different parameters are not in the same orders of magnitude. Therefore, before employing the parameters in the classification, the normalization of the parameters is necessary. The normalized parameters of these leaves are plotted in
Figure 13. Compared with
Figure 12, the parameters are more balanced. The feature vector composed by the normalized six parameters is employed as the input vector of the SVM classifier. Since we have sixteen different leaves, a SVM classifier with multiple labels was designed and its structure is presented in
Figure 14.
We scanned the leaves at different spectral wavelengths, and the spectral intensities from these wavelengths were randomly selected to generate more samples for training the classifier. For instance, we scanned these leaves at 750 nm eight times for each leaf; then, we randomly selected one measurement from the eight, and the same operation was conducted for other spectral wavelengths and leaf samples. Here, a training dataset was randomly selected from the first 50% of the datasets and the testing dataset was selected from the rest of the 50% of datasets. The training and testing dataset both contained 1000 samples, which was randomly selected from the collected dataset.
Table 6 lists the classification accuracy of the sixteen leaves using the designed multi-label SVM classifier. All the classification accuracies could reach 100% with the dataset. No SVM parameters optimization was carried out; the multi-label SVM classifier was built on the default SVM in MATLAB. The kernel function was revised from “linear” to “RBF”.
4.3. Experiment #03—Influence of the Incident Angle Analysis on Spectral Indexes Results
In this experiment, normalized difference vegetation index (NDVI) and water index (WI) were employed as the representative parameters of the “(A − B)/(A + B)” and “A/B” styles. NDVI and WI results of the
Fraxinus velutina,
Eucommia ulmoides, and
Prunus triloba Lindl green leaves and
Fraxinus velutina yellow leaf are presented in the
Figure 14. A 100% whiteboard was employed as the reference to calculate these NDVI and WI results. As analyzed in the
Section 2.2, the incident angle had no influence on the NDVI and WI results. In
Figure 15, the incident angles were 0°, 10°, 20°, 30°, 40°, 50°, 60°, 70°, and 80°. The distance remained the same while measuring the whiteboard and these leaves. It could be seen that the NDVI and WI values varied at a limited range, which demonstrated that the incident angle almost had no influence on the NDVI and WI results.
As aforementioned, an ASD FieldSpec 4 spectrometer was employed to collect the spectral profiles, which could be regarded as the reference.
Figure 16 presents the difference results between the spectral parameters from the HSL and spectrometer. In aspects of the NDVI difference,
Eucommia ulmoides and
Prunus triloba Lindl green leaves and
Fraxinus velutina yellow leaf NDVI difference were smaller than 0.2 at all the selected incident angles. In addition, the most of the WI difference results were smaller than 0.1.