1. Introduction

As an important active earth observation technology, lidar (light detection and ranging) can acquire high-precision three-dimensional spatial information, and plays an increasingly important role in agricultural monitoring [1,2], forest management [3–6]. With the continuous expansion and refinement in remote sensing detection fields, more effective and accurate plant detection is becoming more complex and difficult. Due to only a single-wavelength intensity acquisition, the traditional lidar will not only cause poor accuracy of identification and classification in certain complex scenes [7,8], but also influence inversion results of plant physiological and biochemical parameters [9]. Thus, based on the advantage of high-precision spatial information acquisition, it is very necessary and significant to develop new types of lidar technologies to enhance the accuracy and availability on plant detection.

To improve spectral acquisition capability with lidars, researchers have carried out various research attempts. Drawing on the thought of multi-source data fusion, the fusion of hyperspectral images and point cloud can enhance the accuracy for plant detection [10,11], yet several applications are still restricted by data matching accuracy, inconsistency of acquisition time, sun illumination [12]. Therefore, referring to the development process of passive optical remote sensing, which increases the number of detected bands to enhance spectral information, various hyperspectral lidars have been successfully designed and structured by several scientific institutions. Gong et al. [13] employed a supercontinuum laser source and a 32-channel detector to obtain rich reflection information and proposed the utilization of reflection information to classify and derive the nitrogen contents of rice. Sun et al. [14] constructed a hyperspectral full waveform lidar prototype with 32 bands for monitoring the fine structure and biochemical parameters of vegetation. Chen et al. [15] designed a tunable hyperspectral lidar system using an acousto-optic tunable filter (AOTF) to classify four leaves of different plants. Due to the advantages in spatial-spectral information integrated acquisition, hyperspectral lidar gradually becomes one of the most active research hotspots, and has been employed widely in target classification [16,17], plant detection [18,19], true-color 3D reconstruction [20]. Especially in the field of plant detection, hyperspectral lidar has great research potential in the future.

In addition to solar-induced fluorescence, laser-induced fluorescence technology as a new remote sensing technology has been applied to plant detection. Different from the physical mechanism of reflection spectrum, laser-induced fluorescence signals can be excited through laser irradiation, and are frequently used to express physiological biochemical characteristics and earlier present the growing status for different plants [21,22]. Therefore, rely on the laser-induced fluorescence technique, researchers have developed many fluorescence lidar systems. Chappelle et al. [23] used a pulsed N2 laser emitting at 337 nm to excite fluorescence signals for identifying five different plants. Yang et al. [24] proposed using 556 nm laser source to construct fluorescence lidar for plant detection was better than 355 nm laser source. These fluorescence lidar systems mostly take high-energy shortwave single-wavelength laser as laser source to excite fluorescence signals, and further employ high-performance detectors to obtain. Fluorescence lidar has become one of the research hot spots all over the world, and has been applied in many fields of plant detection.

In view of remarkable technical advantages of hyperspectral and fluorescence imaging in plant detection, whether the synchronization acquisition of two types of spectral signals achieved by one device has aroused widespread concern in remote sensing. Ariana et al. [25] designed and constructed a set of multi-spectral imaging system, combining different lighting modes and optical filters, which can achieve visible and near-infrared reflected imaging, ultraviolet and visible induced fluorescence imaging to distinguish the health status of apples. Zhao et al. [26] utilized a mobile laboratory to study maize leaves by reflectance and fluorescence measurements for minoring five different species of maize under different fertilizer supplied, and the results demonstrated that both reflectance and fluorescence information could effectively monitor plants under different fertilizer levels. However, most of these imaging techniques are based on passive imaging systems and cannot achieve an integrated acquisition of spatial information, reflection, and fluorescence signals simultaneously. Therefore, based on the electronic transition theory, using a shortwave laser can excite fluorescence signals at longer spectral locations [27]. Through theoretical analyses, optical filters can be used to filter out some wavelengths of outgoing lasers that are sensitive to vegetation variation, then reflection and fluorescence signals at different wavelength locations can be obtained synchronously. However, how to develop and determine the optimal detection bands combination for monitoring plants was prerequisite and essential for designing and constructing the dual-mechanism lidar system. In general, spectral bands combination was determined by spectral characteristics of spectral data with little regard to hardware system performance, yet hardware performances of lidar units also needed to be taken seriously. In this study, considering the spectral characteristic of reflection and fluorescence signals themselves and the hardware performance of lidar units comprehensively, the optimal bands combination can be determined, especially the selection of laser source and optical filters.

This study proposed a two-step wavelength selection method to determine optimal bands combination for constructing the dual-mechanism lidar system. Selecting the spectral bands combination by this method can not only reduce the complexity and the construction cost of dual-mechanism lidar system, but also improve the accuracy of plant detection. The first step was to select the preliminary bands combination based on spectral characteristic of reflection and fluorescence signals themselves by two types of leaves experiments with the adaptive band selection and random forest methods. Then, further considering the band correlation and hardware performance of lidar units, these selected bands were optimized to determine the optimal detected bands. In order to verify the feasibility and reliability of the two-step wavelength selection method, classification accuracies in an indoor scanning experiment were used as the evaluation criterion. The results demonstrated that the proposed method was effectively to determine the optimal bands combination and could provide a new idea for constructing dual-mechanism lidar system.

2. Experimental design

2.1 Instrument description

The physical mechanisms of spectral signals generation between reflection and laser-induced fluorescence information exist distinct differences. The former obtains the radiation signal reflected by the target surface, while the latter obtains the radiation signal excited by lights illumination. For gaining reflection and laser-induced fluorescence information, these plants were detected twice by a same lidar system. The major difference for twice detections was the selection of laser sources.

For reflection spectrum, a supercontinuum laser source, the wide-band “white” laser, generated from high-nonlinearity photonic crystal fibers, is used to emit wide-spectrum laser pulses, and the spectral range covers from 430 to 2400 nm [28]. While for fluorescence spectrum, a laser source emitting at 355 nm, is used as excitation source to excite fluorescence signals [27]. Two kinds of laser sources are mainstream schemes for constructing hyperspectral and fluorescence laser systems currently. Except for two different laser sources, the other hardware components are consistent. A parabolic mirror with high reflectivity, and a Schmidt–Cassegrain telescope are utilized to ensure the signal quality during spectral information acquirement process. Then, acquired signals are guided to a grating spectrometer by a fiber. To ensure the high-precision signal separation and data acquisition, a high-performance blazed grating, and an intensified charge-coupled device (ICCD) are employed. The spectral range is 160 nm and the spectral resolution is 0.5 nm. Finally, the wider spectral information can be obtained by rotating the central location of blazed grating. The optical setup of the lidar system for gaining reflection and fluorescence information is shown in Fig. 1.

 

Fig. 1. Optical setup of the lidar system for gaining reflection and fluorescence information

Download Full Size | PDF

2.2 Experimental design

To demonstrate the feasibility in category classification and growth condition monitoring of plants with reflection and laser-induced fluorescence information, two indoor experiments, leaves experiment and scanning experiment, are conducted in a dark and clean laboratory to avoid the interference of stray light and other atmospheric factors. Especially for fluorescence signals, it is more affected than reflection signals by external environment. The leaves experiment is mainly used to determine the preliminary bands combination based on spectral characteristics of two kinds of spectral information. While the scanning experiment is to explore the effective of optimal bands combination determined by the two-step wavelength method. In this study, the spectral range of reflection and fluorescence signals are 460-780 nm and 620-780 nm, respectively.

The first experiment composed of a variety of leaves gained from school campuses, was used to determine preliminary spectral bands for categories classification and growth condition monitoring of plants. These leaves were divided into two parts of 10 kinds of healthy leaves in different categories(A-J), and 9 kinds of leaves in different growth condition(a-i). This study randomly selected 4 different locations on the surface of these leaves for reducing accidental errors in signals acquisition process. These leaves were illustrated in Fig. 2.

 

Fig. 2. 10 kinds of healthy leaves in different categories (A) and 9 kinds of leaves in different growth conditions (B)

Download Full Size | PDF

The second scanning experiment was composed of 9 different targets, including 3 kinds of plants in different growth status. These targets were as follows: 1) black paper, 2)black iron, 3) healthy aglaonema, 4) wilted aglaonema, 5) healthy evergreen, 6) yellow evergreen, 7) wilted evergreen, 8) healthy scindapsus, 9) wilted scindapsus. These targets, apart from the black paper, were placed on a small horizontal platform, which placed approximately 5 m away from the supercontinuum laser source. A total of 3114 scanned points was measured. Besides, the category label was manually marked via MATLAB software, and different colors presented different categories. The scanning experiment was illustrated in Fig. 3.

 

Fig. 3. The scanning experiment of 9 different targets

Download Full Size | PDF

3. Methods

3.1 Flowchart of wavelength selection with reflection and fluorescence spectra

The overall flowchart of two-step wavelength selection method consisted of three stages, and the flowchart as shown in Fig. 4.

 

Fig. 4. The flowchart of wavelength selection with reflection and fluorescence spectra

Download Full Size | PDF

Whether for reflection and fluorescence signals, the high-precision intensity calibration was prerequisite to ensure the accuracy of the wavelength selection. For reflection signals, the distance and angle-of-incidence were two essential factors that impacted on high-accuracy record of echo intensities. The distance- and angle-calibrated empirical models were often applied to intensity calibration because the empirical model was more convenient and effective [29]. Besides, the backscatter intensity was normalized by SpectraLon. Spectralon is a fluoropolymer that has the highest diffuse reflection of any known material or coating over the ultraviolet, visible, and near-infrared regions of the spectrum [30]. While for fluorescence signals, we employed a standard light source (Halogen lamp) to gain a standard spectrum to correct fluorescence intensities obtained by a ICCD detector.

Considering the criterions of spectral information amount and band importance, this study adopted two wavelength selection methods, the adaptive band selection and random forest, to select the preliminary bands combination. Then, the optimal bands combination was determined by further considering two factors of band correlation and systematic design. This method can avoid effectively spectral information redundant and provide references for the prototype construction of dual-mechanism lidar system.

3.2 Wavelength selection based on spectral characteristics with reflection and fluorescence spectra

To ensure the reliability of wavelength selection, the results determined by the adaptive band selection and random forest were used to contrast verification. Adaptive band selection is an unsupervised wavelength selection method highlighting the amount of information of bands combination, and has been widely applied in remote sensing fields [31]. Random Forest algorithm is an excellent ensemble learn-based algorithm [32], which is not only used for classification and regression, but also has a good performance for wavelength selection, among which band importance values are an important basis for wavelength selection.

For adaptive band selection, the ABS values are obtained by calculating the standard deviation of each band divided by the average of the sum of the correlation coefficients of adjacent bands. Finally, the spectral position with the largest ABS values is selected as the results for wavelength selection. The formulas are expressed as follows.

Due to the high spectral resolution of 0.5 nm, it is easy to cause similar ABS values in adjacent bands, which further leads to the selected bands with the largest ABS values being at the adjacent location. In order to avoid such case, this study calculates the first order derivative of ABS values to divide the spectral range into several intervals. The spectral location of the largest ABS values within spectral intervals is determined to as the wavelength selection results.

As one of the most commonly machine learning algorithms, random forest has been successfully applied to the study of various subjects due to simple, convenient, and low computational cost [33]. It is a variant of Bagging ensemble learning, constructing a certain amount of decision trees by the bootstrap random resampling technique and node randomly splitting technology, and the final result can be obtained by voting. In this study, the Gini Index of random forest algorithm is used as a wavelength selection measure, which measures the impurity of wavelengths with respect to the classes, and each tree will be split according to a certain node when it grows. The basis of the split can be the reduction of the Gini index before and after the split. The trees start with nodes having the greatest impurity decrease, and end up with nodes having the least impurity decrease. The decreases in impurity from every wavelength are averaged, which are known as the mean decrease in impurity. Improvement in the Gini decrease of each individual attribute for every tree in the forest provides surplus variable importance that is often quite consistent with the permutation importance measure. Then, these bands of higher bands importance are selected.

3.3 Wavelength selection based on the system design of dual-mechanism lidar

Due to the different generation mechanisms of two types of spectral information, spectral bands determined by wavelength selection methods may have some spectral overlaps, which increased the hardware requirement and construction difficulty of dual-mechanism lidar system. Therefore, further considering the band correlation of the selected bands combination and hardware performance of dual-mechanism lidar system comprehensively, the optimal band combination for constructing the dual-mechanism lidar can be determined.

The hardware performance of lidar key components has a great influence on the system construction of dual-mechanism lidar. Specially, the selection of laser source and optical filters is extremely essential. Referring the generation mechanism of laser-induced fluorescence information, which a short-wave laser excites fluorescence signals at long-wave locations, this study plans to utilize a supercontinuum laser source that the emission range of 430-2400 nm as the system source. This source can be used to generate reflection information, and excite fluorescence signals from plants synchronously. Besides, based on the advantage of narrow-band optical notch filters reflecting lasers at specific wavelengths and transmitting lasers at other wavelengths, this study will utilize these optical filters to filter lasers at selected fluorescence wavelength positions for gaining laser-induced fluorescence signals. These optical filters own a relatively large bandwidth about 40 nm, which is a notable factor affecting the results of wavelength selection. Therefore, on the premise of quantitatively considering the hardware performance of optical filters, we calculate the spectral band correlation of two kinds of spectra, then set the correlation threshold parameters to optimize the selected bands, and to ensure the availability and rationality for wavelength design of the dual-mechanism lidar system. The brief design diagram of dual-mechanism lidar system was as shown in Fig. 5. A supercontinuum source was used as emitting source and two narrow band optical notch filters (L2 and L3) were used to filter lasers at selected fluorescence wavelength positions. Subsequently, two kinds of spectral signals of reflection and fluorescence at different spectral locations were gained by signal collection, optical splitting, signal detection and acquirement.

 

Fig. 5. The brief design diagram of dual-mechanism lidar system

Download Full Size | PDF

3.4 Target classification and accuracy evaluation

To further validate the availability of the proposed method, we take the classification accuracy of a scanning experiment as assessment criteria. As mentioned in section 3.2, random forest algorithm can also be as an outstanding classification method because it can provide a high classification accuracy without overfitting, and has been applied in target classification widely. Some researchers compared 179 algorithm performances through a large number of experiments, and the results indicated that the classification performance of the random forest algorithm was excellent [34]. In addition, it could also achieve better classification results in the case of small datasets and imbalanced categories, so this algorithm was very suitable for this classification experiment.

For the random forest algorithm, the number of decision trees and the maximum of features were two essential parameters which affected the classification accuracy. After parameter optimization by grid search with cross validation, the number of decision trees was set to 80, and the maximum of features was set to 2. Due to the dataset small, a three-fold cross validation method was employed to ensure the classification accuracy. Besides, we employed three classification evaluation criteria of overall accuracy, average accuracy, Kappa coefficient to evaluate the classification results. The corresponding formula as follows:

4. Results and discussions

4.1 Wavelength selection results based on spectral characteristics with reflection and fluorescence spectra

The reflectance and fluorescence intensity of 10 healthy leaves of different categories and 9 leaves of different growth conditions were shown in Fig. 6. The spectral range of reflectance and fluorescence intensity were 460-780 nm and 620-780 nm, respectively. Figure 6(1) and Fig. 6(3) represented the reflectance by smooth denoising in 10 healthy leaves of different categories and 9 leaves of different growth conditions respectively. Compared with the reflectance of healthy leaves, there were larger differences in different growth conditions. Reflectance curves of wilted leaves basically got up in a straight line. Likewise, Fig. 6(2) and Fig. 6(4) represented the fluorescence intensity in these leaves respectively. Fluorescence signals of wilted leaves can hardly find it. Finally, the preliminary bands combination of two kinds of spectra was determined by the adaptive band selection and random forest.

 

Fig. 6. Reflectance and fluorescence intensity of 10 healthy leaves of different categories and 9 leaves of different growth conditions. (1): reflectance of 10 healthy leaves of different categories; (2) fluorescence intensity of 10 healthy leaves of different categories; (3) reflectance of 9 leaves of different growth conditions. (4): fluorescence intensity of 9 leaves of different growth conditions.

Download Full Size | PDF

Based on two spectra of reflection and fluorescence on 10 healthy leaves of different categories and 9 leaves of different growth conditions, the ABS values on different wavelength location were calculated. Figure 7 illustrated Adaptive values with adaptive band selection on two types of leaves experiments.

 

Fig. 7. Adaptive values of 10 healthy leaves of different categories and 9 leaves of different growth conditions. (a): adaptive values of reflection spectrum of 10 healthy leaves of different categories; (b): adaptive values of fluorescence spectrum of 10 healthy leaves of different categories; (c): adaptive values of reflection spectrum of 9 leaves of different growth conditions; (d): adaptive values of fluorescence spectrum of 9 leaves of different growth conditions.

Download Full Size | PDF

The band combination was determined by selecting the maximum adaptive values. However, to avoid the result falling into local optimums, we divided the spectral range based on the first derivative results on reflection and fluorescence signals. The spectral position from negative to positive in the first derivative was used as the separation point to divide the whole spectral range. Figure 8 presented the results of first derivative of reflection and fluorescence information on two types of leaves experiments. According to the location of dividing points, (e), (f), (g) and (h) were divided into 4, 2, 2 and 2 spectral intervals respectively. These spectral intervals were (e):460-526 nm, 526-637 nm, 637-723 nm,723-780 nm; (f):620-717 nm, 717nm-780 nm; (g):460-675 nm, 675-780 nm; (h):620-721 nm, 721-780 nm. The spectral bands corresponding to the maximum adaptive values in respective spectral intervals were determined, thus these spectral bands can be determined: (a): 481 nm, 575 nm, 711.5 nm, 758.8 nm; (b): 686.5 nm, 737 nm; (c): 542.5 nm, 705 nm, 780 nm; (d): 687 nm, 736 nm. Because the category and growing status of two leaves experiments different, there were some discrepancies in the wavelength selection results of reflection information, while for fluorescence information, two spectral bands near 687 nm and 737 nm determined by wavelength selection were relatively accurate.

 

Fig. 8. First derivative values of adaptive values on two types of leaves experiments. (e): first derivative values of reflection spectrum of 10 healthy leaves of different categories; (f): first derivative values of fluorescence spectrum of 10 healthy leaves of different categories; (g): first derivative values of reflection spectrum of 9 leaves of different growth conditions; (h): first derivative values of fluorescence spectrum of 9 leaves of different growth conditions.

Download Full Size | PDF

As a supervised method of wavelength selection, random forest selected these bands with more importance in classification process to determine bands combination. Different leaves samples were manually labeled with different categories to select appropriate spectral bands. Figure 9 illustrated the results of spectral band importance values on two types of leaves experiments. From the band importance values distribution, these spectral bands were selected and as follows: (i): 480.5 nm, 541 nm, 685.5 nm, 775.5 nm; (j): 687.5 nm, 738 nm; (k): 541 nm, 713 nm, 780 nm; (l): 686.5 nm, 737.5 nm.

 

Fig. 9. Spectral band importance results of reflection and fluorescence spectra on two types of leaves experiments. (i): band importance of reflection spectrum of 10 healthy leaves of different categories; (j): band importance of fluorescence spectrum of 10 healthy leaves of different categories; (k): band importance of reflection spectrum of 9 leaves of different growth conditions; (l): band importance of fluorescence spectrum of 9 leaves of different growth conditions.

Download Full Size | PDF

Compared with the above two wavelength selection results, the spectral location of fluorescence bands was almost perfectly consistent, which further indicated that the two fluorescence bands belonged to the characteristic bands of plant detection. For the selected reflection bands, there were some minor differences because of the principle of two methods selecting spectral bands different. The former highlighted the amount of information of bands combination, while the latter emphasized the importance of selected bands. Therefore, it was necessary to consider comprehensively two wavelength selection results to determine the optimal bands combination.

4.2 Wavelength selection results based on system design of dual-mechanism lidar

To acquire reflection and fluorescence signals synchronously, the spectral bands should be selected as much as possible, where these bands did not have overlaps in two spectral space and were far away from each other. Therefore, on the basis of the 4.1 wavelength selection results, these selected bands could be optimized to decrease spectral overlap by considering the band correlation and system design. To optimize the selected bands and eliminate these bands with higher band correlation, band correlation principle was utilized by Pearson correlation method. Figure 10 showed the band correlation results calculated by Pearson correlation method on two types of leaves experiments.

 

Fig. 10. Band correlation results with reflection and fluorescence spectra on two types of leaves experiments. (m): band correlation of reflectance spectrum of 10 healthy leaves of different categories; (n): band correlation of fluorescence spectrum of 10 healthy leaves of different categories; (o): band correlation of reflection spectrum of 9 leaves of different growth conditions; (p): band correlation of fluorescence spectrum of 9 leaves of different growth conditions.

Download Full Size | PDF

The band correlation at closer spectral positions was higher. The correlation threshold of 0.93 was taken as a distinction of strong correlation of spectral bands. For the first leaves experiment, band correlations of selected reflection bands were as follow: 480.5 nm-481nm: 1; 541nm-575nm: 0.9385; 685.5nm-711.5nm: 0.6756; 758.5nm-775.5nm: 0.994. The band correlations of fluorescence bands were as follow: 686.5nm-687.5nm: 1; 737nm-738nm: 1. Similarly, for the second leaves experiment, band correlations of selected reflection bands were as follow: 541nm-542.5nm: 1; 705nm-713nm: 0.9777; 780nm-780nm: 1. The band correlations of fluorescence bands were as follow:686.5nm-687.5nm: 0.9999; 737nm-738nm: 1. Considering band correlation results comprehensively, reflection bands of 481 nm, 541 nm, 685.5 nm, 711.5 nm, 775.5 nm and fluorescence bands of 686.5 nm and 737 nm were determined.

Based on the hardware design of dual-mechanism lidar, the two spectral locations of reflection band 685.5 nm and fluorescence band 686.5 nm were exceptionally close, and the existing photoelectric detection technology cannot achieve accurate separation and acquirement. Besides, to achieve better fluorescence signals acquirement, the dual-mechanism lidar system was planned to use narrow band optical notch filters to filter reflection signals near at 686.5 nm, and further obtain fluorescence signals at same location. Likewise, the fluorescence signal at 737 nm was obtained by the same method. Given all that, the optimal bands combination of reflection bands of 481 nm, 541 nm, 711.5 nm, 775.5 nm and fluorescence bands of 686.5 nm, 737 nm was determined as the detection bands for the dual-mechanism lidar system.

4.3 Classification result

To validate the availability and reliability of optimal bands combination determined by the two-step wavelength selection method, five different target classification strategies (I-V) were constructed and the classification accuracy of a scanning experiment was as the evaluation criteria. Table 1 showed the five different target classification strategies with different bands combinations. Table 2 presented the classification results calculated by random forest classification algorithm on five different target classification strategies. The overall accuracy (OA), average accuracy (AA), Kappa coefficient, precision rate (P) and recall rate (R) were used to evaluate the classification results respectively. Predicted classification results of 9 different targets were shown in Fig. 9, and the green cloud point indicated the misclassified targets.

Table 1. five different target classification strategies

View Table | View all tables in this article

Table 2. Classification accuracy summary

View Table | View all tables in this article

In the first classification strategy I, overall accuracy, average accuracy, and Kappa coefficient were 89.31%, 71.29% and 0.8210 respectively, and the predicted results as shown in Fig. 11(A). Most of the point clouds could be correctly classified, yet due to the presence of salt and pepper noise and the complex spatial structure, the correct classification accuracy of certain categories was lower, less than 0.6. In addition, the classification accuracy was slightly reduced by using 4 reflection bands selected from 640 reflection bands, yet this classification result indicated the effectiveness of wavelength selection. The overall accuracy, average accuracy and Kappa coefficient were 88.18%, 65.89% and 0.8013 respectively, and the predicted results as shown in Fig. 11(C).

 

Fig. 11. Classification results of five different target classification strategies, and different colors represent different targets. (A): 640 reflection bands; (B): 320 fluorescence bands; (C): 4 reflection bands; (D): 2 fluorescence bands; (E): 4 reflection bands and 2 fluorescence bands.

Download Full Size | PDF

Compare with the classification results of reflection information, the classification accuracy using 320 fluorescence bands of strategy II was not relatively high, and the overall accuracy, average accuracy, and Kappa coefficient were 78.61%, 53.99% and 0.6343 respectively, and the predicted results as shown in Fig. 11(B). The precision rate of two-thirds categories was less than 0.6, and even the third class of wilted aglaonema was only 0.15. Similarly, the classification accuracy by using 2 fluorescence bands selected from 320 fluorescence bands was also relatively poor, and the overall accuracy, average accuracy and Kappa coefficient were 75.24%, 44.63% and 0.5828 respectively, and the predicted results as shown in Fig. 11(D).

Compared to strategy III (4 reflection bands) and strategy IV (2 fluorescence bands), the fifth strategy of combining 4 reflection bands and 2 fluorescence bands got the best classification results. The overall accuracy and average accuracy were improved by 2.51%, 15.45%, and 7.8%, 29.06% respectively. However, especial the third class of wilted aglaonema and the ninth class of wilted Scindapsus, the precision rate still was low, less than 0.6. From Table 2, this phenomenon also existed in the other four classification strategies. The specific reasons were analyzed as follows: (1) the spectral signals were greatly affected by surface geometry factors because of the spatial structure more complex; (2) the spectral signals were inaccurate when the laser irradiated the edge of targets or only part of laser irradiated to targets.

From classification results of different classification strategies, the band combination, combining 4 reflection bands of 481 nm, 541 nm, 711.5 nm, 775.5 nm and 2 fluorescence bands of 686.5 nm, 737 nm, could obtain the best classification accuracy. Specially, the misclassification phenomenon of certain plant categories can greatly improve by adding fluorescence signals on the basis of reflection spectrum. This classification results also demonstrated the availability of the proposed wavelength selection methods, and provided the references for constructing dual-mechanism lidar system.

5. Conclusions

As important spectral properties of plants, reflection, and fluorescence spectra have great research potential and commercial values in plant detection. To make full use of the spectral advantages of reflection and fluorescence spectra, a concept of dual-mechanism lidar system that can obtain spatial information, reflection, and fluorescence signals simultaneously is raised. In order to better design and construct the dual-mechanism lidar system, determining an appropriate spectral bands combination is prerequisite. Therefore, this study proposed a two-step wavelength selection method to determine optimal bands combination by considering the spectral characteristic of reflection and fluorescence signals themselves, and the hardware performance of lidar units comprehensively. The optimal band combination of 4 reflection bands of 481 nm, 541 nm, 711.5 nm, 775.5 nm, and 2 fluorescence bands of 686.5 nm, 737 nm was determined. Besides, compared to a single type of spectral information, the optimal bands combination got the best classification results in an indoor scanning experiment. Overall accuracy and average accuracy were improved by 2.51%, 15.45%, and 7.8%, 29.06% respectively. Increasing the fluorescence information can improve the misclassification phenomenon of certain plant categories. The results demonstrated adequately the reliability of two-step wavelength selection method, and provided references for the development of dual-mechanism lidar system.