Feature Extraction on the Difference of Plant Stem Structure Based on Ultrasound Energy

Abstract

Plant growth is closely related to the structure of its stem. The ultrasonic echo signal of the plant stem carries much information on the stem structure, providing an effective means for analyzing stem structure characteristics. In this paper, we proposed to extract energy features of ultrasonic echo signals to study the structure of the plant stem. Firstly, it is found that there are obvious different ultrasonic energy changes in different kinds of plant stems whether in the time domain or the frequency domain. Then, we proposed a feature extraction method, density energy feature, to better depict the interspecific differences of the plant stems. In order to evaluate the extracted 24-dimensional features of the ultrasound, the information gain method and correlation evaluation method were adopted to compute their contributions. The results showed that the mean density, an improved feature, was the most significant contributing feature in the four living plant stems. Finally, the top three features in the feature contribution were selected, and each two of them composed as 2-D feature maps, which have significant differentiation of the stem species, especially for grass and wood stems. The above research shows that the ultrasonic energy features of plant stems can provide a new perspective for the study of distinguishing the structural differences among plant stem species.

1. Introduction

The study of variability in the internal structure of plant stems is one aspect of research for exploring the physiological functions of plant stems [1]. The marked interspecific differences in the structure of plant stems lead to variations in the distribution and proportion of the composition of the stem. The plant stem body is organized, from the outside to the inside, by bark, formative layer, xylem, and pith. The differences in stem cells lead to different stem characteristics. Due to many lignified cells, woody stems have a firm and tall stem texture. Herbaceous stems have a less developed or undeveloped xylem, so the stem is juicy and soft in texture. At the same time, as the growth of the internal structure of the stem is a dynamic and organic change process [2], the search for non-destructive online stem-condition detection has received increasing attention from researchers [3,4]. In recent years, researchers have initiated a large number of studies on dynamic online plant-stem growth detection and the proposal of new detection methods [5,6]: the γ-ray method [7], nuclear-magnetic-resonance method [8,9,10], X-ray computed tomography [11], resistance method [12], time-domain reflection method TDR [13], frequency-domain capacitance method [14,15] and ultrasonic detection [16,17] methods.

Because of ultrasound’s accuracy and nondestructive characteristics in detecting the internal structure and properties of substances, it is widely used in medical and industrial testing. In biomedicine, ultrasound on different tissues and organs of the human body with different reflection and absorption characteristics can effectively detect the structural properties of tissues and organs and their dynamic characteristics. At present, it can effectively detect the thickness of the fat layer [18,19], masses [20,21], the shape and flow rate of human blood vessels [22], and the water content when water accumulation occurs in organs [23].

Ultrasonic testing has been widely used in inspecting wood since it was found that the velocity of low-frequency waves in the longitudinal direction of wood can effectively measure wood strength property indexes such as the smooth grain flexural modulus and the flexural strength [24]. Researchers have found that the ultrasonic velocity decreases with the increasing moisture content of wood [25,26]. In detecting wood by ultrasound, its axial ultrasound signal has the most negligible energy attenuation and the fastest ultrasound echo velocity, while the radial attenuation is the largest and the ultrasound echo velocity is the most minor [27,28]. Recent studies have pointed out that the ultrasonic attenuation characteristics of wood are closely related to the structure of the plant stem tissue [26]: because the plant stem bears the transport of water and organic matter in the axial direction, forming a direct connection or perforated connection between the axial cell tissues, so the ultrasonic impedance is small, while complete cell walls separate the radial cell tissues, so the ultrasonic impedance is large [28,29]. Therefore, it is important to extract features, which reflect the attenuation of the ultrasound energy in plant stems, to discover the performance of living plant stems. However, there is little research on this aspect.

At present, the study of plant stem structure difference by ultrasound still uses ultrasonic velocity. The freezing condition of plant stems was first studied by Guillaume in 2014 using the ultrasound echo signals [30], which detected the process of freezing and thawing of sampled plant stems by the velocity of the ultrasound echo signal. However, researchers still only adopted the features of the ultrasound echo velocity and energy decay in the time domain to study the freeze–thaw phenomenon of plant stems. Since the experiments were performed with intercepted stem samples, it was impossible to study the stress response of living stems after freezing. At the same time, because the ultrasound signals in the experiments were not ultrasonic radio frequency (RF) signals, more feature extraction could not be achieved to study the differences in living plant stems on ultrasonic. Lv [31] used the ultrasound velocity to track the moisture changes of living stem plants. They adopted velocity to analyze the change of plant stems, which assumed that plant stem structure is uniform. However, the truth is that because the plant stem is non-uniform and anisotropic, it is difficult for the velocity to reflect such internal structural differences. Therefore, it is necessary to extract new features that can reflect the plant stem structure other than velocity.

It is possible to explore the sensitivity of the ultrasound propagation to the density and structure of plant stems by ultrasound RF [32]. In this paper, we focused on extracting the ultrasound energy features from the ultrasound RF in the time and frequency domain to discover the difference of plant stem structure. What is more, the evaluation of the feature contribution was ranked by information gain and correlation.

2. Methods

2.1. Common Energy Features of Ultrasound in Plant Stems

2.1.1. Mean and Variance of Ultrasound Echo Signals

The change in the ultrasound echo signal’s mean value reflects the plant stem conditions. As the ultrasound propagates through the plant stem, its energy is absorbed, attenuated, and scattered by the material in the stem, resulting in an exponential attenuation of the echo signal envelope. Different plant stems have different attenuation exponents, which ultimately lead to variations in the mean and variance of the ultrasound echoes. In this regard, the expression for calculating the mean value of the ultrasonic echo is presented in Equation (1).

$${\mu }_x=E\left\{x\right\}=\underset{\mathrm{T}\to \infty }{\lim }\frac{1}{T}{{\displaystyle \int }}_0^{T}x\left(t\right)dt\approx \frac{1}{N}{\displaystyle \sum }_{n=1}^{N}x\left[n\right]$$

(1)

where $x\left(t\right)$ is the ultrasonic echo continuous signal, and $x\left[n\right]$ is the ultrasound echo signal sequence values. $T$ is the observation time of the ultrasound echoes. $n, n=1,2,\dots ,N$ is the sampling point for the ultrasound echoes.

The ultrasonic echo signal variance reflects the fluctuating characteristics of the signal above and below the average ultrasonic energy, reflecting the changing energy characteristics of the ultrasonic stem signal. The mathematical expression for calculating the variance of an ultrasonic echo sampling signal is

$${\sigma }_x^2=E\{{\left|x\left(n\right)-{\mu }_x\right|}^2\}$$

(2)

2.1.2. Spectral Features

The propagation of the ultrasound waves is closely related to the density and structure of the material. Hence, when the structure of the plant stem changes, it causes changes in the spectrum of the ultrasonic echo signal.

The FFT is used to get the spectrum $X\left(k\right)$ of the ultrasound echo sequence $x\left(n\right)$. The energy in the frequency domain is expressed in Equation (3).

$$G_x\left(k\right)=\frac{1}{N}{\left|X\left(k\right)\right|}^2$$

(3)

where $X\left(k\right)$ is the FFT sequence value of the ultrasound echo, and $k$ represents the kth spectral line. The relationship between spectral line k and each harmonic frequency of the signal is

$$f=\frac{k}{N{T}_s}$$

(4)

where $f$ is the sampling frequency, which is 5–7 times the center frequency of the ultrasonic probe, and $T_s$ is the sampling interval.

2.2. Improvement of Ultrasound Features

The traditional ultrasonic feature is the ultrasonic velocity computed by the envelope of the ultrasonic RF signal, which is the signal retained after removing the high-frequency components in the FR ultrasonic. It is easy to identify the difference of the structure by the different velocity in the uniform medium. However, because the plant stem is non-uniform and anisotropic, the traditional feature velocity has limitations in reflecting structural differences. Therefore, we tried to extract new features based on the ultrasonic RF signal, carrying the inhomogeneous and anisotropic structural characters.

When ultrasonic energy is used as features, the processing of the detection depth is a key problem. Different plant diameters at height can cause various interference with ultrasound detection. This interference is manifested in the following ways: the internal structure of the same plant is similar, and its difference is mainly due to the different detection positions, forming different detection plant stem diameters, which cause different ultrasound energy parameters, such as mean and variance. While, if the plant diameters are the same, the different parameters of the ultrasound would show the difference of the plant stem structure. In live plant stem detection, at the radial direction, the stem diameter at the detection site is the detection depth of the ultrasound. As the structures of the plant stems at breast height can vary significantly depending on the species and growing period, this causes the difference in the depth of the ultrasound detection and the attenuation of the ultrasound. Suppose the depth is determined and is considered constant during the detection period, such as in a day. In this case, changes in the ultrasonic parameters can be caused by changes in the plant structure, such as water deficiency or lack of water. At the same time, to reduce the interference caused by the different detection depths, this paper proposes to use the thoracic diameter parameter to obtain the feature density of the ultrasound energy to suppress or eliminate the interference of the thoracic diameter on the premise of obtaining the standard ultrasound features.

$$de{n}_f=\frac{f_{eature}}{d}$$

(5)

where $de{n}_f$ is the density of the ultrasonic feature, $f_{eature}$ is the ultrasonic feature, and $d$ is the diameter at breast height at the plant detection.

Based on the above method, the improved features in this paper are extracted, which are energy density, mean density, variance density, and first to third resonance peak amplitude density.

The equations of all the extracted features are shown in

Table 1. Equations of the extracted features.

No.	Feature Name	Equation
1	Energy	$E={\displaystyle {\displaystyle \sum }_{i=1}^N}\left|x_i\right|, i=1\cdots N$
2	Energy density	$de{n}_E=\frac{E}{d}$
3	Mean value	$M=mean\left(x_i\right), i=1\cdots N$
4	Mean density	$de{n}_M=\frac{M}{d}$
5	Variance	$v_{ar}=\mathit{var}\left(x_i\right),i=1\dots .N$
6	peak-to-peak value	$peak=\left|\mathit{max}\left(x\right)-\mathit{min}\left(x\right)\right|$
7	Spectrum DC	$S_{DC}=Z_0,Z=abs\left(fft\left(x_i\right)\right),i=1,\dots ,N; Z=Z_0,\dots ,Z_{N-1}$
8	Spectral DC density	$de{n}_{S_{DC}}=\frac{Z_0}{d}$
9	1st formant value	$forman{t}_1=max\left(Z\right)$
10	1st formant density	$de{n}_{forman{t}_1}=\frac{forman{t}_1}{d}$
11	1st formant frequency	$Inde{x}_{forman{t}_1}=index\left(max\left(Z\right)\right)$
12	2nd formant value	$forman{t}_2=secon{d}_{max\left(Z\right)}$
13	2nd formant density	$de{n}_{formant2}=\frac{forman{t}_2}{d}$
14	2nd formant frequency	$Inde{x}_{forman{t}_2}=index\left(secon{d}_{max\left(Z\right)}\right)$
15	3rd formant value	$forman{t}_3=thir{d}_{max\left(Z\right)}$
16	3rd formant frequency	$Inde{x}_{forman{t}_3}=index\left(thir{d}_{max\left(Z\right)}\right)$
17	removed DC energy	$E_{-DC}=E-mean\left(x\right),x=x_1,\dots ,x_N$
18	removed DC variance	$Va{r}_{E_{DC}}=var\left(E_{-DC}\right)$
19	removed DC 1st formant value	$forman{t}_1=max\left(Z_{-DC}\right),Z_{-DC}=Z_1,\dots ,Z_{N-1}$
20	removed DC 1st formant frequency	$Inde{x}_{forman{t}_1}=index\left(max\left(Z_{-DC}\right)\right)$
21	removed DC 2nd formant value	$forman{t}_2=secon{d}_{max\left(Z_{-DC}\right)}$
22	De-rectified 2nd formant frequency	$Inde{x}_{forman{t}_2}=index\left(secon{d}_{max\left(Z_{-DC}\right)}\right)$
23	removed DC 3rd formant value	$forman{t}_3=thir{d}_{max\left(Z_{-DC}\right)}$
24	removed DC 3rd formant frequency	$Inde{x}_{forman{t}_3}=index\left(thir{d}_{max\left(Z_{-DC}\right)}\right)$

.

2.3. The Evaluttion of the Contribution of the Feature

In order to achieve a multi-perspective study of the energy features of the ultrasound stem differences, it is necessary to evaluate the most discriminating features from the multidimensional ultrasound energy features. When multiple parametric features are extracted from the time and frequency domains, compelling ultrasound features or feature sets are first selected so that they can effectively reflect the structural differences in the plant stems. Tracking the structural properties within the stem body is also a crucial issue in feature selection. Therefore, this paper uses the information gain and correlation evaluation method to rank the ultrasound energy features of the plant stems by their contribution.

2.3.1. Information Gain

Information gain as shown in Equation (6) indicates the extent to which the information of a known feature vector A reduces the information uncertainty of the partitioned dataset D [33].

$$\mathrm{Gain}\left(D,A\right)=H\left(D\right)-H\left(D|A\right)$$

(6)

$H\left(D\right)$ is the information entropy of the dataset $D$. $H\left(D|A\right)$ is the conditional entropy of $D$ given feature $A$. $A$ is the feature vector, and $D$ is the dataset. $H\left(D\right)$ is defined as shown in Equation (7).

$$H\left(D\right)=-{{\displaystyle \sum }}_{j=1}^M\frac{\left|C_j\right|}{\left|D\right|}lo{g}_2\frac{\left|C_j\right|}{\left|D\right|}$$

(7)

where $M$ is the number of categories $C$ in dataset $D$. $\left|C_j\right|$ is the number of samples of category $j$. $\left|D\right|$ is the total number of samples of dataset $D$.

For a given dataset $D$, different features have different information gains. The greater the information gain, the more information the feature contains that helps in classification and the stronger the classification ability. The method of feature contribution is based on information gain. For a given dataset $D$, the information gain of each feature $A$ is calculated, and the feature with the highest information gain is selected.

2.3.2. Evaluation of Correlation

Correlation is mainly used to assess the contribution of features to classification by measuring the correlation between feature vector $A$ and category vector $C$ [34].

The correlation coefficient $r$ is calculated as in Equation (8).

$$r=\frac{\sum AC-\frac{\sum A\sum C}{\left|D\right|}}{\sqrt{(\sum A^2-\frac{{(\sum A)}^2}{\left|D\right|})(\sum C^2-\frac{{(\sum C)}^2}{\left|D\right|})}}$$

(8)

where $C$ is the category vector, and where the dataset $D$ is related to the sample category vector $C$ as in Equation (9).

$${\displaystyle \sum }_{j=1}^M|C_j|=\left|D\right|$$

(9)

where $C_j$ is a vector of $M$ categories, $j=1,2,\dots ,M.$

The correlation coefficient indicates the degree of linear correlation between the two variables, feature vector $A$ and category $C$. When $0<\left|r\right|<1$, it indicates a certain degree of linear correlation between the two variables. Moreover, the closer $\left|r\right|$ is to 1, the closer the linear relationship between the two variables. The closer $\left|r\right|$ is to 0, the weaker the linear correlation between the two variables.

In this paper, the multiple stem ultrasound energy features were extracted based on time-domain acoustic and frequency-domain amplitude-frequency signals, combined into a multidimensional set of plant stem ultrasound energy features. The optimal ultrasound energy features or feature sets were selected to represent different stem structure differences using the information gain attribute evaluation and correlation evaluation methods.

3. Results

3.1. Simulation Experiment

The ultrasonic plant stem detection system is the same as the paper [31] shown in Figure 1. The system includes a non-metallic ultrasonic probe, ultrasonic pulse signal generator and receiver (CTS-8077PR). It also includes a Nextkit data acquisition instrument and computer. The acquisition function of Nextkit is developed on labView.

(1) CTS-8077PR parameters: pulse repetition frequency PRF is set to 1 kHz to get 1000 ultrasonic RF echo signals in each second, the pulse width is 1000 ns, and the conversion voltage is 200 v.

(2) Ultrasonic probe parameters: probe diameter is 30 mm, and the primary frequency is 1 MHz, which is suitable for ultrasonic testing of nonmetals.

(3) Data acquisition equipment parameters: sampling depth is 2000 points to record each ultrasonic RF echo signal, and the sampling frequency is 10 MHz to get more details of higher frequency of the signal.

Radial ultrasonic testing was conducted on four healthy plant stems: basho, palm, magnolia denudata, and potted sunflowers shown in Figure 2.

Figure 1. The ultrasonic plant stem detection system.

Figure 1. The ultrasonic plant stem detection system.

Figure 2. Testing samples. (a) basho; (b) palm; (c) sunflower; (d) magnolia denudata.

Figure 2. Testing samples. (a) basho; (b) palm; (c) sunflower; (d) magnolia denudata.

3.2. Different Ultrasonic Manifestations of Stem and Body Structure of Different Tree Species

The structural characteristics of the stem bodies of these four types of plants are shown in

Table 2. Attributes of plant stems.

Varieties of Trees	Stem Texture	Growth Characteristics	Plant Category	Vascular Properties
Basho	Herbaceous stems	Perennial	Monocotyledon	Vascular bundles dispersed within the stem, no formative layer within bundles
Sunflower	Herbaceous stems	Annual	Dicotyledon	Vascular bundles forming a circular shape, without a forming layer within the bundle.
Magnolia denudata	Woody stem	Deciduous plant	Dicotyledon	Vascular bundles forming a circular shape, with a formative layer within the bundle.
palm	Woody stem	Casuarina	Monocotyledon	Vascular bundles scattered within stem, with forming layer within bundle.

. The testing took place on 9 May 2019 from 5:00 p.m. to 5:30 p.m.

The plant stem body ultrasound signal acquisition was completed using the experimental platform in the literature [31]. Assay parameters are shown in

Table 3. Parameters of test data.

Samples	Basho 1	Sunflower 1	Sunflower 2	Palm 1	Palm 2	Magnolia Denudata 1	Magnolia Denudata 2
Height of the test point from the ground (cm)	48.00	20.00	20.00	57.00	79.00	75.00	80.00
Test circumference (cm)	20.00	5.00	7.00	49.00	50.00	46.30	63.20
Samples of ultrasound pulse echoes (pcs)	57	33	42	61	53	60	48

.

3.2.1. Ultrasonic Signals of Plant Stems in Time and Frequency Domain

Ultrasound Signal in Time Domain

The ultrasound signals of the four samples in the time domain are shown in Figure 3.

The ultrasonic echo signals from the stem bodies of the four plant species in Figure 3 showed an obvious difference in the envelope of their signals. The peaks of the signals of the sunflowers ranged from −2 to 2 v, while the peaks of the other three species ranged from −4 to 4 v. Moreover, waveforms of these four plants varied in the time domain. It showed that the signals displayed fluctuation at about 20 ns for the stems of magnolia denudata and palm, which was caused by the bark of these two kinds of stems. Because the texture structure of the bark has some extensive differences with the xylem of these stems, the waveform energy of ultrasonic RF signal has obvious increase, rather than constant attenuation. While this phenomenon did not appear in the stems of sunflower and basho, because these two kinds of stems have no obvious barks. However, it is a challenge to describe the difference in the time domain.

Frequency Domain Ultrasound Echo Amplitude and Frequency Signals

In order to further analyze the energy distribution characteristics of the ultrasonic wave of the four plant species stems, the experiments used the Fourier transform to obtain the amplitude-frequency signals. In the experiment, the sampling frequency was 10 MHz, and the spectrum could be analyzed from 0 to 5 MHz. The experimental results are shown in Figure 4.

Figure 4 shows that the energy distribution of the four plants was prominent at 1 MHz and 3 MHz, but their spectral envelopes had apparent differences.

Figure 5 segments the spectrum to show the resonance peaks are mainly distributed near 1 MHz and 3 MHz.

The results showed that the stems of the monocotyledon (basho and palm) showed a transparent envelope of single peaks around the first resonance frequency (1 MHz). In contrast, the dicotyledon (sunflower and magnolia) showed a clear multi-peak feature. Comparing the energy peaks around 1 MHz and 3 MHz, the peak energy distribution of the monocotyledonous stems around 3 MHz was less than 4 to 8 times that around 1 MHz. In contrast, the distribution of the peak energy at the two frequencies for the dicotyledon sunflower and magnolia was closer, between 1 and 2 times, as shown in

Table 4. Energy distribution of resonance peak in ultrasonic spectrum of plant stems.

Name of Tree	Envelope around 1 M.	Envelope around 3 M.	P1: Peak Value around 1 M (mv).	P3: Peak Value around 3 M (mv).	Times of P1/P3
Basho	Single peak	Multi-peak	32.00	4.20	8.00
Sunflower	Multi-peak	Twin peaks	2.60	2.30	1.00
Magnolia denudata	Multi-peak	Single peak	20.10	8.70	2.00
Palm	Single peak	Multi-peak	28.80	6.80	4.00

.

3.2.2. Ultrasonic Energy Features and Feature Contribution of Plant Stems

(1)

Energy features in the time domain

The energy features of the radial ultrasound of the stems of the above four plant species were extracted in the time domain, as shown in

Table 5. Ultrasound energy features of plant stems in time domain.

Name of Sample	Circumference (cm)	Energy (v)	Energy Density (v/cm·10−4)	Average Value (v)	Average Density (v/cm)	Variance	Peak-to-Peak Value (v)	Removed DC Energy (v)	Removed DC Variance
Basho 1	20.00	31.57	5.44	0.11	0.16	1.43	7.86	31.54	1.43
Sunflower 1	5.00	21.64	1.75	0.01	0.43	0.16	5.41	21.63	0.16
Sunflower 2	7.00	21.90	3.61	0.03	0.31	0.16	5.56	21.88	0.16
Magnolia denudata 1	63.50	31.76	0.74	0.05	0.05	1.51	7.84	31.76	1.51
Magnolia denudata 2	46.50	32.23	1.50	0.07	0.07	1.68	7.82	32.22	1.68
Palm 1	49.50	31.22	1.97	0.10	0.06	1.32	7.83	31.19	1.32
Palm 2	50.50	31.31	1.89	0.10	0.06	1.34	7.83	31.29	1.34

. The main features are conventional energy features as well as improved energy features. The conventional energy features mainly contained total energy, mean value, variance, and peak-to-peak values. The improved energy features were mainly the energy density and the mean density.

(2)

Energy features in the time domain

The frequency-amplitude signals based on the Fourier transform were used to extract the radial ultrasound frequency energy features in the frequency domain, as shown in

Table 6. Ultrasound features of plant stems in frequency domain.

Name of Sample	Basho 1	Sunflower 1	Sunflower 2	Magnolia Denudata 1	Magnolia Denudata 2	Palm 1	Palm 2
Average (dB)	125.65	8.78	27.08	28.73	29.95	70.37	109.92
Average density (dB/cm)	0.63	0.18	0.39	0.04	0.05	0.15	0.22
1st formant value (dB)	254.02	24.86	28.94	243.92	246.71	269.51	215.13
1st formant value density (dB/cm)	1.27	0.50	0.41	0.36	0.39	0.58	0.43
sampling position of 1st formant	184.00	186.00	191.00	184.00	185.00	183.00	183.00
1st formant frequency (MHz)	0.92	0.93	0.95	0.92	0.92	0.92	0.91
2nd formant value (dB)	83.52	25.82	34.43	83.52	84.99	97.19	75.24
2nd formant density (dB/cm)	0.42	0.52	0.49	0.11	0.13	0.21	0.15
sampling position of 2nd formant	579.00	583.00	585.00	585.00	584.00	595.00	574.00
2nd formant frequency (MHz)	2.90	2.91	2.92	2.92	2.92	2.97	2.87
3rd formant value (dB)	33.97	49.72	43.38	38.53	36.71	32.88	35.18
3rd formant density (dB/cm)	0.17	0.99	0.62	0.08	0.06	0.07	0.07
sampling position of 3rd formant	826.00	782.00	777.00	845.00	845.00	801.00	822.00
3rd formant frequency (MHz)	4.13	3.91	3.88	4.23	4.23	4.00	4.11
Circumference	20.00	5.00	7.00	63.50	46.50	49.50	50.50

. The main features were the mean value, each resonance peak’s energy value, and frequency location. At the same time, the mean value density and resonance amplitude density parameters were computed.

(3)

Feature combination

The 24-dimension ultrasound energy features were extracted from the time and frequency domains, as shown in

Table 7. Features of plant stems and feature numbers.

Feature No.	Feature Name	Feature No.	Feature Name
1	Energy	13	2nd formant density
2	Energy density	14	2nd formant frequency
3	Mean value	15	3rd formant value
4	Mean density	16	3rd formant frequency
5	Variance	17	removed DC energy
6	peak-to-peak value	18	removed DC variance
7	Spectrum DC	19	removed DC 1st formant value
8	Spectral DC density	20	removed DC 1st formant frequency
9	1st formant value	21	removed DC 2nd formant value
10	1st formant density	22	De-rectified 2nd formant frequency
11	1st formant frequency	23	removed DC 3rd formant value
12	2nd formant value	24	removed DC 3rd formant frequency

.

The ultrasound energy features were selected based on information gain attributes and correlation evaluation methods. The ultrasound energy features were selected to reflect the differences in the stem structure of the different plant species. The number of samples used in the final samples is shown in

Table 8. Sample number of radial features of stems of four species.

Detecting Tree Species	Basho	Sunflower	Palm	Magnolia Denudata	Total Number of Samples (pcs)
Ultrasound samples (pcs)	52	62	35	40	189

.

(4)

Results of feature contributions

The information gain and correlation evaluation methods were used to compute the 24-dimensional feature to identify the most distinguishing energy features. The evaluated features are shown in

Table 9. Contribution of ultrasound features of four plant stems.

Feature Sorting	Feature No.	Information Gain	Feature No.	Correlation Values	Feature No.	Mean Value of Contributions
1	4	1.71	18	0.71	4	1.20
2	3	1.55	5	0.71	3	1.11
3	7	1.29	9	0.70	7	0.96
4	5	1.14	19	0.70	5	0.92
5	18	1.14	4	0.69	18	0.92
6	17	1.12	1	0.69	17	0.90
7	1	1.10	17	0.69	1	0.89
8	2	1.10	6	0.69	9	0.85
9	8	1.07	12	0.68	19	0.85
10	13	1.00	21	0.68	6	0.84
11	9	1.00	13	0.67	12	0.84
12	12	1.00	3	0.67	21	0.84
13	19	1.00	7	0.64	13	0.83
14	21	1.00	20	0.41	2	0.66
15	6	1.00	11	0.41	8	0.63
16	20	0.83	23	0.34	11	0.62
17	11	0.83	15	0.34	20	0.62
18	10	0.45	10	0.29	10	0.37
19	22	0.35	2	0.23	15	0.33
20	14	0.35	24	0.23	23	0.33
21	23	0.32	16	0.23	14	0.25
22	15	0.32	8	0.20	22	0.25
23	16	0.13	14	0.16	16	0.18
24	24	0.13	22	0.16	24	0.18

The final contribution is the mean contribution that is the average of the information gain and correlation.

.

The results showed that the top five feature contributions by the information gain method were the mean density, mean value, spectral DC variance, and removed DC variance. The top five feature contributions by the correlation evaluation method were the removed DC variance, variance, 1st formant value, removed DC 1st formant valued, and mean density. The extracted features were ranked by the mean contribution of the two methods. The mean density was the most compelling feature for structural differences, followed by variance and 1st formant value.

The three features (the mean density, variance, and 1st formant value) were analyzed in 2-dimensional feature maps to explore plant species differentiation, and the experimental results are shown in Figure 6.

Figure 6 shows that the combination of features with the mean density and variance and features with the mean density and 1st formant value, as shown in Figure 6a,b, were linearly distinguishable between herbaceous and woody plant stem bodies with good differentiation. In contrast, the combination of features using variance and the 1st formant value, as shown in Figure 6c, was linearly indistinguishable between the herbaceous and woody classes. However, the difference between sunflowers and the other three stem types was still evident.

4. Discussion and Conclusions

In this paper, the ultrasonic energy signals of the differences in the stems of basho, sunflower, palm, and magnolia denudata were analyzed in the time and frequency domains, which provide a basis for the feature extraction from an ultrasonic energy perspective.

In order to reduce the interference caused by different detection depths, this paper proposed improved ultrasound energy features, that is, the ultrasound energy feature density, which constitutes energy density, mean density, variance density, and resonance peak density features.

The 24-dimensional energy features were extracted from the time and frequency domains. They were evaluated by the information gain and correlation method. Based on the top three energy features in feature contribution, the 2-dimensional feature maps were formed to discover the structure differences of the plant stems. The feature map of the variance and the first formant had good differentiation between herbaceous and woody plant stems. Our future study will test more stem plants to verify the universality of these extracted features.

Based on these extracted features, the feature selection, the filtered or wrapper method, is our future study. The classifier is intended to select random forest (RF), support vector machine (SVM) and Linear discriminant analysis (LDA).