J. For. Res.

(2022) 33:579–589
https://doi.org/10.1007/s11676-021-01362-4

ORIGINAL PAPER

Leaf recognition using BP‑RBF hybrid neural network

Xin Yang1 · Haiming Ni1 · Jingkui Li1 · Jialuo Lv1 ·
Hongbo Mu1 · Dawei Qi1

Received: 24 December 2020 / Accepted: 5 April 2021 / Published online: 26 June 2021
© The Author(s) 2021

Abstract Plant recognition has great potential in forestry Introduction

research and management. A new method combined back
propagation neural network and radial basis function neural Forest resources play an important role in the development
network to identify tree species using a few features and of economies, societies and the environment (Yang and Kan
samples. The process was carried out in three steps: image 2020). However, in recent years, the natural environment
pretreatment, feature extraction, and leaf recognition. In has been severely damaged in many countries, resulting in
the image pretreatment processing, an image segmentation the reduction of forest area and the extinction of many spe-
method based on hue, saturation and value color space and cies. However, there is a new awareness of the importance
connected component labeling was presented, which can of protecting plant species. A fundamental premise of plant
obtain the complete leaf image without veins and back- protection is the accurate recognition and classification of
ground. The BP-RBF hybrid neural network was used to plants (Gong and Cao 2014). On the one hand, recognition
test the influence of shape and texture on species recogni- of plant species can help understand the forest ecosystem
tion. The recognition accuracy of different classifiers was and forest economy (Nevalainen et al. 2017); on the other,
used to compare classification performance. The accuracy of it helps strengthen the management and protection of for-
the BP-RBF hybrid neural network using nine dimensional est resources, and improve the public’s awareness of forest
features was 96.2%, highest among all the classifiers. protection.
Plant recognition and classification based on image fea-
Keywords Leaf recognition · BP-RBF neural network · tures is an important research focus in biodiversity informat-
Image processing · Feature extraction · Machine learning ics, and it is beneficial to explore the evolutionary rules and
relationships of plants and establish a taxonomic database.
The recognition of tree species is often difficult because of
the existence of numerous species and the similarities among
Project funding This work is supported by the Fundamental
Research Funds for the Central Universities (No.2572020BC07) some species. The amount of information contained in leaves
and the Project of National Science Foundation of China is considered and manifested in the colors, shapes, textures,
(No.31570712). veins and edges. Leaves are easy to collect and process with
digital equipment. Based on the above characteristics, leaves
The online version is available at http://​www.​sprin​gerli​nk.​com.
are often used for species recognition (Rahman et al. 2019).
Corresponding editor: Yanbo Hu. Information on species composition of an urban forest
is essential for its management. However, this information
* Hongbo Mu is increasingly difficult to obtain due to limited taxonomic
[email protected]
expertise of the urban managers. Traditional leaf recogni-
* Dawei Qi tion methods require a significant professional knowledge,
[email protected]
and in its absence, may result in low efficiency. With the
1
College of Science, Northeast Forestry University, rapid development of computer technology, researchers
Harbin 150040, People’s Republic of China are able to combine the image processing method, pattern

13
Vol.:(0123456789)
580 X. Yang et al.

recognition and machine learning technology with plant This system is a software solution for automatic recog-
morphology to ascertain the automatic recognition of leaf nition and classification of plant species. The scheme is
images. Wu et al. (2007) extracted 12 digital morphologi- divided into four main steps: (1) color space conversion of
cal features on Flavia dataset (a widely used leaf dataset) images; (2) image pretreatment; (3) leaf feature extraction;
and used the probabilistic neural network (PNN) to test the and (4) design of classifier and recognition. The basic flow
accuracy of their algorithm. The result was very similar of the leaf identification system is shown in Fig. 1.
to other systems (90%). Turkoglu et al. (2019) proposed
that leaf feature extraction may be completed by dividing
the leaf image into two or four parts. The accuracy of their Materials and methods
method with the Flavia dataset was 99.1%. In their study,
image processing based on feature extraction methods such Sample preparation
as color, veins, Fourier descriptors (FD), and gray-level
co-occurrence matrix (GLCM) were used. Tang (2020) In order to achieve the recognition and classification of
used the grey cluster analysis method to establish a quan- species, the necessary preparation work is to establish a
titative feature system of leaves, and used a probabilistic leaf database (Backes et al. 2009; Kumar et al. 2012). The
neural network for classification. The results were to eval- research area was the Experimental Forest Farm of the
uate model performance and the influence of core features Northeast Forestry University, with geographic coordinates
of the model. The results showed that an accuracy of the 45°71′ − 45°72′ N, 126°62′ − 126°63′ E. The dataset com-
GBDT-PNN model using 12 core features was 92.7%, and prised 366 images of leaves belonging to 15 common spe-
the accuracy with all 35 features was 93.5%. cies in Northeast China. The scientific names and sample
Although there have been numerous advances on leaf numbers are shown in Table 1.
classification based on machine learning, there are still Leaves with common shapes, complete fronds, spotless,
some shortcomings. Few researchers have analyzed the and without pests were chosen, including petioles. Dust was
influence of different features on recognition. Most stud- removed from leaves; LEDs were used to illuminate leaves,
ies have selected too many features, which faced challenges and all leaf samples were photographed with a Nikon D850
in practical application. In the process of leaf image pre- digital camera. Leaf images of each species are shown in
treatment, the binary image obtained contains noise after Fig. 2.
image segmentation, usually caused by highlights on the leaf
surface and dust particles scattered on the image acquisi- Image pretreatment
tion device. Traditional denoising methods usually mistake
noise-polluted leaf surface for background. The RGB color space does not distinguish between bright-
Given these problems, this study proposed an image seg- ness and color information and so the images are converted
mentation method based on HSV color space and connected to HSV color space which has good linear scalability and
component labeling, which can completely extract leaves is directly oriented to human visual perception (Perona and
without petiole. The extracted image had a good denoising Malik 1990). Based on HSV images, the background was
performance. In addition, shape and texture were extracted, well separated, leaf contours were extracted and the noise
and the leaf recognition performance of various machine was initially removed. The Otsu algorithm (Chang et al.
learning methods was compared, and a BP-RBF hybrid neu- 2018; Yu et al. 2019) was selected as the threshold segmen-
ral network was newly established. tation method proposed by Otsu (1979). The basic principle

Fig. 1  Flow diagram of leaf recognition

13
Leaf recognition using BP‑RBF hybrid neural network 581

Table 1  Statistics of the species from the training and test sets Shape features
Label Tree species Number of samples
Shape is one of the most important features for character-
Training set Test set Total
izing a leaf because it can be perceived by humans (Wang
1 Acer negundo L 17 5 22 et al. 2008). According to the extracted leaf contour, sev-
2 Ulmus pumila L 23 6 29 eral geometric parameters (Wu et al. 2007) were calcu-
3 Armeniaca sibirica (L.) Lamb 21 6 27 lated, including leaf area (S), the smallest rectangular area
4 Salix matsudana Koidz 25 7 32 surrounding the leaf (S0), the perimeter of the leaf area
5 Populus davidiana Dode 20 5 25 (L), length (b) and width (a) of the minimum enclosing
6 Amygdalus triloba (Lindl.) 15 4 19 rectangle of the leaf, the coordinates (x0, y0) of the center
Ricker of mass of the leaf, the coordinates (x, y) of the upper left
7 Tilia mandshurica Rmpr. & 19 5 24 corner of the rectangle, the length (X and Y) of the rec-
Maxim tangle in the x and y directions, the maximum deflection
8 Fraxinus mandshurica Rupr 17 5 22 angle (m) and the total number of groups (M) of the leaf
9 Betula platyphylla Suk 16 4 20 profile. These geometric parameters were used to further
10 Tilia amurensis Rupr 26 7 33 calculate the following five shape features:
11 Juglans mandshurica Maxim 23 6 29 Rectangularity: the ratio of leaf area (S) to the smallest
12 Acer mono Maxim 16 5 21 rectangle surrounding the leaf (S0)
13 Quercus mongolica Fisch. ex 16 4 20
Ledeb S
14 Lonicera maackii (Rupr.) 16 5 21
E1 =
S0 (1)
Maxim
15 Phellodendron amurense Rupr 17 5 22 Roundness: the similarity between leaf contour and cir-
cle.and circle
4𝜋S
E2 = (2)
is to find the best threshold to maximize the variance within L2
or between clusters to accurately classify background and where, S is leaf area and L is the perimeter.
foreground content. However, the Otsu algorithm is very Aspect ratio: the ratio of length (b) to width (a) of the
sensitive to noise, so noise should be eliminated using image smallest enclosing rectangle
smoothing algorithms.
b
E3 = (3)
Feature extraction a
Deviation degree: the offset degree of the leaf centroid
In the process of feature extraction, shape features (Wu et al.
relative to the smallest enclosing rectangle
2007; Wang et al. 2008) and texture features (Haralick 1973)
are the two most commonly used recognition features.

Fig. 2  Samples of leaf images

used for classification (The
numbers correspond to the
names in Table 1)

13
582 X. Yang et al.

{ (x0 −x)Y
∑∑
, X≥Y E= (P(i, j))2
X(y0 −y) (8)
E4 = (y0 −y)X (4) i j
Y(x0 −x)
, X<Y
Homogeneity: It represents the local uniformity of the
where, (x0, y0) is the coordinates of the center of mass of image
the leaf, (x, y) is the coordinates of the upper left corner of
∑∑ P(i, j)
the rectangle, X and Y are the length of the rectangle in the H=
1 + (i − j)2 (9)
x and y directions. i j
Sawtooth degree: the ratio of the maximum deflection
angle (m) to the total number of leaf profiles (M) In the above formulas, i and j are coordinates (row and
column) of a pixel in the image. P(i, j) is the gray value of
m the pixel located at coordinates (i, j) in a leaf image.
E5 =
M (5)

Support vector machine

Texture features
Support vector machine (SVM), originally developed by
Texture features can be used to quantitatively describe the Vladimir Vapnik, is a powerful tool for solving nonlinear
texture information. The secondary statistics obtained by classification, function estimation, and density estimation
the gray level co-occurrence matrix (Haralick 1973) reflects problems (Zhang et al. 2018). SVM has a great advantage
the texture features and is based on the relation between two in solving the problems of small samples, and nonlinear and
neighboring pixels in a gray image. This study selected the high-dimensional pattern recognition because the test error
following four features: for the independent test set is smaller than other machine
Contrast: It reflects the sharpness of the image and the learning algorithms. Plant leaf recognition and classifica-
depth of the texture. tion is a complex classification problem. However, due to
∑∑ the limitation of leaf numbers, it is difficult to collect a large
C1 = 𝛿(i, j)2 P(i, j) number of image samples for each kind of foliage plants,
(6)
i j
which is possible with the SVM classifier. Hence, this paper
where 𝛿(i, j) = |i − j| is the gray difference between neigh- built a SVM classifier model. The main process is shown
boring pixels, P(i, j) is the gray value of the image. in Fig. 3.
Correlation: It indicates the gray level similarity in the SVM is based on the principle of risk minimization,
row or column direction. which means that the empirical risk and confidence inter-
vals are quite large. Therefore, the output of the model is the
∑ ∑ (i − 𝜇i )(j − 𝜇j )P(i, j)2 optimal solution (Nelson et al. 2008; Roy and Bhattacharya
C2 = (7) 2010; Tarjoman et al. 2012). The basic idea is to use a non-
i j
𝜎i 𝜎j
linear mapping algorithm to convert the linearly inseparable
where 𝜇i , 𝜇j , 𝜎i and 𝜎j are the means and standard deviations samples of the low-dimensional input feature space into a
of the rows and columns of the gray value P(i, j). high-dimensional feature space, making it linearly separable.
Energy: It is a measure of the stability degree of the gray- The optimal hyperplane can be obtained by solving the
scale change of the image texture. following quadratic optimization problem:

Fig. 3  Flow diagram of SVM classifier

13
Leaf recognition using BP‑RBF hybrid neural network 583

1 BP-RBF hybrid neural network in this article. The network

min Φ(w) = ‖w‖2 (10)
2 structure is shown in Fig. 4.

s.t. yi (w ⋅ xi ) + b ≥ 1 , y = 1,2, ⋯ ⋯ ,n (11)

Results
In the case of a particularly large number of features, this
problem can be transformed into its dual problem: Image pretreatment
n
∑
max W(𝛼) = 𝛼i 𝛼j yi yj (xi ⋅ xj ) (12) The leaf RGB image was converted to HSV color space
i=1 (Fig. 5). The optimal threshold obtained by the Otsu algo-
rithm was used to convert the S component image into a
n
∑ binary image. Traditional denoising methods mainly include
w∗ = 𝛼i yi xi (13) median and mean filtering and Gaussian low-pass filtering.
i=1
None of the smoothing methods can completely remove
background noise and leaf veins. There was still unremoved
b1 = y1 − w ⋅ x1 (14) noise inside the leaves in Fig. 6b. For the purpose of elimi-
nating noise, this research presented a method based on con-
n
∑ nected component labeling (Fig. 6c).
s.t. 𝛼i yi = 0 , 𝛼i ≥ 0 , i = 1, 2, ⋅ ⋯ ⋯ , n (15) The petiole needs to be removed because it exceeds the
i=1
extraction range of the leaf features and interferes with the
where, 𝛼 = (𝛼1 , ⋯ ⋯ , 𝛼n ) is the Lagrange multiplier, w∗ is the calculation results. The main vein could be distinguished
normal vector of the optimal hyperplane, and b1 is the offset from the mesophyll according to the H component (Fig. 7a).
of the optimal hyperplane. In the solution and analysis of To keep the background clean, the denoising binary image of
this type of optimization problem, the Karush–Kuhn–Tucker H and S components were subjected to matrix point opera-
(KKT) condition will play a very important role. In the sec- tions and the non-background H component image could
ond constraint formula, the solution must satisfy Eq. (16): be obtained as shown in Fig. 7b. This image was subjected
{ } to brightness stretching and binarization. An appropriate
𝛼i yi (w ⋅ x + b) − 1 = 0, i = 1, 2, ⋯ ⋯ , n (16) morphological algorithm was then executed, and the binary
image without petiole could be obtained (Fig. 7e).
where, the samples of 𝛼i > 0 are called support vectors. The S component has a better performance at displaying
The final classification function is as follows. the minor veins, while the H component shows the midveins
n
∑ clearly. S and H components were processed by dot multipli-
f (x) = 𝛼i yi (x ⋅ xi ) + b1 (17) cation. The gray value was normalized afterwards, and the
i=1 median filter was used for image enhancement. In Fig. 8, the
processed image retains the texture features of all the leaf
veins (Larese et al. 2014).

BP‑RBF neural network Feature extraction

A neural network is a complex machine learning algorithm Matlab built-in functions were used to obtain the M groups
used for prediction analysis. It is trained with a set of inputs of coordinate values located in the leaf contour (Fig. 9). For
and outputs, and implicit relationships between inputs and
outputs are extracted. The back propagation neural network
(BPNN) is a type of multilayer forward neural network,
which has strong data compression and fault tolerance abil-
ities (Rumelhart et al. 1986). It is commonly used in the
fields of pattern recognition, data classification and predic-
tive analysis because of its good adaptability and robustness
(Xu et al. 2018; Yang and Kan 2020).
The radial basis function neural network (RBFNN) has
the ability to approximate functions with arbitrary precision.
Based on the previous study using BPNN to identify plant
leaves, BPNN and RBFNN are connected in series to form a Fig. 4  BP-RBF neural network structure

13
584 X. Yang et al.

(a) RGB image. (b) H component. (c) S component. (d) V component.

Fig. 5  RGB image and H, S and V components of leaf in HSV color space

Fig. 6  Segmentation and

denoising of S component
image

(a) S component. (b) Binary graph. (c) Denoising.

(a) (b) (c) (d) (e)

Fig. 7  Main process of image segmentation a H component; b background removal; c brightness stretch; d binary image; e petiole removal

each vector formed by two adjacent groups of coordinate Recognition results

values, the large angle deflection number m was calculated,
and its ratio was taken as the feature. It should be noted that In this study, the combination of 5-dimensional shape
after the petioles are removed, there are several connected features and 4-dimensional texture features is defined as
components in the binary image of A. negundo (Fig. 9). fusion features. KNN, SVM, BPNN and BP-RBF were
Therefore, the average of multiple groups of ratios is calcu- used to classify leaves by training above three types of
lated. For texture features, several secondary statistics of the features.
gray level co-occurrence matrix were selected as the input
vectors. The mean of shape and texture features are shown
in Tables 2 and 3.

13
Leaf recognition using BP‑RBF hybrid neural network 585

Fig. 8  Texture from H, S and

gray images

(a) H component (b) S component (c) Gray image

(a) (b) (c)

Fig. 9  Shape features extraction: a shape features; b contour extraction of Betula platyphylla; c contour extraction of Acer negundo

Table 2  Mean of shape features Label Shape features

Rectangularity Roundness Aspect ratio Deviation degree Sawtooth degree

1 0.4121 0.06222 1.179 1.224 0.3299

2 0.6268 0.1324 2.49 0.9003 0.3266
3 0.5656 0.1179 1.496 0.843 0.3385
4 0.5961 0.09547 5.133 0.9148 0.317
5 0.6382 0.1968 1.091 1.075 0.3291
6 0.5992 0.1236 1.656 0.9879 0.3387
7 0.6529 0.1259 1.255 0.9132 0.3346
8 0.6196 0.1433 2.952 0.8794 0.316
9 0.5963 0.1631 1.228 0.8394 0.3258
10 0.601 0.1236 1.281 0.8356 0.3317
11 0.6913 0.1651 2.716 0.9064 0.3235
12 0.4348 0.08122 1.245 1.139 0.3299
13 0.5783 0.1362 1.725 1.089 0.3315
14 0.6077 0.1955 1.983 0.9106 0.3384
15 0.6088 0.1605 2.578 0.9015 0.3248

13
586 X. Yang et al.

Table 3  Mean of texture features not completely recognized, while the remaining tree species
Label Texture features
were all correctly recognized.

Contrast Correlation Energy Homogeneity

BPNN and BP‑RBF
1 0.1416 0.9764 0.7437 0.9626
2 0.05388 0.9904 0.6701 0.9765 There are several important steps in the establishment of BP
3 0.1385 0.9712 0.7864 0.9664 neural network:
4 0.1061 0.9552 0.7853 0.9691
5 0.213 0.977 0.4925 0.9308 (1) Determination of the number of neuron nodes in the
6 0.04346 0.9928 0.762 0.9825 input and output layers. Taking 9-dimensional fusion
7 0.09316 0.9926 0.4632 0.9612 features as an example, there were 9 neuron nodes in
8 0.1207 0.9822 0.6404 0.9581 the input layer and 15 in the output layer.
9 0.1045 0.9879 0.5006 0.9542
10 0.1127 0.9852 0.5839 0.955
11 0.1261 0.9822 0.5709 0.9549 (2) Choice of the number of neuron nodes in the hidden
12 0.1278 0.9811 0.662 0.9596 layer. The number of hidden layer nodes which influ-
13 0.1236 0.9866 0.5375 0.953 ences the training performance
√ can be determined by
14 0.04101 0.993 0.7767 0.9837 the empirical formula j = i + k + c , where j is the
15 0.07374 0.9911 0.565 0.9708 number of hidden layer neural nodes, i and k are the
number of neurons in the input and output layers,
respectively. As mentioned above, i = 9 and k = 15. c is
KNN and SVM the regulation constant in [1, 10] (Xu et al. 2018). The
result was j ∈ [6, 14]. Figure 11 shows the change of
A KNN classifier model (Muhammad et al. 2019) was estab- mean square error (MSE) with j. The number of hid-
lished after normalizing the features. The recognition results den layer nodes was selected as 11, and then the local
were tested using the test set data. Euclidean distance was optimal solution of neural network MSE was obtained.
chosen as the parameter to calculate the distance, and the K (3) The establishment, training and testing of BP neural
value was set to 1 (1-NN). The rate of recognition accuracy networks. Based on the “newff” function in matlab, a
of shape, texture and fusion features were 87.3%, 48.1% and three-layer neural network with the structure of "9–11-
92.4%, respectively. For the SVM classifier, the linear kernel 15" was established. The training function was Leven-
function was used. The SVM rate of recognition accuracy berg Marquardt algorithm. “Learngdm” and MSE were
of the three features were 50.6%, 16.5% and 86.1%, respec- selected for the adaptive learning function and perfor-
tively. The test results of the fusion features are shown in mance function. The transfer functions of the hidden
Fig. 10. The results show that the leaves with the label 8, and output layers were both “logsig”, and “trainParam.
10, 11, 12 by KNN and the label 8, 11, 12, 15 by SVM were goal” was set to 0.001.

Fig. 10  Fusion features recognition results of KNN and SVM: a KNN; b SVM. The ordinate is the label of tree species and the abscissa is the
sample order number of test set

13
Leaf recognition using BP‑RBF hybrid neural network 587

0.040 unchanged, and then the BPNN is connected with an RBF

network. In the process of RBF network training, the most
0.035 important parameter is “spread”. It is vital to select “spread”
reasonably, and its value should be large enough to make the
0.030 RBF neural network respond to the interval covered by the
Mean square error

input vector, and make the prediction performance smoother.

0.025 However, an overly large spread may cause numerical prob-
lems. The calculation of MSE and recognition accuracy rate
0.020
with “spread” is shown in Fig. 13. When “spread” is in the
range of 3.3–3.7, there is a small MSE and a large accuracy.
Finally, 3.4 was selected as the ideal value of “spread”.When
0.015
the 5, 4 and 9-dimensional features were imported into the
BP-RBF neural network for training, the highest recognition
0.010
5 6 7 8 9 10 11 12 13 14 15 accuracies of the test set were 88.6%, 49.4% and 96.2%,
Number of nodes respectively. Compared with KNN, SVM and BPNN, the
new BP-RBF hybrid neural network can improve the accu-
Fig. 11  Relationship between mean square error and number of neu- racy rate of leaf recognition.
ron nodes in hidden layer

Discussion
The training process of the BP neural network and the
MSE are shown in Fig. 12. For the same dataset configura- Table 4 shows the recognition accuracy rates of different
tion, the recognition accuracy rate may change because the classifiers, which provided the opportunity for performance
weight generated by each training was not a certain value, comparisons. As can be seen, whether shape, texture or
which was different from KNN and SVM. When shape and fusion features were used, BP-RBF had the highest recog-
texture features were used as input, the average accuracy rate nition accuracy among all the methods. In contrast, SVM
was 88.1% and 50.6%, respectively, and the highest accuracy had the lowest. The BP-RBF network was used to optimize
rate was 89.9% and 55.7%, respectively. The highest recogni- the previous methods, and the fusion feature recognition
tion accuracy rate of BPNN was 94.9% when using fusion accuracy reached 96.2%, 1.3% higher than the BPNN. The
features as input. contribution of various features to recognition rate can be
The parameter selection of the BP-RBF network can be compared by selecting shape and texture features as the
divided into two parts. The first is the parameter selection input. When using texture features, the recognition rates of
of the BP neural network. The parameters of BPNN are kept all methods were generally low (less than 50%). It’s worth

Fig. 12  Training process of BPNN (left) and mean square error of training (right)

13
588 X. Yang et al.

4.0 Table 5  Comparison of the proposed method with other studies

Mean square error 90
3.5 Accuracy Method Number of Leaf features Accuracy (%)
88 leaves
3.0
86 PNN 1800 12 90.0
Mean square error

2.5 LDC 1907 15 94.0

Accuracy (%)
84
2.0
SVM 660 15 87.0
82 GBDT-PNN 1600 35 93.5
1.5 SSA based SVM 1600 36 96.7
80
1.0
BP-RBF 366 9 96.2
78
0.5
76
0.0 performance in plant recognition systems using fewer sam-
0 1 2 3 4 5
Spread
ples and features.
Plant classification methods have great potential in for-
est studies and management. There were reports that plant
Fig. 13  Mean square error of recognition results and recognition
accuracy of BP-RBF identification error for professionals was 10–20% to the spe-
cies level (Gray and Azuma 2005; Crall et al. 2011). Gener-
ally speaking, leaf structure allows closely related taxa to
Table 4  Comparison accuracy results of different classifier differentiate from each other (Merrill 1978; Sajo and Fls
Features Accuracy rate of classifiers (%)
2002; Espinosa et al. 2006). At the same time, leaf shape and
texture are extracted from leaf structure. The system in this
KNN SVM BPNN BP-RBF study can automatically preprocess leaf images, extract fea-
Shape 87.3 50.6 87.3 88.6 tures, and realize the identification of the species. Its accu-
Texture 48.1 16.5 48.1 49.4 racy is comparable to the work of professionals, and can be
Fusion 92.4 86.1 94.9 96.2 used to develop a portable forest tree species recognition
KNN k-nearest neighbors, SVM Surport vector machine, BPNN Back
system helpful to non-professionals. The advantage is that
propagation neural network, BP-RBF Back propagation and radial the selected features are not affected by translation, rotation
basis function neural network or scale of the leaf images. Although the recognition system
proposed in this study has excellent performance, there is
still room for improvement. Future work is to optimize the
noting that the recognition rate was significantly improved classification methods. On the one hand, the recognition per-
after using texture features combined with shape features. formance of the system can be further improved by enriching
This indicates that the contribution of shape features was the leaf features. On the other hand, more classification mod-
obviously higher than that of texture features in this study. els and algorithms, such as convolutional neural network,
Therefore, it is necessary to fuse texture features with other require further exploration and experimentation. And the
types of features for plant identification. number of tree species in the dataset needs be increased.
According to Fig. 10, KNN and SVM were not able to
recognize some specific types of leaves in the sample. The
results of the above classifiers were unsatisfactory when Conclusion
identifying the tree species labeled 8, 11 and 12, while the
other tree species were all recognized. The reason might This study used image processing technologies and machine
be that the features of these plants are so similar to others learning algorithms to identify 15 kinds of plant leaves. A
that the current features are not enough to distinguish these new BP-RBF hybrid neural network was proposed to further
leaves. The next task is to extract other suitable features improve the recognition accuracy rate. The conclusions of
to increase the differences between different kinds of plant this study are as follows.
leaves. In this study, a leaf database of common tree species in
Up to now, researchers have proposed numerous effec- Northeast China was established. An image segmentation
tive methods for species recognition. Commonly used leaf method, based on HSV color space and connected compo-
recognition methods include PNN (Wu et al. 2007), LDC nent labeling was presented, which can obtain the complete
(Kalyoncu and Toygar 2015), GBDT-PNN (Tang 2020), and leaf image without veins and background. Leaf shape and
SVM (Salman et al. 2017; Ahmed and Hussein 2020). As texture were extracted using feature extraction algorithms.
seen in Table 5, the BP-RBF neural network achieved high With all the leaf samples in our database, the recognition

13
Leaf recognition using BP‑RBF hybrid neural network 589

rates of KNN, SVM, BPNN and BP-RBF methods in the Merrill E (1978) Comparison of mature leaf architecture of three types
test set were 92.4%, 86.1%, 94.9% and 96.2%, respectively. in Sorbus L. (Rosaceae). Bot Gaz 139(4):447–453
Muhammad AFA, Lee SC, Fakhrul RR, Farah IA, Sharifah RWA
Accordingly, the proposed BP-RBF hybrid algorithm had (2019) Review on techniques for plant leaf classification and rec-
higher recognition accuracy than the other algorithms. For ognition. Computers 8(4):77
each method, the recognition contribution of shape features Nelson JDB, Damper RI, Gunn SR, Guo B (2008) Signal theory for
was greater than that of texture features. Compared with SVM kernel design with applications to parameter estimation and
sequence kernels. Neurocomputing 72(1–3):15–22
single-class features, the highest recognition rate can be Nevalainen O, Honkavaara E, Tuominen S, Viljanen N, Hakala T, Yu
obtained using fusion features. The BP-RBF neural network XW, Hyyppa J, Saari H, Polonen I, Imai NN, Tommaselli AMG
can achieve high recognition accuracy rates with fewer fea- (2017) Individual tree detection and classification with UAV-
tures and leaf samples compared with the other methods. In based photogrammetric point clouds and hyperspectral imaging.
Remote Sensing 9(3):185
future studies, the performance of this proposed method will Otsu N (1979) A threshold selection method from gray-level histo-
be improved by using other feature extraction techniques grams. IEEE Trans Syst Man Cybern 9(1):62–66
and classifiers. Perona P, Malik J (1990) Scale-space and edge detection using
anisotropic diffusion. IEEE Trans Pattern Anal Mach Intell
Open Access This article is licensed under a Creative Commons 12(7):629–639
Attribution 4.0 International License, which permits use, sharing, adap- Rahman MM, Islam MS, Sassi R, Aktaruzzaman M (2019) Convolu-
tation, distribution and reproduction in any medium or format, as long tional neural networks performance comparison for handwritten
as you give appropriate credit to the original author(s) and the source, Bengali numerals recognition. SN Appl Sci 1(12):1660
provide a link to the Creative Commons licence, and indicate if changes Roy K, Bhattacharya P (2010) Improvement of iris recognition per-
were made. The images or other third party material in this article are formance using region-based active contours, genetic algorithms
included in the article’s Creative Commons licence, unless indicated and SVMs. Int J Pattern Recognit Artif Intell 24(8):1209–1236
otherwise in a credit line to the material. If material is not included in Rumelhart DE, Hinton GE, Williams RJ (1986) Learning representa-
the article’s Creative Commons licence and your intended use is not tions by back-propagating errors. Nature 323(6088):533–536
permitted by statutory regulation or exceeds the permitted use, you will Sajo MG, Rudall PJ (2002) Leaf and stem anatomy of Vochysiaceae in
need to obtain permission directly from the copyright holder. To view a relation to subfamilial and suprafamilial systematics. Bot J Linn
copy of this licence, visit http://​creat​iveco​mmons.​org/​licen​ses/​by/4.​0/. Soc 138(3):339–364
Salman A, Semwal A, Bhatt U, Thakkar VM (2017) Leaf classification
and identification using Canny Edge Detector and SVM classi-
fier. In: 2017 International Conference on Inventive Systems and
References Control (ICISC) pp. 1-4
Tang ZX (2020) Leaf image recognition and classification based
Ahmed A, Hussein SE (2020) Leaf identification using radial basis on GBDT-probabilistic neural network. J Phys Conf Series
function neural networks and SSA based support vector machine. 1592:012061
PLoS ONE 15(8):e0237645 Tarjoman M, Fatemizadeh E, Badie K (2012) An implementation of a
Backes AR, Casanova D, Bruno OM (2009) A complex network- CBIR system based on SVM learning scheme. J Med Eng Technol
based approach for boundary shape analysis. Pattern Recogn 37(1):43–47
42(1):54–67 Turkoglu M, Hanbay D (2019) Recognition of plant leaves: an approach
Chang ZY, Cao J, Zhang YZ (2018) A novel image segmentation with hybrid features produced by dividing leaf images into two
approach for wood plate surface defect classification through and four parts. Appl Math Comput 352:1–14
convex optimization. J for Res 29(6):1789–1795 Wang XF, Huang DS, Du JX, Xu H, Heutte L (2008) Classification
Crall AW, Newman GJ, Stohlgren TJ, Holfelder KA, Graham J, Waller of plant leaf images with complicated background. Appl Math
DM (2011) Assessing citizen science data quality: an invasive Comput 205(2):916–926
species case study. Conserv Lett 4(6):433–442 Wu SG, Bao FS, Xu EY, Wang YX, Xiang QL (2007) A leaf recogni-
Espinosa D, Llorente J, Morrone JJ (2006) Historical biogeographical tion algorithm for plant classification using probabilistic neural
patterns of the species of Bursera (Burseraceae) and their taxo- network. In: 2007 IEEE International Symposium on Signal Pro-
nomic implications. J Biogeogr 33(11):1945–1958 cessing and Information Technology, pp. 11–16
Gong DX, Cao CR (2014) Classification of plant leaves based on con- Xu ZH, Huang XY, Lin L, Wang QF, Liu J, Yu KY, Chen CC (2018)
volutional neural network. Comput Mod 4:12–15 BP neural networks and random forest models to detect damage
Gray AN, Azuma DL (2005) Repeatability and implementation of a by Dendrolimus punctatus Walker. J Forestry Res 31:107–121
forest vegetation indicator. Ecol Ind 5(1):57–71 Yang R, Kan J (2020) Classification of tree species at the leaf level
Haralick RM, Shanmugam K, Dinstein IH (1973) Texture features for based on hyperspectral imaging technology. J Appl Spectrosc
image classification. IEEE Trans Syst Man Cybern 3(6):610–621 87(1):184–193
Kalyoncu C, Toygar O (2015) Geometric leaf classification. Comput Yu HL, Liang YL, Liang H, Zhang YZ (2019) Recognition of wood
vis Image Underst 133:102–109 surface defects with near infrared spectroscopy and machine
Kumar N, Belhumeur PN, Biswas A, Jacobs DW, Kress WJ, Lopez vision. J Forestry Res 30:2379–2386
IC, Soares JVB (2012) Leafsnap: A computer vision system for Zhang JW, Song WL, Jiang B, Li MB (2018) Measurement of lumber
automatic plant species identification. European Conference on moisture content based on PCA and GS-SVM. J Forestry Res
Computer Vision, pp. 502–516 29(2):1–8
Larese MG, Namías R, Craviotto RM, Arango MR, Gallo C, Granitto
PM (2014) Automatic classification of legumes using leaf vein Publisher’s Note Springer Nature remains neutral with regard to
image features. Pattern Recogn 47(1):158–168 jurisdictional claims in published maps and institutional affiliations.

13