Hindawi

Computational Intelligence and Neuroscience

Volume 2022, Article ID 7179477, 9 pages
https://doi.org/10.1155/2022/7179477

Research Article
Land Resource Use Classification Using Deep Learning in
Ecological Remote Sensing Images

Bin Xia ,1 Fanyu Kong ,2 Jun Zhou ,3 Xin Wu ,1 and Qiong Xie 1

1
Department of Management, Chengyi University College, Jimei University, Xiamen, Fujian 361021, China
2
Chongqing Engineering Technology Research Center for Development Information Management,
Chongqing Technology and Business University, Chongqing, 400067, China
3
Chongqing Business Vocational College, Chongqing, 401331, China

Correspondence should be addressed to Fanyu Kong; [email protected]

Received 21 February 2022; Revised 22 March 2022; Accepted 13 April 2022; Published 21 April 2022

Academic Editor: Ziya Uddin

Copyright © 2022 Bin Xia et al. This is an open access article distributed under the Creative Commons Attribution License, which
permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Aiming at the problems that the traditional remote sensing image classification methods cannot effectively integrate a variety of
deep learning features and poor classification performance, a land resource use classification method based on a convolutional
neural network (CNN) in ecological remote sensing images is proposed. In this study, a seven-layer convolution neural network is
constructed, and then the two fully connected layer features of the improved CNN network training output are fused with the fifth
layer pooled layer features after dimensionality reduction by principal component analysis (PCA), so as to obtain an effective
remote sensing image feature of land resources based on deep learning. Further, the classification of land resources remote sensing
images is completed based on a support vector machine classifier. The remote sensing images of Pingshuo mining area in Shanxi
Province are used to analyze the proposed method. The results show that the edge of the recognized image is clear, the clas-
sification accuracy, misclassification rate, and kappa coefficient are 0.9472, 0.0528, and 0.9435, respectively, and the model has
excellent overall performance and good classification effect.

1. Introduction and fast update. Therefore, how to accurately extract useful

land information from massive remote sensing image data
The remote sensing image is a comprehensive image by computer to achieve efficient land use is a key problem to
reflecting various surface information obtained by sensors. be solved [4].
The research on target classification of large-area remote For remote sensing image target classification, the
sensing images is not only an important way to obtain land computer automatically distinguishes the attributes of pixels
cover information but also provides important basic support in remote sensing images with patterns representing certain
for its application in the fields of sea situation monitoring, features through a pattern recognition system, so as to
urban planning, environmental supervision, rescue, disaster obtain the classification information of remote sensing
relief, and military reconnaissance; it is of great significance images [5]. In the research of land resource use classification
both from the perspective of social economy and ecological using remote sensing images, researchers mostly used visual
environment [1]. With the continuous development of re- interpretation and traditional pattern recognition classifi-
mote sensing technology, remote sensing images now show cation methods at first. Visual interpretation is simple, but it
the characteristics of hyperspectral, high space, and high takes a long time, and there are personal differences,
resolution. The information obtained from the images is resulting in inaccurate classification [6]. Traditional classi-
more and more comprehensive, and its application field is fication methods include a minimum distance method,
also expanding [2, 3]. The remote sensing image of land maximum likelihood method, etc. [7, 8]. Reference [9]
resources has a large amount of data, complex information, studies feature extraction based on high-resolution remote
2 Computational Intelligence and Neuroscience

sensing images for coastal land use planning. Through the particular, it is difficult to meet the needs of current ap-
research and analysis of space motion remote sensing image plications in classification efficiency and speed and cannot
sequence, the characteristic parameters of land environment well solve many problems of high spectral remote sensing
and moving objects are obtained, but the consideration of images for land resources [19]. Therefore, this study pro-
ecological factors is relatively single, which does not have poses a land resource use classification method using deep
good popularization. Reference [10] compares the classifi- learning in ecological remote sensing images. The innova-
cation results of remote sensing images in specific areas by tions of this study are summarized as follows:
four methods: random forest, support vector machine, re-
(1) In this study, three high-level features of remote
gression tree, and minimum distance. Reference [11] pro-
sensing images are extracted by using the convolu-
posed a method of normalized differential vegetation index
tional neural network (CNN), and a variety of depth
(NDVI) using time series. Using a time series NDVI da-
image features are fused in series. The fused features
tabase to modify the classification results can significantly
cover more complete information and have stronger
improve the classification accuracy of land cover products,
discrimination.
but this method increases the amount of calculation and
cost. Reference [12] used the object-oriented classification (2) To further improve the classification performance,
method combined with fuzzy classification and cart (clas- the proposed method designs a remote sensing
sification and expression tree) decision tree classification image classifier based on SVM, which combines deep
method to classify the land information of Dongjiang River learning features with a deep classifier to solve the
Basin and obtained a more accurate classification effect than problem of poor classifier performance.
the maximum likelihood method and unsupervised classi-
fication method. Although the pattern recognition classifi- 2. Study Area and Data
cation method overcomes some shortcomings of visual
interpretation, it is not good at extracting spatial informa- The visible shortwave infrared hyperspectral camera carried
tion and has poor flexibility. by the “Gofen 5” (GF-5) satellite has a spectral resolution of
With the development of remote sensing technology and 5–10 nm, a spatial resolution of 30 m, and a width of 60 km.
computer technology, many new classification methods are The camera can simultaneously obtain the spatial infor-
gradually emerging, mainly including artificial neural network mation and spectral information of 330 continuous spectral
(ANN), support vector machine (SVM), and fuzzy theory and segments of ground objects in the range of 400–2500 nm.
expert system [13, 14]. Reference [15] proposed a remote The collected data are mainly composed of two parts: visible
sensing image classification method combining SVM and near infrared (VNIR) and short wave infrared (SW). Among
k-nearest neighbor. Using the class separability of SVM and the them, VNIR has 150 bands and SW has 180 bands, a total of
spatial and spectral characteristics of remote sensing data, a 330. The VNIR band range is about 0.39–1.03, the spectral
distance formula is designed as the measurement standard resolution is 5 nm, the SW band range is about 1.0–2.5, and
considering vector brightness and direction, which effectively the spectral resolution is about 10 μm.
realizes the accurate classification of remote sensing images, but The study area of this study is located in the Pingshuo
the classification efficiency is low. Reference [16] uses the land mining area, Shanxi Province, covering about 400 km2,
segmentation method of remote sensing image based on the N39°24′52″–39°37′15″, and E 112°16′29″–112°33′43″. This
convolutional neural network to realize the correct marking of area is the largest open-pit coal mine in China, and the
different land cover types. However, for remote sensing images ecological environment has been damaged due to perennial
with complex background, more and larger database learning mining. Therefore, it is of great significance to study the land
and training are needed to better complete the classification cover types in this area. The data used in the proposed
task. Aiming at the problem that traditional remote sensing method are visible short wave infrared hyperspectral data of
image classification methods are vulnerable to the loss of spatial GF-5 satellite, with a total of 4 images. The corresponding
features, reference [17] proposed an image semantic seg- high-resolution image of “Gaofen-2” satellite with the closest
mentation method based on a dense coordinate transformation region (the spatial resolution of the fused image is 0.8 m) and
network, which improves the accuracy of semantic segmen- the global 30 m land cover type map were obtained free of
tation of high-resolution remote sensing images but still has a charge from Tsinghua University. First, atmospheric cor-
certain dependence on the training data set. Reference [18] rection is carried out to remove the impact of atmosphere on
proposed a feature integration network including multiscale the image. Then, referring to the thematic map of land cover
features and enhancement stages for the classification of land types with 30 m spatial resolution, the land cover types are
remote sensing images and used two-dimensional extended manually drawn on high-resolution images using the En-
convolution with different sampling rates for each scale feature vironment for Visualizing Images (ENVI) platform. Finally,
layer to realize image classification with higher accuracy than the coverage type map is downsampled to 30 m resolution as
ordinary depth learning methods, but the classification effi- the real label of land cover in this area, as shown in Figure 1.
ciency needs to be further improved.
The improvement of most classification algorithms can 3. Research Method
improve the accuracy of land resources classification, but
still, there are problems such as too large processing scale, 3.1. System Model. The proposed method first designs a
complex calculation, and easy to fall into the minimum. In seven-layer CNN and then inputs high-resolution remote
Computational Intelligence and Neuroscience 3

is represented by Pool1, Pool2, and Pool5, respec-

tively. Both the third and fourth layers have only one
convolution layer, which is represented by Conv3
and Conv4. The sixth and seventh layers are all
connected layers, represented by FC6 and FC7, re-
spectively. The overall architecture of remote sensing
land image classification method based on a 7-layer
CNN network structure is shown in Figure 2.
(4) The training samples of remote sensing images are
Original image Label image used to train the CNN. First, the remote sensing
E 112°0′0″ E 113°0′0″ image training set is input into the built CNN to
N calculate the output value of each neuron of CNN.
W E
Ulanqab City Assuming that layer l is a convolution layer, the cal-
S
Hohhot City culation of the j feature map ylj of layer l is as follows:
N 40°0′0″

N 40°0′0″
Datong City ⎝ 􏽘 yl−1 ∗ κl + bl ⎞
⎠
ylj � δ⎛ i ij j , (1)
i∈Ml− 1

where ∗ is convolution operation, yl−1 i is the i feature map

of layer l − 1, κlij is the convolution kernel used for con-
Shuozhou City
nection between yl−1 i and ylj , blj is the offset of ylj , δ is the
activation function, and Ml− 1 is the number of feature maps
of layer l − 1.
N 39°0′0″
N 39°0′0″

Assuming that layer l is a pool layer, the calculation of

Xinzhou City
the j feature map ylj of layer l is as follows:
0 10 20 40 60 80
Km ylj � δ􏼐αlj · f􏼐yl−1 l
(2)
j 􏼑 + bj 􏼑,
E 112°0′0″ E 113°0′0″
where αlj is the pooling parameter of ylj , yl−1
i is the j feature
Figure 1: Geographical location of the study area. map of layer l − 1, f is the pooling function, and blj is the
offset of ylj .
sensing image samples into the network for training. The Assuming that layer l is a fully connected layer, the
specific steps are as follows: calculation of the j feature map ylj of layer l is as follows:
(1) Build multiclass remote sensing image sample data ylj � δ􏼐yl− 1 + blj 􏼑, (3)
set I � [I1 , I2 , . . . , Ii , . . . , IN ] and make the corre-
sponding sample label L � [L1 , L2 , . . . , Li , . . . , LN ], where yl− 1 is the weighted result of all feature maps of layer
where N represents a class N remote sensing image, l − 1 and blj is the offset of ylj .
Ii (i � 1, 2, . . . , N) represents a collection of class i Second, the overall loss function of CNN is calculated.
remote sensing images, and Li represents a label Any one of the label samples Gi (i � 1, 2, . . . , N × m) is set in
collection of class i remote sensing images. the remote sensing image training set. The label of Gi is
(2) The remote sensing image data set is divided into actually a one-of-N label. That is, for sample Gi , its classi-
training set and test set. For each type of remote fication label is as follows:
sensing image in the data set, the total number of 1, Gi actually belongs to category k,
samples is n, from which m images are randomly ϑki � 􏼨 (4)
selected to build a training set, and the other n − m 0, Gi is not actually a class k.
images to build a test set, as follows: Tr � [Tr1 , Tr2 ,
For label sample Gi , if the probability of model identi-
. . . , Tri , . . . , TrN ] and Te � [Te1 , Te2 , . . . , Tei ,
fication of class k(k � 1, 2, . . . , N) is pki , then the error is
. . . TeN ], where Tri represents the training set of
defined as Ei :
class i remote sensing images, including m images;
Tei represents the test set of class i remote sensing 1 N k 2
images, including n − m images. Ei � 􏽘 􏼐pi − ϑki 􏼑 . (5)
2 k�1
(3) Build a seven-layer CNN model. The first five layers
are represented by the first layer, the second layer, Based on the errors of all training samples, the loss
the third layer, the fourth layer, and the fifth layer, function φE of the model is calculated as follows:
respectively. The first layer, the second layer, and the
N×m
fifth layer contain convolution layer and pool layer. 1 N×m N k 2
φE � 􏽘 E i � 􏽘 􏽘 􏼐pi − ϑki 􏼑 . (6)
Each convolution layer is represented by Conv1, i�1
2 i�1 k�1
Conv2, and Conv5, respectively, and each pool layer
4 Computational Intelligence and Neuroscience

CNN design based on Fusion of The function image corresponding to TReLU is shown in

Characteristics of the fifth pool layer

deep learning depth features Figure 3 (assuming that β is 1).

after dimensionality reduction

The initial value of β is set to 1. As can be seen from

Feature dimension reduction

Remote sensing image Serial fusion
data set Figure 3, the function is approximately linear at the origin

based on PCA
Fusion and has a fast convergence speed [22, 23]. Compared with
Seven layer CNN based
features the existing activation functions Sigmoid, ReLU, and
on deep learning PReLU, the proposed improved activation function has the
following advantages:
SVM
Characteristics of the fifth (1) he Problem of Gradient Disappearance. When x > 0,
pool layer
Classification the derivative value of the function is always 1, so the
Characteristics of the sixth results TReLU function maintains the gradient without
full connection layer
Remote sensing attenuation at x > 0, which alleviates the problem.
image
Characteristics of the seventh
classification (2) Activation of Negative Values. The TReLU function
full connection layer
Depth feature extraction from
based on SVM retains some gradient values in the negative half-axis
remote sensing images unsaturated region. When the activation value falls
Figure 2: The overall architecture of the proposed method. into the unsaturated region, it can still obtain ef-
fective activation and retain the characteristics of the
image. At the same time, the size of the unsaturated
Finally, the gradient descent algorithm is used to min- region is controlled by parameter β to activate the
imize the loss function and update the parameters in the negative value feature more effectively [24]. In the
network. The purpose of training CNN is to find the optimal actual training, with the continuous training, by
parameters to minimize the loss function φE . The parameters automatically adjusting the parameters of β, more
of CNN are κlij , αlj , and blj . ψ represents the above three eigenvalues falling on the negative axis can be ac-
parameters, that is, ψ � (κlij , αlj , blj ); after the CNN is trained tivated and more information can be transmitted to
through the remote sensing image training set, a set of the front layer, alleviating the phenomenon of gra-
parameters ψ ∗ can be obtained as follows: dient disappearance [25].
(3) Approximation to 0-Means Distribution. The TReLU
ψ ∗ � arg min φE . (7) function has an active value on the negative half axis,
ψ
which ensures that the mean value of the output is
The gradient descent algorithm is used to update the approximately 0. The average value of the output of
parameter ψ of CNN and minimize the loss function φE : the upper layer is approximately 0, which effectively
zφ 􏼌􏼌􏼌 alleviates the offset of the ReLU activation function,
ψ (i) � ψ (i− 1) − ε E 􏼌􏼌􏼌 , (8) and the weight can be updated quickly, so as to
zψ ψ�ψ(i)
obtain a faster gradient descent speed.
where ε is the learning rate of CNN, which determines the (4) Robustness to Noise. The TReLU function has soft
adjustment range of each step; ψ (i) is the updated parameter saturation on the negative half axis when the output
of group i; ψ (i− 1) is the parameter of group i − 1; zφE /zψ is range of the function is [0, 1). Soft saturation means
the partial derivative of parameter ψ for loss function φE . that the function can reduce the change of infor-
mation output to the next layer, which is robust to
noise and reduces complexity.
3.2. Improved Activation Function TReLU. In this study, a
TReLU activation function combining the advantages of
tanh function and parametric ReLU (PReLU) function with 3.3. High-Level Feature Extraction. The designed CNN is
parameters is proposed. The TReLU activation function not used to extract multiple depth features of remote sensing
only retains the advantages of fast convergence speed of images. First, the whole data set, including all sample images
PReLU function and can alleviate the disappearance of in the training set and test set, is input into the trained CNN,
gradient but also uses tanh function to introduce negative and the first five layers of features of all sample images are
half axis activation value and its soft saturation character- automatically learned through the CNN model. Among
istics to prevent “neuron death” and offset and is more them, the convolution kernel of the first layer mainly extracts
robust to noise [20, 21]. the bottom features of the image, such as edges, angles, and
The mathematical expression of TReLU activation curves. The input of the second layer is actually the output of
function is as follows: the first layer. The filter of this layer can be used to detect the
combination of bottom features, such as semicircle and
x, x > 0, quadrilateral and these information correspond to the color,
g(x) � 􏼨 (9) edge, contour, and other features of the image. The third
tanh(βx), x ≤ 0,
layer is the image texture feature. The fourth layer of learning
where β is a variable parameter used to control the unsat- obtains more distinctive features, which reflect the differ-
urated region of the function. ences between classes. The fifth layer of learning obtains
Computational Intelligence and Neuroscience 5

10 are arranged from large to small according to the

eigenvalues, and the first q eigenvectors are selected
8
according to the accuracy to form the principal
component matrix Mq .
(6) The dimension of the output of the fifth pool layer is
6
reduced, and the feature after dimension reduction is
F 5 � C · Mp . F 5 is the third high-level feature of
y 4 remote sensing image.
After using PCA to reduce the dimension of deep fea-
2 tures, the enhanced deep learning features are used to train
an SVM model. A multiclass SVM is defined as follows:
0 N
1�� ��2
min ��ω′ �� + τ 􏽘 sa ,
ω,b,s 2
-2
a�1 (10)
-10 -5 0 5 10
x s.t. δ ω′ κ za 􏼁 + b􏼁 ≥ 1 − sa , sa ≥ 0, a � 1, 2, . . . , N,
Figure 3: TReLU function. where ω′ is the projection of multiclass SVM model; τ is the
penalty parameter, set it to 0.01; sa is a nonnegative relax-
complete and discriminative key features, which are a class ation variable; za is the enhanced feature after PCA di-
of objects with significant differences in remote sensing mensionality reduction. The depth feature of CNN is further
images. Finally, the output result of CNN layer 5 pooling enhanced, then the feature will be used to continue the
layer can be obtained, which includes all the characteristic training of the SVM classifier, and the trained SVM classifier
diagrams calculated by layer 5 pooling layer [26]. is tested with the test set.
Then, using equation (3), the output results F 6 and F 7
of CNN layer 6 and 7 full connection layers FC 6 and FC 7 4. Experiment and Analysis
can be obtained, including all the characteristic diagrams
calculated by FC 6 and FC 7 layers. Finally, F 6 and F 7 are 4.1. Experimental Environment. This experiment is based on
two different high-level features of remote sensing images. a Tensorflow framework. Tensorflow is a powerful visuali-
zation suite of low-level and high-level interfaces (Tensor-
board) for huge and active community network training. It
3.4. Feature Dimensionality Reduction and Classification. can track network topology and performance, making
For the output of the fifth pool layer of CNN, the principal debugging easier and more convenient. The specific ex-
component analysis (PCA) method is used to reduce the perimental environment is listed in Table 1. On the Ubuntu
dimension, and the reduced dimension result is used as the 16.04 operating system, some dependent libraries are first
third high-level feature of remote sensing image. PCA di- installed, such as Python and open CV, and then the python
mensionality reduction process is as follows: environment and Tensorflow are installed. In addition, the
(1) Matrix Deformation. The output result of the fifth graphics processing unit (GPU) mode is used. After the
pool layer of CNN is transformed into a two-di- environment is configured, the network is built according to
mensional matrix C, and each row of the two-di- the designed network structure, including network struc-
mensional matrix reflects the feature vector ture, convolution kernel size, step length, and the number of
corresponding to a remote sensing training sample. feature maps of each layer.
(2) Zero Mean. Each column of two-dimensional matrix
C is zeroed to obtain a new matrix C0 , and the
4.2. Evaluating Indicator. The evaluation indexes include
average value of each column of C0 is 0.
classification accuracy Acc, misclassification error, and
(3) Covariance Matrix M0 of C0 . Covariance matrix M0 kappa coefficient. The classification accuracy and misclas-
reflects the relationship between the feature vectors sification rate are calculated as follows:
of all remote sensing image training samples: if the
covariance of the two sample feature vectors is not 0, TP
it indicates that there is correlation between them; Acc � ,
Num
otherwise, it means irrelevant. The final calculated (11)
covariance matrix M0 is a diagonal matrix with size FP
Error � 1 − Acc � ,
d × d. Num
(4) Find d eigenvalues and d eigenvectors of covariance where TP is the number of correctly classified images in the
matrix M0 . remote sensing image test set, Num represents the total
(5) Select the principal component of covariance matrix number of images, and FP is the number of incorrectly
M0 . The eigenvectors of the covariance matrix M0 classified images.
6 Computational Intelligence and Neuroscience

Table 1: System experimental environment parameters. 1.0 5

Environment Parameter setting
Operating system Ubuntu16.04 0.8 4
GPU GTX TITAN X (12G)
CPU Intel E5-2600 v3

Training accuracy

Training loss
Deep learning framework Tensorflow 0.6 3
Memory 32G
Computer language Python 3.6
0.4 2

Assuming that the actual number of samples is

0.2 1
c1 , c2 , . . . , cN and the predicted number is η1 , η2 , . . . , ηN , the
kappa coefficient is defined as follows:
Acc − Pe 􏼁 0.0
RKappa � , 10000 20000 30000 40000 50000
1 − Pe 􏼁 Number of iterations
(12)
c η + c2 η2 + · · · + cN ηN Training accuracy
Pe � 1 1 , Training loss
Num × Num
Figure 4: Model training accuracy and loss function.
where Acc is the actual accuracy and Pe is the theoretical
accuracy. The higher the kappa coefficient, the better the
overall classification accuracy of the method.

4.3. Model Training. The training set and verification set

after PCA dimensionality reduction are used to train the
improved CNN model. With the increase of the number of
iterations, the variation trend of the training accuracy of the
model and the loss function value of the training objective
function are shown in Figure 4.
As can be seen from Figure 4, with the increase of it-
erations, the training accuracy of CNN model gradually
tends to be stable, 93% of the training accuracy can be
obtained at the highest, and the training loss gradually
decreases and tends to be flat, indicating that the model has
good convergence. Cultivated land Forest
Road Others
Residential area
4.4. Land Classification Result Map. On the test set samples,
the proposed method is used to extract five types of land Figure 5: Land classification results.
resources. The proposed method extracts three high-level
features from remote sensing images and fuses them to
generate the final classification map. The results are shown in As can be seen from Figure 6(a), cultivated land has
Figure 5. regular graphics and clear edges in the image and accounts
As can be seen from Figure 5, the five land resource use for a very large proportion of the whole image. Reference
types are clearly identified, especially residential areas, which [12] adopts the traditional fuzzy and decision tree classifi-
are relatively scattered and irregular, but the location of cation, which can identify a large area of land types, but the
residential areas can be clearly seen in the identification identification effect of small land resources is poor, and the
result figure. The distribution of roads and cultivated land is misclassification phenomenon is obvious. Reference [15]
very regular, and the overall recognition effect is ideal. combines SVM and k-nearest neighbor to complete culti-
vated land recognition. Because it is not suitable for pro-
4.5. Comparison with Other Methods cessing complex remote sensing images, there are many
missing points at the edge, and there are many missing
4.5.1. Comparison of Cultivated Land Classification Results. points in light-colored cultivated land. In reference [17], the
In order to more intuitively evaluate the performance of the depth CNN model is used to identify the cultivated land
proposed method in cultivated land recognition, it is image. The most important cultivated land position is
compared with the recognition results obtained by the extracted accurately, but there is a case of wrong points and
methods used in references [12, 15, 17]. In the experiment, missing points. The proposed method can better identify the
the trained model is applied to the Gaofen-5 for recognition, cultivated land image, and the contour is clear, which is
and the results are shown in Figure 6. better than other comparison methods.
Computational Intelligence and Neuroscience 7

(a)

(b)

(c)

(d)

(e)

Figure 6: Comparison of cultivated land classification results. (a) Original image. (b) Reference [12]. (c) Reference [15]. (d) Reference [17].
(e) Proposed model.

4.5.2. Comparison of Evaluation Indicator. The performance Table 2: Evaluation indicator values of four methods.
of the four classification methods is quantitatively analyzed.
Reference Reference Reference Proposed
The classification accuracy Acc, misclassification error, and [12] [15] [17] method
kappa coefficient are listed in Table 2.
Acc 0.8655 0.8907 0.9286 0.9472
It can be seen from Table 2 that the classification ac-
Error 0.1345 0.1093 0.0714 0.0528
curacy, misclassification rate, and kappa coefficient of the RKappa 0.8629 0.8839 0.9193 0.9435
proposed method are 0.9472, 0.0528, and 0.9435, respec-
tively, which are better than other comparison methods. The
proposed method adopts the 7-layer CNN network struc-
ture, improves the activation function, reduces the di- network for image recognition based on the depth CNN
mension by PCA, and improves the classification accuracy. model, but it has not been optimized in terms of dimen-
Reference [17] proposed a dense coordinate transformation sionality reduction and activation function. Compared with
8 Computational Intelligence and Neuroscience

important role in the rational development of land resources

2.0 and the development of precision agriculture. Therefore, a
land resource use classification method based on deep
learning in ecological remote sensing images is proposed.
1.6
The remote sensing image samples are input into the seven-
layer CNN model. The activation function of the model
Time consuming (s)

1.2 adopts the TReLU function, and the three high-level image
features are fused in series and then input into the SVM
classifier to complete the classification of land resources
0.8 remote sensing images. The remote sensing images of
Pingshuo mining area in Shanxi Province are used to analyze
0.4 the proposed method. The results show that the improved
CNN model can achieve rapid convergence, and the image
edges recognized by the proposed method are clear. The Acc,
0.0 error, and kappa coefficients are 0.9472, 0.0528, and 0.9435,
Ref.[12] Ref.[15] Ref.[17] Proposed
method respectively, and the training and testing times are 1.8 s and
Training time 0.95 s, respectively. The overall performance is better than
Testing time other comparison methods.
Remote sensing images often contain complex geometric
Figure 7: Training and testing time of different methods.
and semantic information. The next research work needs to
consider not only the semantic information contained in the
the proposed method, the recognition accuracy is reduced by image itself but also some more complex factors such as
0.0186. Reference [15] combines SVM and k-nearest occlusion, blur, and distortion. In addition, in terms of data
neighbor for remote sensing image classification. For amplification, the subsequent work can consider using the
complex land resource use types, the classification perfor- GAN model to generate some data with the same distri-
mance is poor, and the kappa coefficient is only 0.8839. The bution as the real remote sensing image, so as to meet the
method used in Reference [12] is more traditional, so the requirements of the deep learning model for a large amount
classification effect is not ideal. of training data.

Data Availability
4.5.3. Comparison of Training and Testing Time.
Classification efficiency is another important indicator of land The data included in this paper are available without any
resource use classification. The time consumption of the four restriction.
methods on the training set and test set is shown in Figure 7.
As can be seen from Figure 7, the proposed method takes Conflicts of Interest
the longest time in the training phase, which is 1.95 s. This is
because the method used in Reference [12] is relatively The authors declare that they have no conflicts of interest.
simple. The training stage of reference [15] includes only the
training of k-nearest neighbor model, whereas the training Acknowledgments
stage of reference [17] includes the training of convolutional
neural network and the process of feature extraction. The This article was supported by National Natural Science
training stage of the proposed method includes not only the Foundation of China (no. 71702015), Social Science Plan-
training of convolutional neural network and the extraction ning Major Application Project in Chongqing (no.
of three depth features but also the fusion of three depth 2017ZDYY51), Fundamental Science and Frontier Tech-
features. In addition, in the test stage, due to the simple nology Research Project in Chongqing (no. cstc2017jcy-
calculation in reference [12], the test time is only 0.72 s. The jAX0130), Humanities and Social Sciences Research
methods used in reference [14, 17] are complex, and the test Program of Chongqing Education Commission (no.
time is more than 1.2 s. After training, the proposed method 19SKGH078), Science and Technology Research Projects of
has the best performance in the test, and after PCA di- Chongqing Education Commission (no. KJQN201900812),
mensionality reduction, the calculation speed is faster, and and Artificial Intelligence Application Collaborative Inno-
the test time is about 0.95 s. Overall, the proposed method vation Center of Chongqing Business Vocational College
has the best overall performance and has certain practica- and AI trainer Master Studio of Chongqing Business Vo-
bility in the application of land resource use classification. cational College.

References
5. Conclusion
[1] A. I. R. Cabral, S. Silva, P. C. Silva, L. Vanneschi, and
Using a deep learning model to segment and extract eco- M. J. Vasconcelos, “Burned area estimations derived from
logical remote sensing images can obtain high-precision landsat ETM+ and OLI data: comparing genetic program-
land use classification information, which plays an ming with maximum likelihood and classification and
Computational Intelligence and Neuroscience 9

regression trees,” ISPRS Journal of Photogrammetry and Re- of spatial features,” Sensors, vol. 19, no. 12, pp. 2792–2803,
mote Sensing, vol. 142, no. 8, pp. 94–105, 2018. 2019.
[2] S. Kaplan, “Identification of genetic markers related to milk [18] S. Dong, Y. Zhuang, Z. Yang, L. Pang, H. Chen, and T. Long,
fat in anatolian buffaloes,” Fresenius Environmental Bulletin, “Land cover classification from VHR optical remote sensing
vol. 29, no. 7, pp. 5786–5791, 2020. images by feature ensemble deep learning network,” IEEE
[3] M. A. Khan, S. Saeed, and N. Ullah, “Effect of nickel on the Geoscience and Remote Sensing Letters, vol. 17, no. 8,
germination and biochemical parameters of two rice varie- pp. 1396–1400, 2020.
ties,” Fresenius Environmental Bulletin, vol. 29, no. 2, [19] L. Ghayour, A. Neshat, S. Paryani, H. Shahabi, and A. Ahmad,
pp. 956–963, 2020. “Performance evaluation of sentinel-2 and landsat 8 OLI data
[4] K. Kundu, P. Halder, and J. K. Mandal, “Urban change de- for land cover/use classification using a comparison between
tection analysis during 1978–2017 at Kolkata, India, using machine learning algorithms,” Remote Sensing, vol. 13,
multi-temporal satellite data,” Journal of the Indian Society of no. 1349, pp. 1–23, 2021.
Remote Sensing, vol. 48, no. 11, pp. 1535–1554, 2020. [20] A. M. El-Tantawi, A. Bao, C. Chang, and Y. Liu, “Monitoring
[5] A. Zafar, Z. I. Khan, K. Ahmad, M. Nadeem, and H. Bashir, and predicting land use/cover changes in the Aksu-Tarim
“Appraisal of chromium contents in wheat grains irrigated River Basin, Xinjiang-China (1990–2030),” Environmental
with wastewater,” Fresenius Environmental Bulletin, vol. 29, Monitoring and Assessment, vol. 191, no. 8, pp. 1–18, 2019.
no. 5, pp. 3894–3904, 2020. [21] W. Li, Z. Li, J. Sun et al., “Spear and shield: attack and de-
[6] R. D. Kangabam, M. Selvaraj, and M. Govindaraju, “Spatio- tection for CNN-based high spatial resolution remote sensing
temporal analysis of floating islands and their behavioral images identification,” IEEE Access, vol. 7, no. 99,
changes in Loktak Lake with respect to biodiversity using pp. 94583–94592, 2019.
remote sensing and GIS techniques,” Environmental Moni- [22] P. Zhang and S. Hu, “Fine crop classification by remote
toring and Assessment, vol. 190, no. 3, pp. 118–214, 2018. sensing in complex planting areas based on field parcel,”
[7] A. H. Anbar, T. Antary, J. Sawwan, H. Khawaldah, and Nongye Gongcheng Xuebao/Transactions of the Chinese Society
M. Abu-Dalhoum, “Changing rainfall trends and the impact of Agricultural Engineering, vol. 35, no. 21, pp. 125–134, 2019.
on cereal farming in Jordan. Fresenius environmental bul- [23] A. R. Udgata, P. M. Sahoo, T. Ahmad, A. Rai, and G. Krishna,
letin,” vol. 29, no. 12, pp. 10980–10996, 2020. “Remote Sensing and Machine Learning techniques for
[8] F. Li, Q. Qin, H. Wang, X. F. Hu, and H. Zhao, “Extraction of acreage estimation of mango (Mangifera indica),” Indian
planting information of winter wheat in a Province based on Journal of Agricultural Sciences, vol. 90, no. 3, pp. 551–555,
GF-1/WFV images,” Meteorological and Environmental Re- 2020.
search, vol. 9, no. 4, pp. 104–109, 2018. [24] J. Mallick, S. Alqadhi, S. Talukdar, B. Pradhan, and
[9] W. Liu, “Coastal land use planning and beach sports image A. S. Dajam, “A novel technique for modeling ecosystem
recognition based on high-resolution remote sensing images,” health condition: a case study in Saudi arabia,” Remote
Arabian Journal of Geosciences, vol. 14, no. 11, pp. 1–14, 2021. Sensing, vol. 13, no. 13, pp. 1–20, 2021.
[10] J. Bofana, M. Zhang, M. Nabil, B. Wu, and C. Moyo, [25] C. Sun, Y. Bian, T. Zhou, and J. Pan, “Using of multi-source
“Comparison of different cropland classification methods and multi-temporal remote sensing data improves crop-type
under diversified agroecological conditions in the zambezi mapping in the subtropical agriculture region,” Sensors,
River Basin,” Remote Sensing, vol. 12, no. 13, pp. 1–23, 2020. vol. 19, no. 10, pp. 2401–2413, 2019.
[11] S. Liao and Y. Yue, “Method of improving classification ac- [26] R. M. V. Malladi, A. Nizami, M. S. Mahakali, and
curacy of land cover based on time series NDVI database,” B. G. Krishna, “Cloud masking technique for high-resolution
Transactions of the Chinese Society of Agricultural Engineering, satellite data: an artificial neural network classifier using
vol. 34, no. 7, pp. 241–248, 2018. spectral & textural context,” Journal of the Indian Society of
[12] H. Li, J. Wu, and X. Wang, “Object oriented land use clas- Remote Sensing, vol. 47, no. 4, pp. 661–670, 2019.
sification of Dongjiang River Basin based on GF-1 image,”
Nongye Gongcheng Xuebao/Transactions of the Chinese Society
of Agricultural Engineering, vol. 34, no. 10, pp. 245–252, 2018.
[13] A. Shakya, M. Biswas, and M. Pal, “Parametric study of
convolutional neural network based remote sensing image
classification,” International Journal of Remote Sensing,
vol. 42, no. 7, pp. 2663–2685, 2021.
[14] Y. Boualleg, M. Farah, and I. R. Farah, “Remote sensing scene
classification using convolutional features and deep forest
classifier,” IEEE Geoscience and Remote Sensing Letters,
vol. 16, no. 99, pp. 1944–1948, 2019.
[15] G. Alimjan, T. Sun, Y. Liang, H. Jumahun, and Y. Guan, “A
new technique for remote sensing image classification based
on combinatorial algorithm of SVM and KNN,” International
Journal of Pattern Recognition and Artificial Intelligence,
vol. 32, no. 7, Article ID 1859012, 2018.
[16] R. Ķ�eniņš, “Land cover classification using very high spatial
resolution remote sensing data and deep learning,” Latvian
Journal of Physics and Technical Sciences, vol. 57, no. 1-2,
pp. 71–77, 2020.
[17] X. Yao, H. Yang, Y. Wu et al., “Land use classification of the
deep convolutional neural network method reducing the loss