IET Communications

Special Section: UAV-Enabled Mobile Edge Computing

Accurate landslide detection leveraging UAV- ISSN 1751-8628

Received on 28th October 2019

based aerial remote sensing

Revised 7th February 2020
Accepted on 12th March 2020
E-First on 21st May 2020
doi: 10.1049/iet-com.2019.1115
www.ietdl.org

Shanjing Chen1, Chaocan Xiang2,3 , Qing Kang1, Wei Zhong1, Yanlin Zhou2, Kai Liu2
1Department of Military Facilities, Army Logistics University of PLA, Chongqing, 401311, People's Republic of China
2College of Computer Science and State Key Laboratory of Mechanical Transmission, Chongqing University, Chongqing 400044, People's
Republic of China
3Key Laboratory of Dependable Service Computing in Cyber Physical Society (Chongqing University), Ministry of Education, People's Republic

of China
E-mail: [email protected]

Abstract: Remote sensing by unmanned aerial vehicles (UAVs) is significantly important in emergency rescue applications and
operations. Particularly, the on-site images from UAVs can provide valuable information for hazard identification and disaster
assessment. In this study, the authors propose a novel method by using back propagation neural networks with feature fusion to
detect landslides from UAV images. Specifically, the authors first construct a fundamental shape model of landslides and devise
a scale-invariant feature transform algorithm for feature matching and transformation. By fusing the spatial shape features and
spectral features of the landslide, the suspected landslide object from UAV images can be detected initially. Next, the change
features of a pre/post-landslide object are extracted by using the satellite sensing images (before landslide) and the UAV image
(after landslide). The authors further feed the change features into the proposed model to enhance the precision and accuracy
of landslide detection. They conduct numerous experimental studies with aerospace remote sensing data in two real-world
landslide scenarios. The evaluation results show that the proposed method outperforms baseline algorithms by achieving over
91% accuracy in landslide detection.

1 Introduction UAVs for the emergency surveys of individual geological hazard.

In this work, they arranged the ground control points in the region
In remote sensing, unmanned aerial vehicle (UAV) platforms have of interest, designed the flight plan, and acquired remote sensing
tremendous advantages over satellites, including high flexibility, images by a UAV. According to the processing results on-the-spot,
handy operation, high-resolution sensing and low cost etc. an individual landslide can be recognised within a wide range. The
Particularly, UAVs are crucially important for many remote sensing practical test showed that their method for the emergency survey of
applications, including geological disaster detection, disaster individual Geo-hazard can be useful and capable of achieving fast
analysis, urban crowdsensing [1–3], and disaster emergency rescue and accurate detection on landslides. Wang [8] used data
operation. For example, the Southwest of China is a mountainous processing techniques with remote sensing, GPS, 3D laser
region where the probability of occurrence of landslides is high. scanning, GIS and other ground observations to monitor, analyse,
Deploying UAVs to patrol and monitor important areas in the and forecast the landslide disasters in the mining area. Targeting
above region can essentially improve the success rate of early the landslide deformation monitoring in the mine area, a
landslide detection and gain more time for rescue. Moreover, monitoring system combining GPS and 3D laser scanning was
equipped with multi-type optical equipment for remote sensing, proposed. An early risk warning model of Sina Map and a
UAVs can take multi-spectral, multi-angle and high-resolution weighted risk factor decision model were further established to
photos for landslide disasters, providing significant information for achieve multi-source monitoring data fusion and key feature
disaster assessment, rescue decision making and post-disaster extraction for surface hazard detection in mining areas.
reconstruction [4]. Chen [9] combined the GPS data of a landslide area with
In recent years, with the rapid development of aerospace images from a UAV with a mini single-lens reflex camera to obtain
technology and imaging technology, many researchers have the high-resolution DEM and DOM images that greatly preserved
presented different methods to detect landslides over a large area, the rich spectral and texture information. They further conducted
by using high-resolution, multi-band, multi-platform and multi- object-oriented classification and information extraction by
temporal remote sensing data. For example, Liu [5] proposed a decision tree based on the fuzzy classification and support vector
method of Geo-hazards monitoring based on UAV and geographic machine (SVM) algorithm. According to the spatial features and
information system (GIS) data. They combined remote sensing shape of the target area, their method can identify the high-risk
data collected by UAVs with GIS spatial information to extract areas for landslides. Tang [10] proposed a method for landslide
hazards information, then further used methods of radiometric emergency monitoring with the remote sensing data of UAVs and
correction and image mosaic to determine the area and volume of a domestic high-resolution optical imaging satellites. They utilised
landslide. Moreover, Huang [6] proposed an interpretation method the UAV data and satellite data of the disaster area before and after
on geological mine hazards based on remote sensing data acquired 6.24 Maoxian landslide, then further combined the information of
by UAVs. By combining the characteristics of the landslide on the landslide accident and disaster situation, and other empirical
colour, they presented the interpretation signs of a landslide with a information. The affected areas around the four suspected disaster
digital elevation model (DEM) and a three-dimensional (3D) locations were identified at the initial stage. Fu [11] proposed an
model based on the orthophotos from UAVs. This interpretation information extracting method of landslides caused by earthquake
result of the landslide by UAV remote sensing have a wide range of based on high-resolution UAV images. They established the sample
applications in monitoring geological disasters of mines. In library of earthquake landslides based on UAV images, then
addition, Huang [7] further developed an application of micro introduced the transfer learning mechanism to carry out the

IET Commun., 2020, Vol. 14 Iss. 15, pp. 2434-2441 2434

© The Institution of Engineering and Technology 2020
 17518636, 2020, 15, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-com.2019.1115, Wiley Online Library on [22/08/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
earthquake landslide information extraction from UAV images. assess dustfall distribution in the city of Beijing. Reference [26, 27]
This method satisfies the requirements of earthquake emergency proposed a method based on the integration of BPNNs and genetic
response such as macroscopic investigation of landslide disasters, algorithm, to map the swampland from multi-spectral remote
and monitoring of landslide disaster body. sensing imagery. Prasad [28] employed a BP artificial neural
Koschitzki [12] proposed a multi-temporal photogrammetric network (BPANN) to retrieve the crop parameters like biomass,
method to identify the landslide in Las Minas Basin by processing leaf area index, average plant height, and soil moisture content by
aerial images of UAVs. This method established a model of major using the X-band scattering coefficients at different growth stages
geomorphological changes based on multi-temporal aerial images of this crop, which was helpful in monitoring and estimating the
of the river, Las Minas, from 2014 and 2015, and then through yield of useful and economical crops. Chen [29] established a
analying the height differences, it can automatically detect and BPANN to retrieve land surface emissivity, land surface
monitor landslides in the river valley. Hsu [13] used high- temperature, atmospheric transmittance, from the hyperspectral
resolution aerial photography, terrain models, landslide inventories thermal infrared data. The experiment showed that, without giving
to identify the landslide hazard-prone sections along the South any prior knowledge, the proposed BPANN can achieve reasonable
Link Railway in Taiwan. Giordan [14] proposed a methodology for results under the different land surface and atmospheric conditions.
landslide hazard information acquisition and processing that In general, the application of optical remote sensing, SAR and
considers the landslide typology with remote sensing data of UAVs LiDAR on ground object information acquisition and target
in the northwest part of Italy. Nicola [15] introduced a new recognition is quite mature. The fusion of various remote sensing
approach for rapid detection of newly triggered landslides by an data, the application of basic geographic information, and the
object-oriented change detection technique. The digital terrain design of excellent data processing algorithms are the important
models, acquired by spaceborne optical very high resolution data directions for object recognition in remote sensing imagery. The
and synthetic-aperture radar (SAR) data for pre- and post-disaster, precise mapping of landslides requires highly precise data from
were compared to detect any morphological change between the remote sensing platforms, sensors, and basic geographic
three acquisitions and further characterise the landslides. Their information. However, landslide detection for disaster
experiments demonstrated this method was effective for post- identification and extraction can be carried out based on a large
disaster damage assessment, landslide detection and rapid amount of general aerospace remote sensing data, which is more
mapping, and the definition of states of activity and updating of time-effective and more useful in hazard emergency response.
landslide inventory maps. Recently, UAVs have been widely used in all walks of life [30].
Saito [16] studied the method of landslide detection by using A large number of references also show that it is feasible to apply
the advantages of the high-definition data, repeatability and the technology of remote sensing with UAVs to landslide disaster
mobility provided by the unmanned aerial system and the structure- identification and information extraction. The optical remote
from-motion multi-view stereo photogrammetry. They obtained sensing images of UAVs have provided large amounts of high-
orthorectified images and digital surface models with a spatial resolution data. How to automatically identify landslides from
resolution of 0.06 m before and after the 2016 Kumamoto UAV remote sensing images is particularly important in disaster
earthquake. Then, by determining the differences between the data rescue applications.
obtained in March and May 2016, which corresponded to periods The BP neural network is a widely accepted learning machine
before and after the earthquake, respectively, they detected 54 model especially for classification and object detection from
coseismic landslides with volumes ranging from 9.1 to 3994.6 m3. remote sensing imagery. In addition, multi-feature fusion based on
Li [17] proposed a landslide hazard detection method based on BP neural network is still in the process of improvement. Multi-
bag-of-words for UAV images. The bag-of-words features and information fusion between target and background, efficient and
SVM were applied into the landslide interpretation model to accurate architecture design, and transfer function selection are
achieve rapid localisation and detection of mountain landslides always important directions of neural network algorithm
with large-scale and high-resolution images. Li [18] proposed a improvement. Therefore, according to the requirements of disaster
semi-active landslide hazard area mapping method based on aerial detection and information extraction from UAV remote sensing
remote sensing images. This method can obtain difference images data in emergency rescue operations, we propose a landslide
by analysing the change vectors and extract the candidate landslide detection method based on aerial remote sensing data of UAVs,
hazard region by threshold segmentation. With eliminated errors by which includes recognition of typical spectral and spatial shape
a morphological method, the landslide region can be accurately features of a landslide, change features of the imageries before and
detected by edge grading evolution and regional grading evolution. after the landslide, and an intelligent interpretation model of BP
Chen [19] proposed an automatic landslide detection method by neural network. Our method will play an important role in
data fusion technique using wavelet transform and Taguchi automatic geological hazard recognition, disaster prevention and
methods. The DEM data extracted by airborne LiDAR and images damage mitigation.
acquired by QuickBird satellite were fused by wavelet Our main contributions are as follows:
transformation. The accurate detection of landslides in tropical
urban areas was achieved through object-oriented classification and • First, we design a model of multi-layer BP neural networks to
parameter optimisation by the Taguchi method. Fey [20] combined detect landslides according to the object's spectral features in the
historic topographic data, aerial laser scanning data, and airborne high-resolution, multi-band remote sensing images from
orthoimages, to analyse and monitor the rockslide deformation UAVs, which can quickly extract the suspected landslide object.
information in the Eastern Alps. Mezaal [21] proposed recurrent • Second, by analysing and summarising the shape characteristics
neural networks to use LiDAR-derived data, orthophotos, and of typical landslides, we establish the fundamental shape models,
textural features to detect landslides, and found them to be and further apply a scale-invariant feature transform (SIFT)
effective in combining these features. Sameen and Pradhan [22] algorithm for feature matching and transformation. By fusing
designed residual networks for landslide detection using spectral spatial shape features and spectral features, the landslide can be
(RGB bands) and topographic information (altitude, slope, aspect, detected initially.
curvature), which enhanced the accuracy of landslide detection by • Third, we extract the change features of the landslide area based
traditional CNN. on the satellite remote sensing images before the landslide and the
In terms of the back propagation neural network (BPNN), Liu UAV remote sensing images after it. By combining the change
[23] proposed BPNNs to identify the non-stationary relationship of features, the landslide shape features and the initial detection result
drought conditions based on the remote sensing data. Xie [24] of BP neural networks, our method can achieve accurate
proposed an approach for classification of building type using soft identification and localisation of a landslide. The experimental
BP network, which recursively defuzzified unsure (fuzzy) results based on two real-world landslide scenarios show that the
members according to the information carried by the gradually proposed method outperforms baseline algorithms by achieving
increased sure members. Yan [25] proposed a BPNNs model using over 91% accuracy in landslide detection.
spectral response functions and integrated remote sensing data to

IET Commun., 2020, Vol. 14 Iss. 15, pp. 2434-2441 2435

© The Institution of Engineering and Technology 2020
 17518636, 2020, 15, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-com.2019.1115, Wiley Online Library on [22/08/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Fig. 1 Landslide detection model based on BPNNs
(a) Basic neuron of a BP network, (b) Structure of BPNNs

2 Landslide detection model with BP neural

networks
In this work, we propose a landslide detection method based on BP
neural networks with spectral features. The error BP network is one
of the most widely used in object detection and image
classification. It is a classical machine learning algorithm for
information extraction from post-disaster remote sensing data. BP
neural networks can approximate arbitrary non-linear mapping
relations with good generalisation ability [31]. This study proposes
a landslide detection method based on BP neural networks by
spectral features.
Fig. 2 Fundamental shape models of landslides
(a) Dustpan shape, (b) Leaf shape
2.1 Neuron model of BP network
The basic BP neuron model structure is constructed based on dimensions of remote sensing image. The number of neurons in the
spectral features of remote sensing data. The input data has R first layer is eight, and the transfer function is logsig. The number
dimensions which are the same as the number of spectral of neurons in the second layer is four, and the transfer function is
dimensions in a remote sensing image. Each input is given a tansig. The number of neurons in the third layer is eight, and the
corresponding weight of w. The output of the neuron's transfer transfer function is tansig. The number of neurons in the output
function is composed of a weighted sum of R-dimensional inputs layer is one, and the transfer function is purelin.
and deviation. The basic structure of BP network neurons is shown
in Fig. 1a. 3 Spatial feature fusion for landslide detection
In order to utilise the spatial shape features of a target landslide for
2.2 Transfer function model detection, a SIFT algorithm is applied for feature matching and
In a BP network, the transfer function is significantly important. similarity detection based on landslide detection results by BP
Typical transmission functions include logsig function, tansig neural networks. In this way, accurate landslide classification and
function, purelin function, and exponential linear function. Each of recognition can be achieved by fusing spatial shape features.
the above functions can be chosen for BP neural networks and we
introduce them as follows: * The logsig function is denoted by 3.1 Spatial shape models of typical landslide

1 The surface vegetation of a landslide area will be subject to

ϕ(x) = . (1) severely damage, as the newly exposed soil and rock are
1 + e− x
significantly different from the surrounding vegetation coverage in
terms of brightness, hue and spatial shape. In this work, we select
The tansig function is denoted by two common basic shapes (dustpan shape and leaf shape) of
landslides to construct the fundamental shape models, which are
1 − e−2x shown in Fig. 2.
ϕ(x) = tanh(x) = . (2)
1 + e−2x
3.2 Feature detection and matching based on SIFT algorithm
The purelin function is denoted by
The detection results of BP neural networks are matched with the
ϕ(x) = x . (3) fundamental feature models of landslides by the SIFT algorithm.
The feature points extracted by the SIFT algorithm have scaling
The exponential linear function is denoted by invariance, rotation invariance, and affine invariance. These feature
points also have certain features in anti-illumination and anti-
α ex − 1 , x < 0 viewpoint conversion [32]. Based on the SIFT algorithm, feature
ϕ(x) = . (4) point matching can be divided into five steps: (i) scale space
x, x ≥ 0 extrema extraction, (ii) feature point location, (iii) feature direction
assignment, (iv) feature description extraction and (v) feature
2.3 The structure of BP neural networks matching. For a two-dimensional image I(x, y), ∗ is convolution
A detection model based on BP neural networks is constructed operation. The Gaussian scale space is defined as L(x, y, t), which
according to the requirements of landslide identification and can be computed by convolution of Gaussian kernel G(x, y, t) and
summary of experimental results. The model consists of an input I(x, y) as
layer, an output layer, and three hidden layers, as shown in Fig. 1b.
The dimensions of the input layer are the same as the spectral

2436 IET Commun., 2020, Vol. 14 Iss. 15, pp. 2434-2441

© The Institution of Engineering and Technology 2020
 17518636, 2020, 15, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-com.2019.1115, Wiley Online Library on [22/08/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
G(x, y, t) ∗ I(x, y) t > 0
L(x, y, t) = , (5)
I(x, y) t=0

1 −(x2 + y2)/2t
G(x, y, t) = e , (6)
2πt

where x and y are the pixel position coordinates of L(x, y, t). Here, t
is the scale space factor, which is the variance of the Gaussian
normal distribution that represents the smoothing degree of the
image.
To efficiently detect stable feature points in the scale space,
extrema for difference of Gaussian (DOG) in the scale space is
used as a judgment basis. The DOG operator is defined as follows:

D(x, y, t) = (G(x, y, kt) − G(x, y, t)) ∗ I(x, y)

(7) Fig. 3 The overall framework of landslide detection
= L(x, y, kt) − L(x, y, t)

where k is the scale space rate of two adjacent Gaussian space. the 4 Change feature extraction based on remote
DOG local extreme value is detected when the pyramid sensing images
construction is completed. The extrema point is found and saved as
In order to eliminate the influence of other areas with surface
the key point, by comparing each pixel with 8 pixels in its
damage on the accuracy of landslide identification, change feature
surrounding neighbourhood of the same scale and 18 pixels in the
extraction is carried out by using the satellite remote sensing
corresponding positions of different scale. The gradient model and
images before landslide and the UAV remote sensing images after
direction of each point in the Gauss image are calculated by the
landslide, and then the recognition results of neural networks and
gradient distribution of neighbourhood pixels as (see (8)) . Then,
shape features are combined with change features of the landslide
the direction of a key point is determined by
area, which can further improve the accuracy of landslide
L(x, y + 1) − L(x, y − 1) identification. According to the spectral difference characteristics
θ(x, y) = tan−1 . (9) of pre-/post-landslide, a probability model of multi-spectral change
L(x + 1, y) − L(x − 1, y)
detection of landslides is constructed.
As
The gradient model and direction of the key points
corresponding to the landslide detection results of BP neural
Ipst, R − Ibf , R + Ipst, R − Ipst, B − Ibf , R − Ibf , B
networks and the typical landslide shape model are extracted by the If = , (11)
SIFT algorithm, respectively. The key points of optimal matching max Ipst, R − Ibf , R + Ipst, R − Ipst, B − Ibf , R − Ibf , B
are found by Two-Dimensional norms, and the minimum Two-
Dimensional norm point is selected as the optimal key matching where I f is the normalised probability model of multi-spectral
point for feature matching between the landslide typical shape change detection of landslides. The larger I f is, the greater the
model and neural networks' detection results. probability that the sensing object is a landslide is, and vice versa.
Ibf , R and Ibf , B are R component images and B component images of
3.3 Scale space transformation and similarity detection pre-disaster multi-spectral images, respectively. Ipst, R and Ipst, B are
The optimal key feature matching points of the typical shape model the Rth component images and Bth component images of post-
and landslide detection results by BP neural networks are used to disaster multi-spectral images, respectively.
calculate scale transformation rate and rotation angle. The
fundamental landslide shape model is transformed into a target 5 The framework of neural network-based
template corresponding to the BP neural networks' landslide landslide detection with feature fusion
detection results by scale and rotation transformation. The
correlation detection method (10) is used to detect landslides based In order to make the best use of spectral and spatial shape features
on the results of BP neural networks' detection, which realises the of the remote sensing images, the landslide detection model is
secondary utilisation of spatial shape features and further landslide established based on BP neural networks and typical landslide
detection by spatial shape features. shape feature model after airborne and spaceborne remote sensing
data preprocessing and registration. BP neural networks are trained
m n by typical landslide spectral samples and background spectral
∑ ∑ T(x, y) − T¯ × I(x, y) − I¯ samples. The trained neural networks are used to detect the
ρ(x, y) =
x=1y=1
, (10) landslide object. Then, the Otsu [33] method is used to segment the
m n
2
∑ ∑ T(x, y) − T¯ × ∑ ∑ I(x, y) − I¯
m n
2 results of neural networks. Feature detection and matching using
x=1y=1 x=1y=1 fundamental shape models are done. A specific target template
corresponding to the landslide detection results is formed through
where ρ(x, y) is the correlation coefficient of point (x,y); T(x, y) is scale and rotation transformation from the fundamental shape
the DN value of the fundamental shape model at point (x,y); T¯ is models. With this template, landslide objects are detected initially.
the mean DN value of the fundamental shape model; I is the Based on the background image of the pre-landslide area, the
suspected object with optimal feature matching in neural networks' difference method is used to construct the probability model of
detection results; I(x, y) is DN value of the suspected object at change features. Combined with this model, landslide targets are
point (x,y); I¯ is the mean DN value of the suspected object. m and n further sorted and accurately identified and located. The framework
are the width and height of the suspected object, which have the of the proposed method is shown in Fig. 3.
same values as the size parameters of the fundamental shape
model. 6 Experiments and discussion
In this section, we conduct experimental studies with images of
UAVs on two real-world landslide scenarios. We show the training

m(x, y) = L(x + 1, y) − L(x − 1, y) 2 + L(x, y + 1) − L(x, y − 1) 2 . (8)

IET Commun., 2020, Vol. 14 Iss. 15, pp. 2434-2441 2437
© The Institution of Engineering and Technology 2020
 17518636, 2020, 15, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-com.2019.1115, Wiley Online Library on [22/08/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Fig. 4 Remote sensing images of a landslide area in a university in Chongqing with training and testing samples
(a) UAV image of a landslide area, (b) Distribution of training samples, (c) Distribution of test samples, (d) Image of WorldView-3 before the landslide

Fig. 5 The data processing of BP neural networks and pre-/post-landslide images

(a) Convergence analysis of the proposed method, (b) Regression of proposed BP NNs, (c) The trend of gradient change for our neural networks, (d) The difference image resulted
from pre-/post-landslide images

Fig. 6 The comparison of landslide detection results by baseline methods and proposed method based on a remote sensing image of UAV
(a) Detection result by Bayes classifier, (b) Detection result by SVM, (c) Detection result by BP NNs, (d) Detection result by our method

and regression of BP neural networks and compare the detection function of SVM algorithms is linear, and the quadratic
results of the proposed method with different baseline methods. programming method is used for searching for classification
hyperplane. The training process and results of our BP neural
6.1 Experimental study on a landslide in chongqing networks are shown in Fig. 5a–c. The result of the difference
image is shown in Fig. 5d, which is generated by Formula (11)
In order to evaluate the performance of our proposed method, we based on the UAV remote sensing image after the landslide and the
conduct some comparative experiments on some conventional satellite remote sensing image before the landslide.
object detection methods. A landslide of about 68m × 55m in the The results of Bayes classifier, SVM algorithm, conventional
west of a college in Chongqing is selected as the research object. neural networks algorithm and our method for landslide detection
The multi-band high-resolution aerial image of the landslide area is are shown in Fig. 6. The accuracy and error of landslide detection
acquired by a UAV of DJI. The overall situation of the landslide and time consumption are shown in Table 1, where CDR is the
area by Airborne imaging is shown in Fig. 4a. In order to test our correct detection rate, MAR is the missing alarm rate and FAR is
proposed method in detecting landslides, we implement the the false alarm rate.
algorithm in MATLAB. The detection results by Bayesian As can be seen from Fig. 6 and Table 1, Bayes classifier has the
classifier, SVM and conventional NN algorithm are compared with accuracy of about 89% in landslide detection, but it also has a
the results of proposed algorithms [34–37]. The distributions of higher false alarm rate than other methods, as many non-landslide
selected training samples and testing samples are shown in Figs. 4b pixel points, such as roads and buildings, are incorrectly identified
and c, where the red dots are for landslide samples, green dots are as landslides. The detection rate of SVM is about 83%. Due to the
for background vegetation samples and blue dots are for road and lack of spatial shape information and change features, the correct
building samples. The basic shape model of this landslide is the detection rate of SVM is not high, as 16.13% of the landslides
dustpan (shown in Fig. 2a). The high-resolution satellite remote pixel points are lost. The accuracy of conventional landslide
sensing data of WorldView-3, which was obtained before the detection based on spectral features and BP neural networks
landslide, is selected as the background image (shown in Fig. 4d). achieves up to 98.61%, which is higher than other methods.
The selected BP neural networks model is described in Section Meanwhile, the false alarm rate of conventional BP neural
2.3 where the maximum number of training is 500, the accuracy of networks is almost 27%, which is unsatisfied. The landslide
the training requirement is 0.01, and the learning rate is 0.01. detection method based on spectral features and neural networks
Gradient descent method is used for training the neural networks. cannot accurately identify the building in the upper right corner of
The scale space factor t in the SIFT algorithm is 0.5, and the scale the image and some parts of the road, as some points are
space rate of two adjacent Gaussian spaces k is 2. The kernel misidentified as part of the landslide. Landslide detection method

2438 IET Commun., 2020, Vol. 14 Iss. 15, pp. 2434-2441

© The Institution of Engineering and Technology 2020
 17518636, 2020, 15, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-com.2019.1115, Wiley Online Library on [22/08/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Table 1 Comparison with baseline methods on a landslide in Chongqing
Models
Metrics Bayes classifier SVM BP neural networks Ours
CDR 89.11% 83.87% 97.54% 98.61%
MAR 10.89% 16.13% 2.46% 1.39%
FAR 29.81% 3.47% 27.38% 6.95%
time consumption (s) 2.42 2.55 3.64 4.25

Fig. 7 Remote sensing images of the Pusa village landslide area with training and testing samples
(a) Airborne image of the Pusa village landslide, (b) Distribution of training samples, (c) Distribution of test samples, (d) Image of Plĺȩiades satellite before the landslide

Fig. 8 The data processing of BP neural networks and pre-/post-landslide images

(a) Convergence analysis of the proposed method, (b) Regression of proposed BP NNs, (c) Trend of gradient change for our neural networks, (d) Difference image resulted from
pre-/post-landslide images

Fig. 9 The comparison of landslide detection results by baseline methods and proposed method based on a remote sensing image of UAV
(a) Detection result by Bayes classifier, (b) Detection result by SVM, (c) Detection result by BP NNs, (d) Detection result by our method

based on BP neural networks and spatial-spectral-change features A high-resolution remote sensing image by Plĺȩiades satellite from
fusion can accurately distinguish landslides from building and Google Earth, which was obtained before the landslide, is selected
road. The accurate detection of the landslide area is further as the background image (shown in Fig. 7d).
achieved by spatial feature fusion. Moreover, the proposed method The parameters of the selected BP neural networks model,
clearly identifies the landslide object in the image without any Bayes classifier and SVM algorithm are the same as Section 6.1.
irrelevant part. The overall detection accuracy of our proposed The training process and results of our BP neural networks are
method is over 98%, outperforming other baseline methods. shown in Figs. 8a–c. The result of the difference image before and
after landslide is shown in Fig. 8d. The results of landslide
6.2 Experimental study on pusa village landslide detection by four algorithms are shown in Fig. 9. The accuracy and
error of landslide detection and time consumption are shown in
Our second experiment is established on the basis of a large Table 2.
landslide occurred on 28 August 2017, in Pusa Village, As can be seen from Fig. 9 and Table 2, Bayes classifier has
Zhangjiawan Town, Nayong County, Guizhou Province. This low accuracy in landslide detection, and nearly half of landslide
landslide caused critical damage with three people killed, seven test samples are not detected in the landslide's remote sensing
people injured and > 30 people lost. In this work, we select this image of this area. The correct detection rate and missing alarm
landslide as our second detection object. We use the remote sensing rate of SVM are better than Bayes classifier, but the overall
image acquired by UAV of the rescue team and the pre-disaster effectiveness is still unsatisfied. The accuracy of conventional
remote sensing image of the landslide area in our method. The landslide detection based on spectral features and BP neural
overall situation of the landslide area is shown in Fig. 7a. The networks achieves up to 89.23%, which is higher than Bayes
distributions of selected training samples and testing samples are classifier and SVM. However, according to the Fig. 9c, some road
shown in Figs. 7b and c, where the red dots are for landslide and cloud samples in the remote sensing image of UAV, which are
samples, and blue dots are for background samples. The basic similar to a landslide on brightness, cannot be accurately identified
shape model of this landslide is the leaf model as shown in Fig. 2b. by the conventional BP neural networks. The accuracy of landslide

IET Commun., 2020, Vol. 14 Iss. 15, pp. 2434-2441 2439

© The Institution of Engineering and Technology 2020
 17518636, 2020, 15, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-com.2019.1115, Wiley Online Library on [22/08/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Table 2 Comparison with baseline methods on Pusa village landslide
Models
Metrics Bayes Classifier SVM BP Neural Networks Ours
CDR 53.54% 62.77% 89.23% 91.23%
MAR 46.46% 37.23% 10.77% 8.77%
FAR 4.32% 1.13% 6.86% 0.47%
Time consumption (s) 2.52 2.77 4.12 5.32

detection is further improved by our method, as the proposed images before and after landslide would bring massive change
model fuses the spatial-spectral-change features based on BP information, if the satellite remote sensing image, as the
neural networks. Combined with Fig. 9d, we can see that a small background image, is acquired long before. In this condition, the
number of other ground objects are successfully distinguished from accuracy of our method may be reduced by the interference of
the landslide. As a result, the proposed algorithm achieves higher ineffective change information. Therefore, the fusion processing
overall recognition rate, by making use of spectral and spatial and information complementarity of multi-source remote sensing
shape features and change features of pre/post landslide in the data based on multiple aerospace platforms would be an important
remote sensing images, and it can maintain a low false alarm rate. method of resolution, which is the next focus of our work.
Overall, the performance of landslide detection of our proposed
method shows the best performance in comparison with other three 6.3.3 The convergence and algorithm complexity for our
baseline algorithms. method: Our BP neural networks utilise the spectral characteristics
for landslide detection. Combined with two real-world landslide
6.3 Discussion scenarios (shown in Figs. 4a and 7a and Tables 1 and 2, we can see
that the colour and shape features of the landslide area are visible
6.3.1 The process of spectral-spatial-change feature fusion and stable. The BP neural networks constructed in this study are
for our method: In this study, a method is proposed to solve the mainly based on MATLAB neural network toolbox, and the
problem of automatic recognition of landslides based on BP neural architecture of BP neural networks and their detailed parameters
networks and fusion of spectral features, spatial shape features, and are based on a large amount of test data and engineering
change features in the remote sensing image. The specific experience. According to the Fig. 5a and c, Figs. 8a and c, Tables 1
procedures are as follows: First, according to the characteristics of and 2, we can see that the mean squared error and the gradient
multi-spectral remote sensing data acquired by UAVs, a landslide change of our neural networks tend to be constant after about 450
identification model based on BP neural network is established. iterations, and then the change or difference is going to be very
The model consists of an input layer, three hidden layers and an small. The training result of the proposed model, presetting
output layer, in which the input layer's dimension is the same as the parameters and detection result all satisfy the requirements of
number of bands in the multi-spectral remote sensing image, and landslide identification. Moreover, in theory, the global optimal
the numbers of neurons in the other four layers are 8,4,8,1. The solution would be found, if the number of samples is large enough
model mainly utilises the spectral features of landslides to detect and the time of iteration tends to infinity. But the time consumption
and identify them in the first step. Combined with Tables 1 and 2 and huge amounts of computation would cause negative effects on
(the result of BP neural networks), it can be seen that the correct practical engineering application, especially in hazard emergency
detection rate is fine, but the false alarm rate and missing alarm response. Therefore, the engineering application value of our
rate are too larger, reducing the overall performance of this method would be poor under the wrong pursuit of absolutely high
algorithms. precision. According to the result of the experiment, BP neural
Second, the feature matching and similarity detection of BP networks meet the basic requirements of landslide identification
neural networks recognition results are carried out, based on the after the training of effective samples, and have low computational
fundamental shape models of landslides and the SIFT algorithm. In complexity (seen in Tables 1 and 2) and high time efficiency. In the
this step, the spatial shape information of landslide is integrated subsequent step of this algorithm, we integrate the shape features
into the target recognition algorithm, which finally realises the and change features, and further improve the recognition accuracy
fusion of spectral features and spatial features. This step is mainly of the landslide, which also confirms the validity of the constructed
to remove the interference of similar objects in the detection results BP neural networks for landslide recognition in this study.
of BPNN, which obtains the similar spectral features with Although our method fuses the shape features and change
landslides. According to the differences on the shape of landslides, features into the BP neural networks, the complexity of our method
these interference objects can be effectively removed in this only increases a little compared with other baseline methods.
section. Because the proposed method in our study processes the UAV
Finally, the change feature extraction of the disaster area is remote sensing data for landslide detection in the way of a linear
carried out based on the satellite remote sensing image before the sequence, and the complexity goes up in the linear proportion to
landslide and the UAV remote sensing image after it, and a each processing step, and it is not geometric growth. Therefore,
probability model based on spectral difference characteristics is combined with two experiments, we recorded the time
established. This step combines change features with the results of consumption, which is directly related to the amount of
BPNN and spatial feature fusion detection, and finally realises the computation for our algorithm, to quantify and evaluate the
fusion detection application of spectral features, change features complexity. From Tables 1 and 2, we can see that time
and shape features. According to the two groups of experimental consumption is a little higher than other baseline methods, and they
data in Tables 1 and 2, we can see that correct detection rate of the are all at the same order of magnitude. Furthermore, according to
landslide identification algorithm in this study is further optimised, the engineering applications in two experiments, the time
and the false alarm rate of the final method in this study is consumption of our method is completely satisfied with disaster
respectively reduced by 20.34 and 6.36%. information extraction in remote sensing emergency response. In
general, algorithm complexity for our method is acceptable and
6.3.2 The possible problem in some special conditions: In meets the requirements of practical engineering applications.
terms of landslide detection and disaster information extraction, the
comparative analysis of images before and after landslide is a
common and effective method. However, the satellite remote
7 Conclusion
sensing data before the landslide in a short time may be scarce, due In this study, we propose a fusion detection model based on BP
to the limitation of satellite return period and bad weather in the neural networks to achieve accurate and efficient feature detection
disaster area. In some special conditions, the great difference of on landslide for hazard information extraction in emergency

2440 IET Commun., 2020, Vol. 14 Iss. 15, pp. 2434-2441

© The Institution of Engineering and Technology 2020
 17518636, 2020, 15, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-com.2019.1115, Wiley Online Library on [22/08/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
response. The spectral features and shape features are extracted [17] Li, Z., Li, Y.S., Guo, J.W.: ‘An automatic landslide interpretation model of
UAV imagery based on bow’, Remote Sens. Inf., 2016, 31, (5), pp. 24–29
from remote sensing images of UAVs. We construct a basic shape [18] Li, Z.B., Shi, W.Z., Myint, S.W.: ‘Semi-automated landslide inventory
model and adopt the SIFT algorithm for feature matching and mapping from bitemporal aerial photographs using change detection and level
transformation. By fusing spatial shape features and spectral set method’, Remote Sens. Environ., 2016, 175, (1), pp. 215–230
features, the landslide object can be detected initially. Based on the [19] Chen, W.T., Li, X.J., Wang, Y.X.: ‘Data fusion technique using wavelet
transform and taguchi methods for automatic landslide detection from
initial result, the change features of difference from pre/post- airborne laser scanning data and QuickBird satellite imagery’, IEEE Trans.
landslide is fused into our method, which further improves its Geosci. Remote Sens., 2016, 54, (3), pp. 1610–1622
accuracy of identification and localisation. The experimental [20] Fey, C., Wichmann, V., Zangerl, C.: ‘Reconstructing the evolution of a deep
results show that the proposed method outperforms the baseline seated rockslide (marzell) and its response to glacial retreat based on historic
and remote sensing data’, Geomorphology, 2017, 298, pp. 72–85
algorithms and achieves more than 91% accuracy in landslide [21] Mezaal, M.R., Pradhan, B., Sameen, M.I., et al.: ‘Optimized neural
detection. architecture for automatic landslide detection from high-resolution airborne
In the future, we will study how to use the Deep Learning laser scanning data’, Appl. Sci., 2017, 7, (7), p. 730
algorithms [38] and other machine learning schemes [39, 40] to [22] Sameen, M.I., Pradhan, B.: ‘Landslide detection using residual networks and
the fusion of spectral and topographic information’, IEEE Access, 2019, 7, pp.
further improve the detection accuracy. Moreover, we will consider 114363–114373
more practical scenarios where the UAV sensing data is noisy, [23] Liu, X., Zhu, X., Zhang, Q., et al.: ‘A remote sensing and artificial neural
exploiting the de-noise methods [41, 42]. network-based integrated agricultural drought index: index development and
applications’, Catena, 2020, 186, p. 104394
[24] Xie, J., Zhou, J.: ‘Classification of urban building type from high spatial
8 Acknowledgments resolution remote sensing imagery using extended mrs and soft bp network’,
IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens., 2017, 10, (8), pp. 3515–
This work was partially funded by the National Natural Science 3528
Foundation of China (Grant Nos. 61872447 and 61672038), the [25] Yan, X., Shi, W., Zhao, W., et al.: ‘Mapping dustfall distribution in urban
Science and Technology Innovation Project of Chongqing Social areas using remote sensing and ground spectral data’, Sci. Total Environ.,
2015, 506, pp. 604–612
Work and People's Livelihood Guarantee (Grant Nos. [26] Li, T., Zheng, X., Xiong, T.: ‘Research on the high resolution remote sensing
CSTC2017SHMSA00002), the Science and Technology Research image classification algorithm based on improved neural network model’.
Program of Chongqing Municipal Education Commission (Grant 2016 Int. Conf. on Communication and Electronics Systems (ICCES),
No. KJQN201912905), and the Natural Science Foundation of Coimbatore, India, 2016, pp. 1–6
[27] Li, L., Chen, Y., Xu, T., et al.: ‘Super-resolution mapping of wetland
Chongqing (Grant No. CSTC2018JCYJA1879). The authors would inundation from remote sensing imagery based on integration of back-
like to thank Google Earth and Sina News for providing the images propagation neural network and genetic algorithm’, Remote Sens. Environ.,
of the Pusa village landslide before and after its occurrence. 2015, 164, pp. 142–154
[28] Prasad, R., Pandey, A., Singh, K., et al.: ‘Retrieval of spinach crop parameters
by microwave remote sensing with back propagation artificial neural
9 References networks: a comparison of different transfer functions’, Adv. Space Res.,
2012, 50, (3), pp. 363–370
[1] Qu, Y.B., Tang, S.J., Dong, C., et al.: ‘Posted pricing for chance constrained [29] Chen, M.S., Ni, L., Jiang, X., et al.: ‘Retrieving atmospheric and land surface
robust crowdsensing’, IEEE Trans. Mob. Comput., 2018, 19, (1), pp. 188–199 parameters from at-sensor thermal infrared hyperspectral data with artificial
[2] Xiang, C.C., Yang, P.L., Tian, C., et al.: ‘Carm: crowd-sensing accurate neural network’, IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens., 2019, 12,
outdoor rss maps with error-prone smartphone measurements’, IEEE Trans. (7), pp. 2409–2416
Mob. Comput., 2016, 15, (11), pp. 2669–2681 [30] Liu, D.B., Cao, Z.C., He, Y., et al.: ‘Exploiting concurrency for opportunistic
[3] Guo, S.M., Chen, C., Wang, J.Y., et al.: ‘Rod-revenue: seeking strategies forwarding in duty-cycled iot networks’, ACM Trans. Sen. Netw., 2019, 15,
analysis and revenue prediction in ride-on-demand service using multi-source (3), pp. 1–33
urban data’, IEEE Trans. Mob. Comput., 2019, (Early Access) [31] Li, T., Zhang, Y.X., Lv, Z.Y.: ‘Landslide inventory mapping from bitemporal
[4] Chen, F., Yu, B., Li, B.: ‘A practical trial of landslide detection from single- images using deep convolutional neural networks’, IEEE Geosci. Remote
temporal landsat 8 images using contour-based proposals and random forest: a Sens. Lett., 2019, 16, (6), pp. 1–5
case study of national nepal’, Landslides, 2018, 15, (3), pp. 453–464 [32] Ma, W.P., Wen, Z.L., Wu, Y.: ‘Remote sensing image registration with
[5] Liu, Z.H.: ‘Research on geo-hazards monitoring method based on UAV and modified SIFT and enhanced feature matching’, IEEE Geosci. Remote Sens.
GIS’, Geospatial Inf., 2012, 1, (4), p. 9 Lett., 2017, 14, (1), pp. 3–7
[6] Huang, H.Z., Chen, J.P., Zheng, Y.W.: ‘Interpretation of mine geological [33] Chen, Q., Zhao, L., Lu, J., et al.: ‘Modified two-dimensional Otsu image
hazards based on UAV remote sensing technology’, J. Geol., 2017, 41, (3), segmentation algorithm and fast realisation’, IET Image Process., 2012, 6,
pp. 499–503 (4), pp. 426–433
[7] Huang, H.F., Lin, H.Y., Lv, Y.M.: ‘Micro unmanned aerial vehicle based [34] Majid, M., James, V.A., Rodrigo, A.A.: ‘A machine learning framework for
remote sensing method and application for emergency survey of individual detecting landslides on earthen levees using spaceborne SAR imagery’, IEEE
geohazard’, J. Eng. Geol., 2017, 25, (2), pp. 447–454 J. Sel. Top. Appl. Earth Obs. Remote Sens., 2015, 8, (8), pp. 3791–3801
[8] Wang, Y.J., Wang, J., Huang, Y.: ‘Several key technologies of multi-source [35] Xu, H., Li, X., Gong, W: ‘Research on recognition of landslides with remote
monitoring and analysis of surface disasters in mining areas and their sensing images based on extreme learning machine’. 2017 IEEE Int. Conf. on
applications’, Bull. Surv. Mapp., 2014, 1, (S1), pp. 4–9 Computational Science and Engineering (CSE) and IEEE Int. Conf. on
[9] Chen, T.B., Hu, Z.W., Wei, L.: ‘Data processing and landslide information Embedded and Ubiquitous Computing (EUC), Guangzhou, China, 2017, vol.
extraction based on UAV remote sensing’, J. Geo-Inf. Sci., 2017, 19, (5), pp. 1, pp. 740–747
692–701 [36] Niu, R.Q., Wu, X.L., Yao, D.K., et al.: ‘Susceptibility assessment of
[10] Tang, Y., Wang, L.J., Ma, G.C., et al.: ‘Application of UAV and domestic high landslides triggered by the lushan earthquake, april 20, 2013, china’, IEEE J.
scoring technique in ‘6.24’ maoxian landslide emergency monitoring’, Land Sel. Top. Appl. Earth Obs. Remote Sens., 2014, 7, (9), pp. 3979–3992
Resour. Inf., 2017, 15, (8), pp. 15–19 [37] Kalantar, B., Ueda, N., Lay, U.S., et al.: ‘Conditioning factors determination
[11] Fu, X., Guo, J.W., Liu, X.J., et al.: ‘Method of earthquake landslide for landslide susceptibility mapping using support vector machine learning’.
information extraction based on high resolution unmanned aerial vehicle 2019 IEEE Int. Geoscience and Remote Sensing Symp. (IGARSS),
images’, J. Seismol. Res., 2018, 41, (2), pp. 186–191 Yokohama, Japan, 2019, pp. 9626–9629
[12] Koschitzki, R., Schwalbe, E., Krohnert, M., et al.: ‘Multi-temporal [38] Fan, X.C., Xiang, C.C., Chen, C., et al.: ‘Buildsensys: reusing building
photogrammetric analysis to monitoring the river las minas, punta arenas, sensing data for traffic prediction with cross-domain learning’, IEEE Trans.
chile’, IEEE Latin Am. Trans., 2018, 16, (9), pp. 2481–2489 Mob. Comput., 2020, (Early Access)
[13] Hsu, C.H., Tsao, T.C., Huang, C.M., et al.: ‘Using remote sensing techniques [39] Xiang, C.C., Yang, P.L., Tian, C., et al.: ‘Calibrate without calibrating: an
to identify the landslide hazard prone sections along the south link railway in iterative approach in participatory sensing network’, IEEE Trans. Parallel
taiwan’, Procedia Eng., 2016, 143, pp. 708–716 Distrib. Syst., 2015, 26, (2), pp. 351–361
[14] Giordan, D., Manconi, A., Tannant, D.D., et al.: ‘Uav: low-cost remote [40] Zheng, J.C., Cai, Y.M., Wu, Y., et al.: ‘Dynamic computation offloading for
sensing for high-resolution investigation of landslides’. 2015 IEEE Int. mobile cloud computing: a stochastic game-theoretic approach’, IEEE Trans.
Geoscience and Remote Sensing Symp. (IGARSS), Milan, Italy, 2015, pp. Mob. Comput., 2018, 18, (4), pp. 771–786
5344–5347 [41] Xiao, F., Chen, L., Sha, C.H., et al.: ‘Noise tolerant localization for sensor
[15] Nicola, C., William, F., Stefano, M.: ‘Spaceborne, UAV and ground-based networks’, IEEE/ACM Trans. Netw., 2018, 26, (4), pp. 1701–1714
remote sensing techniques for landslide mapping, monitoring and early [42] Gong, L.Y., Zhao, Y.Y., Xiang, C.C., et al.: ‘Robust light-weight magnetic-
warning’, Geoenvironmental Disasters, 2017, 4, (9), pp. 1–23 based door event detection with smartphones’, IEEE Trans. Mob. Comput.,
[16] Saito, H., Uchiyama, S., Hayakawa, Y.S.: ‘Landslides triggered by an 2018, 18, (11), pp. 2631–2646
earthquake and heavy rainfalls at aso volcano, japan, detected by UAS and
sfM-MVS photogrammetry’, Prog. Earth Planet. Sci., 2018, 5, (1), pp. 15–27

IET Commun., 2020, Vol. 14 Iss. 15, pp. 2434-2441 2441

© The Institution of Engineering and Technology 2020