The use of remote sensing satellite using deep learning in emergency monitoring of high-level landslides disaster in Jinsha River

CNN is a kind of feedforward neural network including convolution computation. CNN imitates the visual perception mechanism of biology, and can be used for supervised learning and unsupervised learning. The convolutional kernel parameter sharing in the hidden layer and the sparse connection between layers can reduce the computation. CNN can also use the two-dimensional structure of input data to reduce a large number of parameters into a small number of parameters in image processing, and does not affect the original characteristics of the image, so it is widely used in the field of machine vision and natural language [7]. Lu et al. (2020) used the remote sensing technology of low altitude unmanned aerial vehicle (UAV) to study UAV images at the place where landslide disasters occurred, and used deep learning algorithm to generate landslide sample database from the obtained images. By comparing various landslide feature models of deep learning, the transfer learning algorithm with the best effect was finally selected for landslide image analysis [8]. Ji et al. (2019) applied deep learning to the detection of building damage after an earthquake, analyzed the high-resolution remote sensing images before and after disasters through the algorithm, and compared the performance of gray matrix texture and CNN features by using random forest classifier. The results show that CNN is better in texture feature recognition of high-resolution remote sensing image [9]. Hua et al. (2020) studied the relationship between landslide sensitivity and time in Badong–Zigui section in the Three Gorges area based on multi-source landslide detection data (geological, topographic, hydrological and remote sensing data), and nine landslide influencing factors were generated according to the monitoring data. Then, the slope element extraction algorithm and neural network algorithm were used to study the dynamic change of landslide sensitivity. The results show that the main factors affecting the landslide are river distance, normalized difference index, relative relief and engineering rock group. Therefore, the fluctuation of reservoir water level and road construction will cause landslide phenomenon [10]. Seydi et al. (2020) proposed a change detection framework based on CNN for the study of remote sensing datasets. Three parallel channels were used to extract the depth features and deep features of the image, and three different types of remote sensing datasets were used for testing. Experimental results show that the overall accuracy and Kappa coefficient of the method are 95.89% and 0.805, respectively [11]. The existing research results show that CNN has a good feature extraction accuracy for remote sensing data, which can help to study landslide phenomenon in remote sensing images.

DBN algorithm is a kind of machine learning, which can be used for both unsupervised learning and supervised learning. DBN is a generative model. Compared with the traditional neural network of discriminative model, the generative model is to establish a joint distribution between observation data and tags. By training the weights among the neurons, the whole neural network can generate training data according to the maximum probability. DBN can be used not only to identify features and classify data, but also to generate data. Therefore, it is widely used in the fields of handwritten character recognition, speech recognition and image processing [12]. Ye et al. (2019) proposed a deep learning framework with constraints to detect landslides in hyperspectral images. First, the spectrum spatial features of landslides were extracted by DBN, and then verified by logistic regression classifier. The results show that compared with the traditional hyperspectral image classification method, the framework has higher accuracy, and the accuracy reached 97.91% [13]. Liu et al. (2020) proposed an automatic landslide recognition model based on improved segmentation network, and realized intelligent extraction of landslide disaster after earthquake by automatically extracting hierarchical features. In the actual model test, the recognition accuracy was 91.3%. The results show that the method has high feasibility for landslide disaster identification after earthquake [14]. Wang et al. (2020) used DBN to extract soil moisture data from medium resolution spectral images of Fengyun-3D (FY-3D) satellite, and constructed a soil moisture inversion model SM-DBN. The SM-DBN model included two sub-networks of temperature and soil moisture. The ground temperature data, normalized difference vegetation index, enhanced vegetation number and soil moisture data were used as the input values of the network. The medium resolution spectral images from 2018 to 2019 were selected as the dataset. The experimental results show that the root mean square error of the overall accuracy of SM-DBN model is 0.032, which is better than the test results of linear regression and back-propagation neural network [15]. Samadi et al. (2019) proposed a change detection algorithm for synthetic aperture radar images based on DBN. The DBN was trained by unsupervised feature learning and supervised network fine-tuning. Morphological operators were used to obtain images to provide appropriate and diverse datasets for DBN. The experimental results show that this method can reduce the computational load without damaging the performance of DBN, and has ideal performance and high accuracy [16]. DBN can effectively identify image features and classify data, so it can be used to identify landslides in remote sensing images.

At present, the detection methods of landslides are to mainly use DBN and CNN to identify and extract the collected images. Therefore, two commonly used landslide detection models, DBN and CDN, are used for comparative analysis to identify and early warn the disaster through the remote sensing images in the area of Jinsha River.

High-level landslides often occur in places with few people and inconvenient transportation. The ecological environment of these places is fragile, and it is very easy to induce large-scale landslide disasters when there is continuous precipitation or earthquake. Therefore, affected by the objective conditions, it is very unfavorable to the emergency monitoring, the analysis and investigation of potential safety hazards and the rescue after the high-level landslide disaster in Jinsha River. The actual factors affecting the landslide in Jinsha River are as follows. (1) Slope shape: It is the prerequisite of slope type rock mass sliding. When there is external disturbance, landslide is easy to occur. (2) Rock mass: Hydrophilic soft soil will slide under the action of gravity under the condition of large amount of precipitation. (3) Terrain structure: The weak surface is parallel to the slope surface, which is prone to creep. (4) Water: Water will change the internal stress of soil and make the soil easy to flow. At the same time, the water will erode the lithology of the slope toe and cause the slope structure damage. (5) Earthquake: Earthquake will lead to the movement of unstable rock mass, resulting in liquefaction of sand and soil, resulting in landslide [17‐20].

In order to study the landslide problem with remote sensing analysis technology, the characteristics of landslides on remote sensing images are analyzed, and the internal and external inducing factors, such as lithology, landform, precipitation, geological structure and hydrological conditions, are transformed into landslide influence factors. The influence factors of landslide are different in different regions. Soil erodibility, slope gradient, vegetation coverage, fault location, river influence, rainfall erosion and earthquake intensity are selected as landslide influence factors. After obtaining the landslide factors, it is necessary to quantify the impact of each factor on the landslide, so as to provide data basis for the application of deep learning algorithm.

(1) Soil erodibility. It describes the degree of sudden erosion by water balls. The larger the value of soil erodibility is, the more likely soil erosion will occur under the same flow condition. The equation is as follows:\(K_{E}\) represents the soil erodibility corrected according to the parameters, and \(K_{{0}}\) is the uncorrected soil erodibility factor.

(2) Fault zone factor. When the geological environment is on the fault zone, landslide is easy to occur. It has been proved that landslides often occur in fault zones with strong geological structures. The influence factors of fault zones can be obtained by superposition of Kriging interpolation.

(3) Earthquake influence factors. The destruction of rock mass caused by earthquake leads to the decrease in soil shear strength, the structural stability of slope and the occurrence of landslide. Through the density analysis of seismic points, the distribution map of earthquake influence factors is obtained by Kriging interpolation.

(4) River influence factor. When there is a river near the slope, the river erosion will destroy the stability of the slope and reduce the anti-sliding ability of the slope. The river influence factor can be obtained by multiplying the distance from the slope to the river and the flow of the river.

(5) Rainfall factors. When there is heavy rainfall on the slope, landslide is easy to occur.\(R\) is the rainfall erosivity, \(i\) is the month, \(P_{i}\) is the current month's precipitation, and \(P\) is the current year's precipitation. Therefore, when the local rainfall intensity is high, the rainfall erosivity of the slope will also increase, which is easy to induce landslide disaster.

(6) Vegetation influence factors. Plants will fix the local soil to reduce large-scale landslides. Through satellite remote sensing images, the local vegetation coverage can be analyzed. The equation of vegetation coverage is as follows:\(NDVI\) is the normalized difference vegetation index (NDVI), which quantifies vegetation by measuring the difference between near infrared (vegetation strong reflection) and red light (vegetation absorption).

(7) Slope influence factor. The slope will affect the erosion and sliding of slope under the influence of gravity. Through the discrete element method, slope calculation is carried out [21‐24].

CNN and DBN can be used to effectively identify the images in remote sensing images. Therefore, in this exploration, CNN and DBN are used to identify the impact of landslides in remote sensing images, and the above landslide influencing factors are taken as constraints. Therefore, first, DBN and CNN are introduced.

(1) The restricted Boltzmann machine is a model based on energy function including two layers of visual layer and hidden layer. All layers are connected, and there is no connection within the layer [25]. When the states of hidden layer and visual layer are determined, the energy function equation and joint probability distribution of nodes in the two layers are as follows:

In the equation, \(\theta\) represents the parameters of the model, \(a\) and \(b\) represent the offset between the visual layer and the hidden layer. \(w\) represents the weight between the \(i\)-th neuron in the visual layer and the \(j\)-th neuron in the hidden layer. \(v\) and \(h\) represents the state values of neurons in visual layer and hidden layer, respectively. The partition function \(\sum\limits_{v} {\sum\limits_{k} E {(}v,h|\theta {)}}\) represents the summation of all possible states of the nodes in the visible layer and the hidden layer.

DBN is composed of multi-layer restricted Boltzmann machines. The output data of the upper layer is taken as the input layer of the next layer and so on. A classifier is added at the top layer to make it have the ability of classification and recognition. Figure 1 shows the structure of DBN. DBN combines the powerful data extraction and retention capabilities of restricted Boltzmann machine, and has the classification accuracy of data [26‐29].

(2) CNN is composed of one or more convolutional layers, top full connected layer, association weight and pooling layer. Figure 2 shows the structure. The convolutional kernel moves regularly on the image, grabs the features with the same structure in the image to make a feature map, which is the basis of image classification [30, 31].

The principle of CNN parameter optimization is as follows. If each convolution layer contains only one convolution kernel, after convolution calculation, the convolution characteristic matrix \(C_{i}\) is output. The equation is as follows.

In the equation, \(W_{i}\) is the weight of convolution kernel, \(\otimes\) is convolution calculation, \(b_{i}\) is node bias, and \(f(^{{}} )\) is activation function. After convolution operation, \(C_{i}\) enters the sampling layer for dimensionality reduction. The characteristic obtained by sampling is \(S_{i}\):

After feature extraction of the image, the feature vector is input into the classifier for classification. The probability distribution of completing the input image for each category is (\(n\) represents the category index):

According to the output value and the actual situation, the loss function is constructed, and the L2-regularization is added to prevent over fitting. The equation is (\(\lambda\) is the weight attenuation coefficient):

To improve the recognition effect of DBN and CNN on remote sensing images, two deep learning algorithms are applied to establish landslide monitoring model based on spectral characteristics and spatial information of remote sensing satellite images. The influence factors of landslides are regarded as constraints, and the detection accuracy of landslides is used as the evaluation criteria. The performance of the two landslide monitoring models is analyzed. The two models adopt different processing flows, and Fig. 3 shows the workflow.

(1) Landslide monitoring model based on DBN. In order to improve the detection of remote sensing image, landslide influence factors are taken as restraint conditions and combined with DBN to construct landslide detection model. The dimension of remote sensing data is reduced by using the algorithm, and the principal component information is selected as the input. The constrained DBN model is shown in Fig. 4.

First, the preprocessed remote sensing image is input into DBN. The equation of the first layer restricted Boltzmann machine is as follows:

The domain spectral vector is used as input (\(n\) is the domain size and \(z\) is the principal component number).

The feature vector input to the classifier is \(L = \left[ {F,O,C} \right]\). Back-propagation algorithm is used to adjust the network parameters, and \(J{(}W^{l,} b^{l}_{y} ,b^{l}_{z} {)}\) is the error function [32‐35].

(2) Landslide monitoring model based on CNN. CNN has good performance in extracting large-scale image information. Therefore, under the constraint conditions, a dual depth detection model convolutional neural-deep belief network (CDN) of CNN and DBN is established. After being processed by DBN and CNN, the input data is combined into a vector and input into the classifier [36]. The process of model checking is shown in Fig. 5.

In the CDN model, DBN is the same as the previous processing method. The output characteristic matrix of CNN is as follows:

Then, it needs to sample and output to the classifier. The equations are:

The equation for establishing CNN is as follows:

The experimental data are high-resolution remote sensing images within 60 km of the landslide area of Jinsha River. Among them, 300 landslides images obtained by remote sensing satellites are used as the training set of the model, and 100 hybrid satellite images (20 non-landslide images) are used as the test set for target detection. Tables 1 and 2 give the parameter settings of DBN and CNN, respectively. Table 3 shows the hardware and software environments in this exploration.

The restricted Boltzmann machine can extract feature information after training, but the number of neurons will affect the extracted data, resulting in data errors. Therefore, it is necessary to make quantitative judgment on the obtained data. The number of hidden neurons in the restricted Boltzmann machine is set to 10, 50, 100, 150 and 200, and the comparative experiment is carried out. The results are shown in Fig. 6.

As can be seen from Fig. 6, with the increase in the number of neurons, the overall error first decreases and then increases. When the number of neurons is 100, the overall error is the smallest. Therefore, the number of neurons in DBN is selected as 100. The depth of DBN has an impact on the extraction of remote sensing image data, and a part of data will be lost in every layer of network. Therefore, it is necessary to verify the error of different network layers, increase the network layer by layer, and record the classification error, as shown in Fig. 7.

As can be seen from Fig. 7, DBN shows high accuracy in classification of remote sensing images. Moreover, with the increase in the number of layers, the classification error first decreases and then increases under different learning rates. However, when the number of layers is 3, the classification error can always reach the minimum. This also shows that in deep learning, it is not the more layers, the better the learning effect, so it is necessary to choose the appropriate number of layers in different situations.

The extraction effect of the two detection models, DBN and CDN, is compared with that of the traditional spectral angle mapper (SAM), spectral information divergence (SID) method, and support vector machine (SVM) algorithm. The detection results of the algorithm are described by using the overall accuracy and landslide detection accuracy. Figure 8 shows the results.

Figure 8 shows that the overall detection accuracy and landslide detection accuracy of CDN algorithm and DBN algorithm are higher than those of other algorithms. The overall accuracy of CDN is 96.02%, which is higher than 94.67% of DBN, indicating that CDN has more advantages in detecting landslide characteristics. However, when the landslide influence factors are added as constraints, the accuracy of landslide detection based on CDN and DBN are 97.63% and 97.56%, respectively. The detection accuracy of the two methods is similar, and both of them are higher than the accuracy without constraints. Fragmentation can lead to wrong information. The fragmentation degree of DBN is 9.02 and that of CDN is 8.89, which shows that DBN and CDN have good effect on image feature extraction, and the loss of remote sensing image information is less.

To sum up, the model parameters and detection results of the two landslide detection models based on DBN and CNN are analyzed when processing remote sensing images. The results show that the accuracy of the two algorithms in landslide detection of remote sensing images is equivalent, both of them can reach more than 97%, and the detection effect of CDN is more accurate. Compared with the landslide detection effect of dynamic network model [37] (error rate is 0.01), the recognition effect is poor, but the model designed in this exploration can identify large-scale landslide phenomenon, while the dynamic network model only detects 15 m landslide phenomenon, so the model designed in this exploration has a better application range.