Study Area

The study area is located in the border region of the Yunnan, Guizhou, and Sichuan Provinces, in addition to a minimal number of rivers flowing through Chongqing Municipality (Fig. 2). The Chishui River Basin originates in Fangba Township, southwestern Zhenxiong County, Yunnan Province. Its main tributaries are the Erdao, Tongzi, and Xishui Rivers, with the rest spread over many counties and cities, including Zhenxiong and Weixin in Yunnan Province; Bijie, Dafang, Jinsha, Zunyi, Rhenhuai, Chishui, Xishui, and Tongzi in Guizhou Province; and Xuyong, Gulin, and Hejiang in Sichuan Province. The Chishui River Basin is located at 104°45′–106°51′E, 27°20′–28°50′N. The basin has a total area of 20,440 km². The terrain of the Chishui River Basin is higher in the northeast and is on the slope from the Yunnan–Guizhou Plateau to the Sichuan Basin. In the Chishui River Basin, the Paleozoic stratum dominates the section above Taipingdu and Sandy Beach, with widely distributed limestone interspersed with shale and coal; meanwhile, the Mesozoic stratum, with sandstone, mudstone, conglomerate, and coal, dominates the section below Sandy Beach. The basin experiences southeastern Gulin area mountain tectonics along with regional tectonics. The section of the river below Unk Indu crosses the shield of the mountain tectonic plate, and the inconspicuous spine part is composed mainly of a series of broad and gentle east– west comb folds with fewer ruptures, where regional tectonic plates are single and stable. The Chishui River Basin is adjacent to a plateau and valley, characterized by a continental climate. Most of the basin is dry and cold in winter, and wet and hot in summer, with an average temperature of 15–20ºC. The atmosphere in the middle and lower reaches of the basin shows a pattern of hot summers and mild winters. The rainy season is from June to September, and the annual rainfall is 900-1000 mm in the upper part of the Basin and 1000-1500 mm in the middle and lower parts.

Data

Selection and Analysis of Evaluation Factors

Analysis of Conditioning Factors

While there is no uniform standard for the number of factors to include, the inclusion of too many factors can lead to an increase in the amount of data and the running time for the evaluation becomes longer; the selection of fewer factors may be inadequate to reveal potential correlations with landslide-triggering factors. Therefore, it is crucial to conduct a preliminary investigation and analysis of the mechanisms that trigger landslides in the study area. The correct selection of applicable evaluation factors can significantly improve the accuracy of vulnerability evaluations for effective disaster prevention and mitigation in disaster-affected areas.

Fig. 3.

Landslide impact factor weighting values.

This study used PCA to screen for landslide conditioning factors in the Chishui River Basin. The attribute data of landslide evaluation factors were extracted based on field investigations and a review of the literature, and a landslide attribute database was established. The attribute database of historical landslides was used as the primary dataset for evaluating landslide conditioning factors. The PCA conducted in SPSS software was used to evaluate the weights and correlations of the factors. The weights of the factors were used to screen appropriate landslide conditioning factors so that the results of landslide susceptibility zoning were as accurate as possible.

The relationships between the weights of the 21 evaluation factors are shown in Fig. 3. The weights of the slope direction, geomorphological type, lithology, soil, TPI, TWI, land use type, and population were 0.042, 0.040, 0.038, 0.044, 0.043, 0.042, 0.037, and 0.035, respectively, with a weight value below 0.045. Therefore, the remaining 13 landslide evaluation factors (DEM; slope; NDVI; temperature; precipitation; distance from roads, rivers, and faults; SPI; surface cut; surface roughness; terrain relief; and elevation coefficient of variation) were applied to landslide susceptibility evaluation and susceptibility mapping in the Chishui River Basin.

Methods for Selecting Non-Landslide Points

To avoid errors caused by anthropogenic factors when selecting non-landslide points, we used the information volume model to take the historical landslide points as the dataset and used the 13 landslide conditioning factors identified by PCA to calculate the information quantity of each factor (Table 1).

Utilizing the natural breakpoint method through the raster superposition value of the 13 information quantities to divide the landslide susceptibility zones into five grades: very low, low, medium, high, and very high susceptibility (Fig 4). In this study, we used an information volume model to predict landslide susceptibility zones and select 724 non-landslide points in very-low-susceptibility zones.

Neural Network Models

(1) BP model

The BP model is a multi-layer feedforward neural network model similar to human neurons for signaling; it consists of an input layer–implicit layer–output layer structure, which has a remarkable ability to process nonlinear data [29]. The construction of the BP model first requires determining the number of neuron nodes in the input layer (n), the number of neuron nodes in the implicit layer (w_ij), and the number of neuron nodes in the output layer (m). Next, the hidden layer threshold (a_j), output layer threshold (b_k), learning rate, number of training iterations, minimum error of training objectives, and excitation function were determined, where the input layer was connected to the hidden layer with weight w_i, and the hidden layer was connected to the output layer with weight w_jk.

In this study, a BP neural network with the structure of “13-6-2” was constructed in MATLAB. The number of neuron nodes in the input layer was 13, the number of neuron nodes in the hidden layer was 6, and the number of neuron nodes in the output layer was 2. The excitation function of the hidden layer was tansig, the function of the output layer was pure-line, and the training function was trainlm, with randomly determined w_ij, w_jk, a_j, b_k. The number of training times was 1000, the learning rate was 0.01, and the minimum error of the training target was 0.000001.

(2) LSTM model

The LSTM model is a deep learning model optimized for RNN models that effectively overcomes the vanishing gradient problem and achieves long-and short-term information memory [30]. With these advantages, scholars in various countries have begun to gradually apply LSTM models to regional landslide susceptibility predictions, allowing information to flow continuously through the point nodes of the model [31]. The structure of LSTM is shown in Fig 5.

The input gate (i_t) determines whether the new input landslide information features are allowed to be updated and saved to memory, the forget gate (f_t) controls whether the memory remembers or forgets the previous landslide information features, the output gate (O_t) determines whether the landslide feature information is allowed to be output, the cellular state (C_t) stores the long-term landslide displacement information, memristors (h_t) is responsible for storing the short-term landslide displacement information, and the candidate state (C˜ _t ) represents the new displacements of landslides summarized by the recurrent neural network [32].

Fig. 4.

Landslide susceptibility zoning map based on informativeness modeling.

In this research work an LSTM model was constructed using MATLAB software, its specific network structure contains an input layer with 13 dimensions, six hidden units, a ReLu layer, a fully connected layer, a Softmax function layer, and finally a binary classification layer. To optimize the prediction accuracy of this model, the MaxEpochs were set to 1000, and the MiniBatchSize parameter was also added with a value of 100, as well as the Adam gradient descent algorithm, which resulted in a learning rate of 0.0001 after 700 training sessions.

(3) CNN model

The CNN model is a neural network model initially proposed by LeCun for digital recognition. It differs from other models mainly in its convolutional operation, and has been widely used for image extraction and classification processing [33, 34]. CNN mainly comprises feed-forward neural networks consisting of one or more convolutional, pooling, and fully connected layers [35, 36]. Their local perceptions and weight sharing make it easy to optimize and reduce model complexity [34, 37-41] The convolutional layer extracts the features of the data, and the convolutional kernel performs convolutional computations to obtain the feature layer of the input image. Assuming that the output layer is an M × 2D image, the convolution formula is:

In Equation (1), y_mn is the output feature map of the convolutional layer, f is the activation function, w is the convolutional kernel of j × i, and b is the bias [34]. Assuming that there are n landslide conditioning factors in the input data and that the convolutional layer filters m convolutional kernels, the layer has n-m+1 feature vectors. Regularization prevents overfitting, which is followed by classification learning using a fully connected layer. The last layer outputs the probability of landslide category a from the normalized exponential function, also known as the softmax function, assuming that there is an array V. V_i denotes the i^th element in V_i, then the softmax value of V_i is

In Equation (2), ε is a natural constant and S_i denotes the ratio of index I to the sum of the indices of V_i.

In this study, a CNN was built using MATLAB software, and the specific network model structure was analyzed, which contains a 13 × 1 × 1 input layer, two convolutional layers, two maximal pooling layers, two ReLu layers, a Softmax’s function layer, and finally a binary classification layer. In addition, based on the results of multiple training and prediction, MaxEpochs was set to 500. Moreover, to make model prediction more accurate, the parameter MiniBatchSize was added, with a value of 128. To avoid overfitting, the regularization parameter 1e-04 was used, and the learning rate was 0.0005.

Fig. 5.

LSTM model structure.

Landslide Susceptibility Zoning Results

The ArcGIS application randomly generated 724 non-landslide points with a proportion of 1:1 to the number of landslide points in the very-low-susceptibility zone under the information quantity model and constructed a point dataset with 724 historical landslide points. The study area was separated into 20604472 rasters, and ArcGIS was utilized for the raster-to-point operation using a tool for multi-value extraction to points in ArcGIS and extracting the values of the 13 landslide conditioning factors for all raster points in the study area. We randomly selected 70% of the dataset as a training set and the remainder as a test set, and imported the divided training set and test set into MATLAB software to complete the training and testing of BP, LSTM, and CNN. The three trained neural network models predicted the probability of a landslide occurring at all grid points in the study area.

After training the neural network models (i.e., BP, LSTM, and CNN), each raster point resulted in a corresponding landslide susceptibility probability value, and the same natural breakpoint method was used to classify the landslide susceptibility probability value into five categories of susceptibility zones (Fig. 6). According to the susceptibility results of BP, LSTM, and CNN models, the historical landslide point map was “spatially connected” in ArcGIS, and the partitioned area of each susceptibility partition and the distribution of historical landslides were obtained respectively (Table 2). The distribution of historical landslide points in landslide susceptibility zones predicted by different models shows that there are 520 historical landslide points in the very high susceptibility zone predicted by the BP model, and there are 20 historical landslide points in the very low susceptibility zone; there are 599 historical landslide points in the very high susceptibility zone predicted by the LSTM model, and there are 25 historical landslide points in the very low susceptibility zone; the number of historical landslide points distributed in the very high susceptibility zone predicted by the CNN model is the number of historical landslides distributed in the very high susceptibility zone predicted by the CNN model is 604, accounting for 83.43% of the total number of historical landslides, and the number of historical landslides distributed in the very low susceptibility zone predicted by the CNN model is 16, which is only 2.21% of the number of historical landslides. Therefore, the prediction result of the CNN model is closest to the actual situation of the number of landslides in the Chishui River Basin and is effective.

Model Accuracy Evaluation

Model accuracy assessment is crucial for landslide susceptibility evaluation. A confusion matrix is commonly used for the prediction accuracy analysis of classification problems. The accuracy, precision, recall, and F1 scores obtained from this matrix are important quantitative metrics for assessing classification results [42-45]. As shown in Table 5, the evaluation indices of the three neural network models were all greater than 0.9, indicating that the evaluation results of these three models had a certain degree of credibility.

Fig. 6.

Landslide susceptibility zoning map under a) BP; b) LSTM; c) CNN models. Fig. 6. Landslide susceptibility zoning map under c) CNN models.

In landslide susceptibility evaluation, previous authors have often used a receiver operating characteristic (ROC) curve in combination with the first four test indicators to validate the susceptibility zoning results and then evaluated the accuracy of the zoning results by calculating the area under the curve (AUC) value [43-45].

Based on the prediction results obtained from training the three neural network models, SPSS was used to generate the ROC curves. As shown in Fig. 7, the AUC values of BP, LSTM, and CNN were all greater than 0.9. In particular, the AUC value of 0.997 for CNN was significantly greater than that of BP (AUC=0.981) and LSTM (AUC =0.984), indicating that this model had the best fitting and prediction ability for this study area.

Validity of the Information Model for Selecting Non-Landslide Points

In landslide susceptibility assessments, the importance of the correct selection of non-landslide points cannot be underestimated. In addition to the informative model used in this study to select non-landslide points, most recent studies have relied on randomly selected non-landslide sites in areas where no landslides have occurred [46, 47]. In general, the possibility of landslides occurring in low-slope areas, such as rivers and gullies, is slight [19, 48], so low-slope areas derived from the DEM data can be utilized to randomly select non-landslide sites within the area. A buffer zone was established at a certain distance from the historical landslide point, and non-landslide points were selected outside the buffer zone [49-52]. To verify the reasonableness of selecting non-landslide points with the information volume model in this study, four non-landslide point-selection methods were used: random selection of non-landslide areas [44], random selection of low-slope areas with slopes less than 3°, random selection of landslide sites outside the buffer of 1000 m [51], and selection of very-low-landslide-susceptibility zones with the informative model [52]. The selected non-landslide points were input into the CNN model with the best-known susceptibility zoning effect.

Fig. 8.

ROC curves for four methods of selecting non-landslide samples.

Fig. 7.

ROC curves for different model performance analyses.

Moreover, based on the ROC curves (Fig. 8), the AUC value of selecting non-landslide points in non-landslide areas was 0.928 and that of selecting non-landslide points in buffer zones 1000 m away from historical landslide sites was 0.924. Setting the slope equal to 3° as a threshold, the AUC value of selecting non-landslide points in areas with ≤3° was 0.989; using the information volume model for the landslide susceptibility zoning, the AUC value of non-landslide points selected in the very-low-susceptibility zone was 0.997, which was the highest AUC value of those found with the four methods. It is clear, feasible, and practical to analyze the landslide susceptibility of the entire Chishui River Basin by randomly selecting the same number of non-landslide points in the very low susceptibility zones of landslides using the information volume model.

In this study, we discarded the previous traditional models and successfully utilized CNN, LSTM, BP, and other neural network models to predict the landslide-prone zones in the Chishui River Basin. Although these three neural network models are more advanced than traditional machine learning models, they may not be able to solve the difficulty when using a large sample size in future research, because the network structure level is not yet deep. Basic neural networks can easily result in lower efficiency of models when training a large amount of data, so we could optimize the model by combining a basic neural network model with an attention module, which only focuses attention on the region of interest and can disregard the unimportant information, which would significantly improve the efficiency of the whole model.