A landslide susceptibility assessment method based on CNN-DF

Abstract

The invention discloses a landslide susceptibility assessment method based on CNN-DF, comprising: performing a multicollinearity test on landslide influencing factors obtained in a study area, and calculating the frequency ratio of the landslide influencing factors based on the landslide catalog; constructing a similar environment by normalizing the frequency ratio results of disaster-prone environmental factors, and constructing a landslide data set by adopting a balanced sampling method under the constraint of the similar environment; wherein the disaster-prone environmental factors are factors selected from landslide influencing factors as environmental constraints; building a landslide susceptibility model to predict the landslide susceptibility of the landslide data set, and generating a landslide susceptibility map; wherein the landslide susceptibility model is built based on a CNN module and a DF module, and high-level features are extracted from a training set by using a trained CNN module, and the high-level features are input into the DF module after expansion for training and testing; and based on the test set, the performance of the trained landslide susceptibility model is evaluated by comparing the accuracy index and the landslide susceptibility map result.

Description

Technical Field

[0001] The present invention relates to the field of deep learning, and relates to but is not limited to a landslide susceptibility assessment method based on CNN-DF. Background Art

[0002] Landslide is a natural geological phenomenon with great hazards. Early identification and prevention of landslides are important means to reduce and avoid disaster losses. Landslide susceptibility assessment can explore the complex relationship between landslides and disaster-prone environments through the selection of models and data. It can also give the probability of occurrence of unknown landslides in similar environments through comprehensive evaluation. It is an effective visualization technology for locating landslide areas and sustainable management of land resources, and can provide strong technical support for decision-makers and managers.

[0003] In landslide susceptibility assessment, the performance of existing deep learning models is closely related to the model structure and hyperparameters, resulting in uneven landslide susceptibility maps. Exploring a more ideal landslide susceptibility model is still a hot topic. Summary of the invention

[0004] In view of this, an embodiment of the present invention provides a landslide susceptibility assessment method based on CNN-DF, which at least solves the problem that the complexity of the existing deep learning model structure and its complex hyperparameters greatly lead to time-consuming and labor-intensive landslide susceptibility mapping with higher accuracy.

[0005] The technical solution of the embodiment of the present invention is achieved as follows:

[0006] The embodiment of the present invention provides a landslide susceptibility assessment method based on CNN-DF, comprising:

[0007] A multicollinearity test was conducted on the landslide influencing factors obtained in the study area, and the frequency ratio of the landslide influencing factors was calculated based on the landslide cataloging; the frequency ratio results of the disaster-prone environmental factors were normalized to construct a similar environment, and a landslide dataset was constructed using a balanced sampling method under the constraint of the similar environment, of which 70% was a training set and 30% was a test set; the disaster-prone environmental factors were factors selected from the landslide influencing factors as environmental constraints; a landslide susceptibility model was built to predict the landslide susceptibility of the landslide dataset, and a landslide susceptibility map was generated; wherein the landslide susceptibility model was built based on the CNN module and the DF module, and the trained CNN module was used to extract high-level features from the training set, which were expanded and input into the DF module for training and testing; based on the test set, the performance of the trained landslide susceptibility model was evaluated by comparing the accuracy index and the landslide susceptibility map results.

[0008] In some possible embodiments, the landslide influencing factors include: elevation, slope, profile curvature, TWI, NDVI, rainfall, land use type, lithology type, soil type, distance to roads, distance to rivers and distance to faults; the disaster-prone environmental factors are lithology type, soil type, distance to faults and distance to rivers.

[0009] In some possible embodiments, the multicollinearity test is performed on the landslide influencing factors obtained in the study area, and the frequency ratio of the landslide influencing factors is calculated based on the landslide catalog, including: using the VIF test method to evaluate the multicollinearity between the landslide influencing factors; using the frequency ratio method to quantitatively describe the correlation between the landslide influencing factors and the historical landslides.

[0010] In some possible embodiments, the frequency ratio results of disaster-pregnant environmental factors are normalized to construct similar environments, and a balanced sampling method is used to construct a landslide dataset under the constraints of similar environments, including: normalizing the frequency ratio results of disaster-pregnant environmental factors; using the curvature watershed method to divide the study area into multiple slope units, each slope unit generates a weighted one-hot encoding of a total of 29 hierarchical categories of 4 disaster-pregnant environmental factors; using the similarity of weighted one-hot encoding to divide all slope units into 205 similar environments; and selecting an equal number of landslide and non-landslide samples from each of the similar environments to establish the landslide dataset.

[0011] In some possible embodiments, the CNN module includes an input layer, a convolutional layer, a pooling layer, a fully connected layer and an output layer, the training set is input into the CNN module, and the output of the fully connected layer is used as the extracted high-level features.

[0012] In some possible embodiments, the input of the DF module is the sequence features of length 256 extracted by the CNN module, and the final output is the probability of the input sample x being classified as 1; the DF module adopts a cascade structure for feature learning, and the number of cascade layers N is automatically determined through the training process, each layer includes two random forests and two completely random forests, and the class vector generated by each forest is generated through K-fold cross-validation; the latter cascade layer receives the feature information processed by the previous level, and then concatenates the processing result with the input vector and outputs it to the next layer; when passing to a new level, the validation set is used to evaluate the results of the model, and if there is no significant performance gain, the training process is terminated.

[0013] In some possible embodiments, the performance of the trained landslide susceptibility model is evaluated based on the test set by comparing the accuracy index and the landslide susceptibility map results, including: classifying the model prediction results of the test set greater than or equal to 0.5 as landslides, and classifying them less than 0.5 as non-landslides, and obtaining a confusion matrix; calculating evaluation indicators based on the confusion matrix, including accuracy, precision, recall rate, F1 score, AUC value under the ROC curve, MSE and RMSE; generating a landslide susceptibility map based on the model prediction results, and counting the proportion of slope units with historical landslides in the extremely high susceptibility level; combining the evaluation indicators and the statistical results to evaluate the performance of the landslide susceptibility model.

[0014] The beneficial effects brought by the technical solution provided by the embodiment of the present application include at least:

[0015] In an embodiment of the present invention, firstly, the spatial distribution of historical landslides is analyzed based on the landslide influencing factors obtained in the study area; secondly, a landslide data set is constructed based on a similar disaster-prone environment; thirdly, the feature extraction capability of CNN and the classification capability of DF are used to realize landslide susceptibility prediction; finally, the performance of different models is evaluated by comparing the accuracy index and the landslide susceptibility map results. The CNN-DF model uses multiple convolutional layers of CNN to deeply mine high-level features, which are used to replace the original data input into the DF model to obtain more accurate prediction results. It solves the problem that the complexity of the existing deep learning model structure and its complex hyperparameters greatly lead to time-consuming and labor-intensive landslide susceptibility mapping with higher accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work, among which:

[0017] Figure 1 A schematic diagram of a flow chart of a landslide susceptibility assessment method based on CNN-DF provided in an embodiment of the present invention;

[0018] Figure 2 The overall technical flow chart of the landslide susceptibility assessment method based on CNN-DF provided by the embodiment of the present invention;

[0019] Figure 3A schematic diagram of landslide influencing factors and their historical landslide distribution provided by an embodiment of the present invention, wherein part (a) is the historical landslide distribution of altitude, part (b) is the historical landslide distribution of slope, part (c) is the historical landslide distribution of profile curvature, part (d) is the historical landslide distribution of terrain moisture index, part (e) is the historical landslide distribution of rainfall, part (f) is the historical landslide distribution of soil utilization type, part (g) is the historical landslide distribution of normalized vegetation index, part (h) is the historical landslide distribution of lithology, part (i) is the historical landslide distribution of soil type, part (j) is the historical landslide distribution of distance to road, part (k) is the historical landslide distribution of distance to river, and part (l) is the historical landslide distribution of distance to fault;

[0020] Figure 4 A schematic diagram of the frequency ratio results of all landslide influencing factors provided by an embodiment of the present invention, in which part (a) is the frequency ratio result of altitude, part (b) is the frequency ratio result of slope, part (c) is the frequency ratio result of profile curvature, part (d) is the frequency ratio result of terrain moisture index, part (e) is the frequency ratio result of rainfall, part (f) is the frequency ratio result of soil utilization type, part (g) is the frequency ratio result of normalized vegetation index, part (h) is the frequency ratio result of lithology, part (i) is the frequency ratio result of soil type, part (j) is the frequency ratio result of distance to road, part (k) is the frequency ratio result of distance to river, and part (l) is the frequency ratio result of distance to fault;

[0021] Figure 5 A schematic diagram of a similar environment balanced sampling method provided by an embodiment of the present invention;

[0022] Figure 6 A schematic diagram of the spatial distribution of a landslide data set provided by an embodiment of the present invention;

[0023] Figure 7 A schematic diagram of the model structure of a landslide susceptibility model provided in an embodiment of the present invention;

[0024] Figure 8 A schematic diagram of the CNN module structure provided by an embodiment of the present invention;

[0025] Fig. 9 A schematic diagram of the DF module structure provided in an embodiment of the present invention, wherein part (a) is the detailed structure of the DF module, and part (b) is an example of a single forest forming class vector process;

[0026] Fig.10 A schematic diagram of a confusion matrix and precision index provided by an embodiment of the present invention;

[0027] Fig.11Landslide susceptibility maps of the landslide susceptibility model and the comparison model provided by the embodiment of the present invention, wherein part (a) is a landslide susceptibility map generated by the SVM model, part (b) is a landslide susceptibility map generated by the CNN-SVM model, part (c) is a landslide susceptibility map generated by the RF model, part (d) is a landslide susceptibility map generated by the CNN-RF model, part (e) is a landslide susceptibility map generated by the DF model, part (f) is a landslide susceptibility map generated by the CNN-DF model, and part (g) is a landslide susceptibility map generated by the CNN model;

[0028] Fig.12 A schematic diagram of the proportion of historical landslide slope units in various landslide susceptibility levels provided by an embodiment of the present invention;

[0029] Fig.13 Schematic diagram of test set accuracy evaluation indicators provided in an embodiment of the present invention, wherein part (a) shows the mean absolute error and root mean square error results of the test set on SVM, RF, CNN, DF, CNN-SVM, and CNN-DF models, respectively; and part (b) shows the true positive rate and false positive rate results of the test set on SVM, RF, CNN, DF, CNN-SVM, and CNN-DF models, respectively. DETAILED DESCRIPTION

[0030] In order to make the purpose, technical scheme and advantages of the embodiments of the present invention clearer, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. The following embodiments are used to illustrate the present invention, but are not used to limit the scope of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0031] In the following description, reference is made to “some embodiments”, which describe a subset of all possible embodiments, but it will be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0032] It should be pointed out that the terms "first\second\third" involved in the embodiments of the present invention are only used to distinguish similar objects and do not represent a specific ordering of the objects. It can be understood that "first\second\third" can be interchanged in a specific order or sequence where permitted, so that the embodiments of the present invention described here can be implemented in an order other than that illustrated or described here.

[0033] Those skilled in the art will appreciate that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as those generally understood by those skilled in the art in the field to which the embodiments of the present invention belong. It should also be understood that terms such as those defined in general dictionaries should be understood to have meanings consistent with those in the context of the prior art, and will not be interpreted with idealized or overly formal meanings unless specifically defined as here.

[0034] Landslide susceptibility assessment is the premise and basis of landslide hazard assessment, which mainly refers to the discovery and identification of the location, scope and possibility of potential landslides at different scales, that is, landslide spatial prediction. Machine learning landslide susceptibility assessment modeling mainly includes many steps such as landslide cataloging, landslide factor extraction, model architecture construction, model training and model evaluation.

[0035] Landslide cataloging generally refers to the systematic recording and classification of known landslide events in a specific area. This concept is very important in geological hazard research and management because it provides historical data on landslide occurrences that can be used to analyze landslide patterns, causes, and possible influencing factors. Landslide cataloging is a dynamic database that needs to be updated regularly to reflect new landslide events and related research results.

[0036] At present, common landslide susceptibility modeling methods include traditional machine learning algorithms and deep learning algorithms. Traditional machine learning algorithms include support vector machines (SVM), random forests (RF), extreme echelon boosting algorithms (XGboost), and artificial neural networks (ANN). Traditional machine learning algorithms have the advantages of high prediction accuracy and low overfitting. However, due to the simple model structure, traditional machine learning algorithms cannot retain the memory of the previous task while completing the current training task, and cannot achieve knowledge reuse and generalization, resulting in the inability to greatly improve the prediction accuracy. Deep learning algorithms include convolutional neural networks (CNN), long short-term memory networks (LSTM), and graph convolutional neural networks (GCN). Deep learning overcomes the limitations of traditional machine learning algorithms. Its multi-layer structure can make full use of high-level features and add nonlinear representations to adapt to abstract features. In particular, CNN is known for its powerful feature extraction capabilities. In the application of CNN in landslide susceptibility assessment, CNN has significantly improved its accuracy compared with traditional machine learning methods, proving the applicability of CNN in landslide susceptibility assessment. Improving model accuracy is an important development direction for CNN in landslide susceptibility assessment. The simplest way is to deepen the network layer. However, deeper networks are more likely to produce gradient vanishing and explosion, making it difficult for the model to converge. An increase in the number of layers means an increase in the number of parameters, which will lead to difficulties in model training and overfitting. Combining CNN with other models is also an important method to improve model accuracy. For example, combining CNN with the gated recurrent unit GRU to extract the neighborhood features and sequence features of the landslide to improve accuracy; building a heterogeneous ensemble learning method that includes multiple deep learning models to improve accuracy. Although the model combination provides comprehensive feature learning, the increase in model size often leads to a significant increase in the number of model parameters and calculations, making model training and fitting difficult, and slow prediction. It is often time-consuming and laborious to obtain the optimal result. Therefore, how to ensure the balanced accuracy and efficiency of the combined model is an urgent problem to be solved.

[0037] Inspired by the layer-by-layer processing of deep learning algorithms, the Deep Forest (DF) algorithm came into being. This algorithm uses a cascade structure to process the raw data layer by layer and perform deep feature extraction. DF has four main characteristics: (1) Compared with deep neural networks, DF does not require complex hyperparameter tuning and can adaptively determine the model complexity based on the input data, making the training process simpler. (2) DF performs well with small sample sizes, which is particularly important in areas where data collection is challenging. (3) Due to its ensemble learning characteristics, DF has stronger noise resistance and generalization performance. (4) DF avoids gradient descent optimization, improves training speed, and reduces dependence on large-scale computing resources. Currently, DF has been widely used in image processing, time series prediction and other fields, and has achieved excellent research results. Although DF performs well in classification tasks, its application in landslide susceptibility assessment has not been explored before our work.

[0038] DF has stronger interpretability and fewer hyperparameters, and has great advantages in landslide susceptibility applications. However, the original data is insufficient to represent the landslide information, and the input DF training may not obtain the optimal prediction results. CNN is good at deep feature extraction, and DF adapts the model complexity according to the input data. Therefore, the present invention designs a landslide susceptibility model combining CNN and DF.

[0039] Figure 1 A flow chart of a landslide susceptibility assessment method based on CNN-DF provided in an embodiment of the present invention is shown in FIG. Figure 1 As shown, the method comprises at least the following steps:

[0040] Step S110, performing a multicollinearity test on the landslide influencing factors obtained in the study area, and calculating the frequency ratio of the landslide influencing factors based on the landslide catalog.

[0041] Here, various factors that affect the occurrence of landslides, including geological conditions, topography, rainfall, vegetation coverage, etc., as well as environmental factors such as lithology, soil type, distance to faults and distance to rivers, are selected as landslide influencing factors. In implementation, landslide influencing factors are obtained based on landslide cataloging and multi-source data, such as historical landslide records and Google Earth images in the target study area. Image data obtained using remote sensing technology can cover a wide range of surface information, and potential landslide areas can be preliminarily identified through image interpretation. Then, through field verification, the actual landslide conditions in these potential areas are confirmed and landslide influencing factors are obtained. Remote sensing technology can provide large-scale, high-resolution surface information, while field verification ensures the accuracy and reliability of the data.

[0042] In order to avoid the problem of collinearity between landslide influencing factors leading to model training errors and low accuracy of landslide susceptibility maps, it is necessary to conduct correlation analysis on each factor to verify its independence. Combining multicollinearity test and frequency ratio method, each landslide influencing factor is analyzed to determine their correlation and the degree of influence on landslide occurrence, so as to identify disaster-prone environmental factors.

[0043] Step S120, construct a similar environment by normalizing the frequency ratio results of the disaster-prone environmental factors, and construct a landslide data set using a balanced sampling method under the constraint of the similar environment, of which 70% is a training set and 30% is a test set.

[0044] Here, the disaster-pregnant environmental factors are factors selected from the landslide influencing factors as environmental constraints. Frequency ratio result normalization is to convert the original data into a unified scale for comparison and analysis. Slope units are selected as evaluation units, and the curvature watershed method is used to divide the study area such as the Bailong River Basin into 222,916 slope units. The landslide influencing factors are processed using zoning statistical tools to obtain the attribute values ​​of the slope units. Similar environments are constructed based on disaster-pregnant environmental factors to consider prior knowledge in the non-landslide sample selection process and complete the data set construction.

[0045] Step S130: construct a landslide susceptibility model to predict the landslide susceptibility of the landslide dataset and generate a landslide susceptibility map.

[0046] Here, the landslide susceptibility model is built based on the CNN module and the DF module. The trained CNN module is used to extract high-level features from the training set, which are then expanded and input into the DF module for training and testing.

[0047] The trained CNN has a stable feature learning ability and does not have the problem of overfitting. It can be used to extract high-quality data features. The only parameter of the DF model is the number of cascade layers, which is automatically determined through the training process. This is the first time that DF has been used for landslide susceptibility assessment, which can effectively improve the prediction accuracy while reducing the complexity of the combined model.

[0048] Step S140 , based on the test set, evaluating the performance of the trained landslide susceptibility model by comparing the accuracy index and the landslide susceptibility map result.

[0049] Here, the overall technical process is as follows Figure 2As shown, after data preparation and landslide data set construction are completed, a landslide susceptibility model, namely a CNN-DF model, is built. Finally, the performance of different models is evaluated by comparing the accuracy index and the landslide susceptibility map results. In the implementation, the comparison models include SVM, RF, CNN, DF, CNN-SVM, CNN-RF, and CNN-DF proposed in the embodiment of the present invention.

[0050] In an embodiment of the present invention, firstly, the spatial distribution of historical landslides is analyzed based on the landslide influencing factors obtained in the study area; secondly, a landslide data set is constructed based on a similar disaster-prone environment; thirdly, the feature extraction capability of CNN and the classification capability of DF are used to realize landslide susceptibility prediction; finally, the performance of different models is evaluated by comparing the accuracy index and the landslide susceptibility map results. The CNN-DF model uses multiple convolutional layers of CNN to deeply mine high-level features, which are used to replace the original data input into the DF model to obtain more accurate prediction results. It solves the problem that the complexity of the existing deep learning model structure and its complex hyperparameters greatly lead to time-consuming and labor-intensive landslide susceptibility mapping with higher accuracy.

[0051] In some embodiments, the landslide influencing factors include: elevation, slope, profile curvature, Topographic Wetness Index (TWI), Normalized Difference Vegetation Index (NDVI), rainfall, land use type, lithology type, soil type, distance to roads, distance to rivers and distance to faults; the disaster-prone environmental factors are lithology type, soil type, distance to faults and distance to rivers.

[0052] Here, lithology, soil type, distance to faults and distance to rivers are fixed environmental characteristics of landslide occurrence and serve as disaster-pregnant environmental factors. The landslide catalog includes 638 landslides that have occurred. The multiple landslide influencing factors and their historical landslides are obtained. Figure 3 shown.

[0053] In some possible embodiments, the above step S110 is further implemented by the following process: using the variance inflation factor (VIF) test method to evaluate the multicollinearity between the landslide influencing factors; using the frequency ratio method to quantitatively describe the correlation between the landslide influencing factors and the historical landslides.

[0054] Here, in statistical analysis, multicollinearity refers to the high correlation between independent variables in the model, which may lead to unstable estimation and difficulty in interpretation of regression coefficients. Multicollinearity can be detected using the VIF test. A VIF value greater than 10 is usually considered a signal of multicollinearity. The VIF formula is:

[0055] Formula (1);

[0056] in, It represents the multiple correlation coefficient of any landslide influencing factor for other landslide influencing factors in regression analysis, and T is the tolerance. The results show that the VIF values ​​of the selected landslide influencing factors are all over 5, reflecting that there is no obvious correlation between the 12 landslide influencing factors.

[0057] Frequency Ratio (FR) is a method for evaluating landslide influencing factors. It evaluates the impact of landslides by comparing the factor frequencies of landslide-affected areas with those of landslide-free areas. The embodiment of the present invention uses the frequency ratio method to quantitatively describe the correlation between the landslide tendency of landslide influencing factors in the classified sub-intervals or sub-categories and the historical landslides that occurred in the study area, and verifies the accuracy of factor classification. The frequency ratio calculation formula is as follows:

[0058] Formula (2);

[0059] In the formula Landslide influencing factor The number of landslide grids under classification category j, is the total number of landslide grids in the study area, Landslide influencing factor The number of grids under classification category j, is the total number of grids in the study area, Landslide influencing factor Landslide frequency ratio under classification category j.

[0060] The frequency ratio results of all landslide influencing factors are as follows Figure 4As shown in the figure, the elevation results show that landslides mostly occur below 3000m above sea level, and the landslide tendency is the largest in the range of 1500-2000 meters, which is consistent with the altitude along the river. The slope range with the largest landslide tendency is 30°-45°. The landslide tendency is similar in the profile curvature range of (-20,0) and (0,20). The range value of TWI is generally negatively correlated with the landslide tendency. The annual average rainfall results show that landslides mostly occur in the range of 600-900 mm / year. Rainfall will destroy the stability of the slope and induce landslides. The land use results show that the forest land has the greatest landslide tendency. The strong evaporation of the forest will induce rainy weather and increase the risk of landslides. The NDVI results show that the landslide tendency is greater in the range of 0.4-0.7, indicating that vegetation coverage indirectly affects slope stability. The lithology results show that under the conditions of mixed sedimentary rocks and carbonate sedimentary rocks, the FR value is larger, indicating that the soft rock slope has a greater landslide tendency. The soil type results show that under the conditions of conical soil and highly active leached soil, the FR value is large, indicating that fine-textured soil has a greater tendency to landslide. These lithology types and soil types with greater landslide tendency are mostly distributed along the river. The results of the distance to roads, rivers and faults show that the landslide tendency and distance value are negatively correlated. Compared with other factors, the FR results of lithology, soil type, distance to rivers and distance to faults show very clear landslide tendencies, which confirms the rationality of the disaster-prone environmental factors selected in the embodiments of the present invention.

[0061] In some possible embodiments, the above step S120 "using the frequency ratio results of the disaster-prone environmental factors to normalize and construct a similar environment, and using a balanced sampling method to construct a landslide data set under the constraints of the similar environment" is further implemented by the following steps:

[0062] S121, normalize the frequency ratio results of disaster-prone environmental factors.

[0063] Here, the normalized expression of the frequency ratio result is as follows:

[0064] Formula (3);

[0065] In the formula, Environmental factors Rating category Environmental factors weights between typical categories of landslides occurring below; Environmental factors In the classification category Maximum frequency ratio under Environmental factors In the classification category The minimum frequency ratio under .

[0066] S122, the curvature watershed method was used to divide the study area into multiple slope units, and each slope unit generated weighted one-hot encoding of a total of 29 hierarchical categories of all disaster-prone environmental factors.

[0067] Here, the disaster-prone environmental factors of each slope unit are weighted and a one-hot encoding is generated. One-hot encoding is a method of converting categorical data into numerical data, and weighted one-hot encoding further considers the importance or weight of different categories. Disaster-prone environmental factors are fixed environmental characteristics of landslides, such as lithology, soil type, distance to faults, and distance to rivers. The distance to faults (in meters) factor includes five classification categories: <2000, 2000~4000, 4000~8000, 8000~18000, and >18000; the distance to rivers (in meters) factor includes <400, 400~1200, 1200~4800, 4800~1 There are six classification categories, including 0000, 10000~24000, and >24000; the lithology types include acidic plutonic rock, acidic volcanic rock, loose sediments, carbonate sedimentary rock, neutral plutonic rock, neutral volcanic rock, siliceous clastic sedimentary rock, mixed sedimentary rock, basic plutonic rock, and metamorphic rock, a total of ten classification categories; the soil types include: anthropogenic soil, chernozem, embryonic soil, thin layer soil, highly active leached soil, black soil, clay pan soil, and loose lithologic soil, a total of eight classification categories.

[0068] S123, all slope units were divided into 205 similar environments using weighted one-hot-encoded similarity.

[0069] Here, a clustering algorithm is used to divide the slope units into multiple similar environments, and the units in each similar environment have similar disaster-prone environmental characteristics. Based on these weighted unique hot codes, the slope units are divided into 205 similar environments according to similarity, which helps to consider the similarity of the environment when selecting samples.

[0070] S124, selecting an equal number of landslide and non-landslide samples from each of the similar environments to establish the landslide data set.

[0071] Here, the embodiment of the present invention adopts a balanced sampling method of landslide samples and non-landslide samples based on the similarity of disaster-prone environments. For example, 638 historical landslide surfaces were processed to obtain 864 landslide slope units. In each similar environment, landslide slope units are evenly sampled as landslide samples and non-landslide slope units are sampled as non-landslide samples. The method for selecting non-landslide slope units based on landslide slope units and similar environments is as follows: Figure 5As shown, loop through each similar environment to determine whether the number of landslide samples is greater than 1. If so, put the landslide samples into the training set and test set at a ratio of 7:3. Further, when the number of unlabeled samples in the similar environment is greater than or equal to the number of landslide samples, randomly select unlabeled samples equal to the number of landslide samples as non-landslide samples. If the number of non-landslide samples is greater than or equal to 2, put them into the training set and test set at a ratio of 7:3, otherwise put them into the training set. Repeat the above process until all similar environments are traversed, and output the final training set and test set. Similar environment sampling can further improve the ability to identify potential landslides. The spatial distribution of the constructed landslide dataset is shown in the figure below. Figure 6 This method takes into account the environmental factors of landslide occurrence and helps the model learn more accurate landslide prediction features.

[0072] In some possible embodiments, the CNN module includes an input layer, a convolutional layer, a pooling layer, a fully connected layer and an output layer, the training set is input into the CNN module, and the output of the fully connected layer is used as the extracted high-level features.

[0073] Here, the present invention establishes a landslide susceptibility model that couples CNN and DF. The model structure is as follows: Figure 7 As shown in Figure 2, the CNN module uses convolution operations to learn high-level feature representations of the original features, expands them into large dimensions, and then uses the expanded features as input to the deep forest. The DF module uses its strong classification ability to achieve landslide susceptibility prediction.

[0074] CNN module: CNN is a special feed-forward deep learning algorithm with powerful feature extraction and nonlinear computing capabilities. It can directly extract obvious features from raw data, thus effectively avoiding the problem that traditional deep learning methods need to use complex mathematical methods to extract data features. Its network structure mainly includes input layer, convolution layer, pooling layer, fully connected layer and output layer. The CNN structure provided in the embodiment of the present invention is as follows: Figure 8 As shown in the figure, the original features are input, and after training, the CNN susceptibility model is obtained, and the sigmoid function is used for classification. The trained CNN model is used for feature extraction, the original features are input, and the output of the first Dense layer, i.e. the fully connected layer, is used as the extracted high-level features.

[0075] In some possible embodiments, the input of the DF module is the sequence features of length 256 extracted by the CNN module, and the final output is the probability of the input sample x being classified as 1; the DF module adopts a cascade structure for feature learning, and the number of cascade layers N is automatically determined through the training process, each layer includes two random forests and two completely random forests, and the class vector generated by each forest is generated through K-fold cross-validation; the latter cascade layer receives the feature information processed by the previous level, and then concatenates the processing result with the input vector and outputs it to the next layer; when passing to a new level, the validation set is used to evaluate the results of the model, and if there is no significant performance gain, the training process is terminated.

[0076] Deep Forest is a deep learning model based on random forest. Similar to the layer-by-layer processing of original features in deep neural networks, Deep Forest uses a cascade structure for feature learning, thereby making full use of deep features to improve classification results. Fig. 9 As shown in part (a), the input of the deep forest model is a sequence feature of length 256 extracted by CNN, and the final output is the probability of the input sample x being classified as 1, that is, the probability of landslide susceptibility. Binner discretizes the feature value into multiple unique values. This module accelerates the construction of decision trees. The model depth is N layers, and each level contains two random forests and two completely random forests. The completely random forest consists of multiple trees, each tree contains all the features, and randomly selects a feature as the splitting node of the splitting tree. The splitting continues until each leaf contains a category. The random forest is also composed of multiple trees, and each tree randomly selects features, and then determine the split node based on the Gini score, the formula is:

[0077] Formula (4);

[0078] In the formula is the number of sample classifications, and in the embodiment of the present invention, the value is 2. is the number of samples in the current node. The proportion of class samples.

[0079] The process of forming a class vector for each forest is as follows Fig. 9 As shown in part (b) of the figure. To reduce the risk of overfitting, the class vector generated by each forest is generated through K-fold cross validation. The next cascade layer receives the feature information processed by the previous level, and then concatenates the processing result with the input vector and outputs it to the next layer. When passing to the new level, the validation set is used to evaluate the results of the model. If there is no significant performance gain, the training process is terminated. Therefore, the number of cascade layers is automatically determined by the model.

[0080] In some possible embodiments, the performance of the trained landslide susceptibility model is evaluated based on the test set by comparing the accuracy index and the landslide susceptibility map results, including: classifying the model prediction results of the test set greater than or equal to 0.5 as landslides, and classifying them less than 0.5 as non-landslides, and obtaining a confusion matrix; calculating evaluation indicators based on the confusion matrix, including accuracy, precision, recall rate, F1 score, AUC value under the ROC curve, MSE and RMSE; generating a landslide susceptibility map based on the model prediction results, and counting the proportion of slope units with historical landslides in the extremely high susceptibility level; combining the evaluation indicators and the statistical results to evaluate the performance of the landslide susceptibility model.

[0081] Here, the confusion matrix is ​​a matrix representation used to measure the performance of the classification model. The model prediction results of the test set are classified as landslide or non-landslide according to whether they are greater than a specific threshold of 0.5. Further, the confusion matrix can be obtained to evaluate the performance of the model. Fig.10 As shown in the figure, based on the confusion matrix, the evaluation indicators Accuracy, Precision, Recall, F1 score, True Positive Rate (TPR) and False Positive Rate (FPR) are calculated. Among them, TP (True Positive) means that the actual positive class is predicted to be positive, TN (True Negative) means that the actual negative class is predicted to be negative, FP (False Positive) means that the actual negative class is predicted to be positive, and FN (True Negative) means that the actual positive class is predicted to be negative.

[0082] FPR is the X-axis of the Receiver Operating Characteristic (ROC), TPR is the Y-axis of the ROC, and the AUC (Area Under the Curve) value is also one of the commonly used performance indicators for evaluating models. The larger the AUC value, the better the performance of the evaluation model and the higher the model accuracy. In order to obtain more scientific evaluation results, the present invention also introduces the error evaluation indicators Mean-Square Error (MSE) and Root Mean Squared Error (RMSE), and the formula is as follows:

[0083] Formula (5);

[0084] Formula (6);

[0085] in: is the number of samples, and Respectively represent The true value and predicted value of a sample. They measure the predictive ability of the model by calculating the difference between the true value and the predicted value. In particular, RMSE is sensitive to outliers and can better reflect the distribution of prediction errors.

[0086] The above-mentioned landslide susceptibility assessment method based on CNN-DF is described below in conjunction with a specific embodiment. However, it should be noted that this specific embodiment is only for better illustrating the present invention and does not constitute an improper limitation to the present invention.

[0087] This specific embodiment takes the Bailong River Basin, which is severely affected by landslide disasters, as the study area. First, based on the landslide catalog and multi-source data, 12 landslide influencing factors are obtained to analyze the spatial distribution of historical landslides; secondly, under the constraint of similar environments, equal amounts of landslide and non-landslide samples are selected to establish a landslide dataset; thirdly, CNN is used to extract high-level features from the original data and input into the DF model for training and testing. Finally, the model is applied to the prediction of landslide susceptibility in the Bailong River Basin.

[0088] The landslide susceptibility diagrams of the CNN-DF model of the embodiment of the present invention and the comparative model are shown in FIG. Fig.11. In order to more clearly show the differences in the model susceptibility results, some areas are enlarged to show more specific details. In general, most of the extremely high susceptibility areas of CNN, DF and CNN-DF are distributed on both sides of the river. Concentrated extremely high susceptibility appears in the river gathering area in the middle of the study area, which corresponds to the spatial distribution of rainfall in the study area. In the results of CNN, concentrated extremely high susceptibility areas also appear in the northwest boundary area and the southeast boundary. Compared with the spatial distribution of rainfall, this area belongs to the area with high rainfall, indicating that the CNN model is more sensitive to the characteristics of rainfall factors. In the CNN-DF model, the area of ​​these concentrated areas has been reduced. Model improvement has played a role in the accuracy of landslide susceptibility map. In terms of the proportion of areas of each susceptibility level, the extremely low susceptibility area of ​​the DF model accounts for only 25.1%, and the extremely high susceptibility area accounts for 6.22%, which is not conducive to landslide prevention and control. However, the proportion of extremely low susceptibility and extremely high susceptibility areas of the CNN model is relatively high, which is also not conducive to landslide prevention and control. The CNN-DF model just solves this problem. Its extremely low susceptibility and extremely high susceptibility area ratios are in the middle of the two model levels, which improves reliability. The same phenomenon also appears in the results of CNN-SVM and CNN-RF. The area ratios of extremely low and extremely high susceptibility of the hybrid model are significantly reduced, indicating that the improvement strategy of the hybrid model is meaningful. The susceptibility results of the DF model are compared with those of the RF, SVM and CNN models. The results of RF and DF are similar. The extremely high susceptibility area of ​​SVM and CNN is too large at the northwest boundary, which affects the reliability of the results.

[0089] The proportion of slope units with historical landslides in the extremely high susceptibility class can also be used as a method to verify the reliability of landslide susceptibility maps. Fig.12 As shown, the embodiment of the present invention counts the proportion of the number of landslide slope units in each level of the seven models. Among them, in the CNN-DF model, 90.76% of the slope units where landslides occurred fell into the extremely high susceptibility area, which is the highest percentage among all models, 0.58% higher than the DF model and 3.51% higher than the CNN model. The same improvement trend was also found in the CNN-SVM and CNN-RF models. These phenomena can also be observed in the detail diagram. It shows that the hybrid model improves the accuracy of model prediction. Compared with the single model, the proportion of historical landslides in the extremely high susceptibility area of ​​the DF model is higher than that of the SVM and RF models, indicating that DF has excellent prediction ability, which is closely related to the optimal performance shown by the CNN-DF model. The statistical results of the landslide slope units are consistent with the index evaluation results of the model, and the performance and accuracy of the CNN-DF model provided by the present invention are optimal.

[0090] The hybrid model is an important improvement idea to improve the performance of machine learning models. The performance of the DF model is related to the performance of the hybrid model. In order to verify the performance of the DF model, three common machine learning models, random forest, support vector machine and CNN, were also used for susceptibility modeling, and their model parameters were determined by trial algorithms. In order to determine whether the features after convolution operation using CNN improve the prediction results of deep forest compared to the original features, the test set accuracy evaluation indicators of CNN model, deep forest model and CNN-DF model are shown in Tables 1 and Fig.13 . Similar to the hybrid model CNN-DF, the present invention also establishes CNN-RF and CNN-SVM models. The accuracy of CNN-DF is not only higher than that of DF and CNN, but also better than all single models and hybrid models. DF is significantly higher than the SVM, RF and CNN models in terms of evaluation indicators, indicating that both the DF model and the CNN-DF model have good fitting accuracy and prediction performance. The features processed by CNN are input into the deep forest classification model, and the prediction performance of the model is greatly improved, which can produce more accurate susceptibility results. Compared with the original features, the deep forest classifier can achieve better recognition with the new features.

[0091] Table 1 Accuracy indicators of the test set in each model

[0092] ;

[0093] The present invention is a landslide susceptibility assessment method based on CNN-DF, which has the advantage that the ability of CNN learning features after training in the CNN module is stable, there is no overfitting problem, and high-quality data features can be extracted by using it. The only parameter of the DF model is the number of cascade layers, and this parameter is automatically determined through the training process. The CNN-DF model of the present invention only needs to pay attention to the structure of CNN to construct a highly robust CNN. Comparing the various accuracy evaluation indicators of the test set, the performance of the CNN-DF model is improved relative to DF. Further comparing the proportion of historical landslide slope units, the results of CNN-DF also showed an improvement of 0.58% relative to DF. The results obtained by the CNN-DF model are optimal.

[0094] It should be noted that, in the embodiment of the present invention, if the above-mentioned landslide susceptibility assessment method based on CNN-DF is implemented in the form of a software function module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present invention, or the part that contributes to the relevant technology, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for enabling an electronic device to execute all or part of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a U disk, a mobile hard disk, a read-only memory (ROM), a magnetic disk or an optical disk. In this way, the embodiment of the present invention is not limited to any specific combination of hardware and software.

[0095] It should be understood that "one embodiment" or "an embodiment" mentioned throughout the specification means that specific features, structures or characteristics related to the embodiment are included in at least one embodiment of the present invention. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. In addition, these specific features, structures or characteristics can be combined in one or more embodiments in any suitable manner. It should be understood that in various embodiments of the present invention, the size of the serial number of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present invention. The serial numbers of the above-mentioned embodiments of the present invention are only for description and do not represent the advantages and disadvantages of the embodiments.

[0096] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0097] In the several embodiments provided by the present invention, it should be understood that the disclosed methods can be implemented in other ways. The methods disclosed in the several method embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method embodiments. The features disclosed in the several method embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method embodiments.

[0098] The above is only an embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A landslide susceptibility assessment method based on CNN-DF, characterized in that: include: Multicollinearity tests were conducted on the landslide influencing factors obtained in the study area, and the frequency ratios of landslide influencing factors were calculated based on the landslide inventory; The frequency ratio of landslide influencing factors is calculated by the following formula: ; In the formula Landslide influencing factor In the grading category The number of landslide grids under is the total number of landslide grids in the study area, Landslide influencing factor In the grading category The number of grids below, is the total number of grids in the study area, Landslide influencing factor In the grading category The landslide frequency ratio under ; The frequency ratio results of disaster-pregnant environmental factors are normalized to construct a similar environment. Under the constraint of the similar environment, a balanced sampling method is used to construct a landslide data set, of which 70% is a training set and 30% is a test set. The disaster-pregnant environmental factors are factors selected from the landslide influencing factors as environmental constraints. A landslide susceptibility model is constructed to predict the landslide susceptibility of the landslide data set and generate a landslide susceptibility map; wherein the landslide susceptibility model is constructed based on a convolutional neural network (CNN) module and a deep forest (DF) module, and high-level features are extracted from the training set using a trained CNN module, which is then expanded and input into the DF module for training and testing; the high-level features are features output by a fully connected layer of the CNN module; Based on the test set, evaluating the performance of the trained landslide susceptibility model by comparing accuracy indicators and landslide susceptibility map results; The method uses the frequency ratio results of disaster-pregnant environmental factors to normalize and construct similar environments, and uses a balanced sampling method under the constraints of similar environments to construct a landslide data set, including: normalizing the frequency ratio results of disaster-pregnant environmental factors; using the curvature watershed method to divide the study area into multiple slope units, each slope unit generates a total of 29 graded weighted one-hot encodings of all disaster-pregnant environmental factors; using the similarity of weighted one-hot encodings to divide all slope units into 205 similar environments; selecting equal amounts of landslide and non-landslide samples from each of the similar environments to establish the landslide data set; the expression of the normalized frequency ratio results is as follows: ; In the formula, Environmental factors Rating category Environmental factors The weights between the categories where landslides occur; Environmental factors In the grading category Maximum frequency ratio under Environmental factors In the grading category The minimum frequency ratio under The input of the DF module is the sequence feature of length 256 extracted by the CNN module, and the final output is the probability of the input sample x being classified as 1; the DF module adopts a cascade structure for feature learning, and the number of cascade layers N is automatically determined through the training process, each layer includes two random forests and two completely random forests, and the class vector generated by each forest is generated by K-fold cross validation; the next cascade layer receives the feature information processed by the previous level, and then splices the processing result with the input vector and outputs it to the next layer; when passing to the new level, the verification set is used to evaluate the model results; The method of evaluating the performance of the trained landslide susceptibility model based on the test set by comparing the accuracy index and the landslide susceptibility map results includes: classifying the model prediction results of the test set greater than or equal to 0.5 as landslides, and classifying them less than 0.5 as non-landslides, and obtaining a confusion matrix; calculating evaluation indicators based on the confusion matrix, including accuracy, precision, recall rate, F1 score, AUC value under the ROC curve, mean square error (MSE) and root mean square error (RMSE); generating a landslide susceptibility map based on the model prediction results, and counting the proportion of slope units with historical landslides in the extremely high susceptibility level; and evaluating the performance of the landslide susceptibility model by combining the evaluation indicators and statistical results.

2. The method according to claim 1, characterized in that The landslide influencing factors include: elevation, slope, profile curvature, terrain wetness index TWI, normalized difference vegetation index NDVI, rainfall, land use type, lithology type, soil type, distance to roads, distance to rivers and distance to faults; the disaster-prone environmental factors are lithology type, soil type, distance to faults and distance to rivers.

3. The method according to claim 1, characterized in that The multicollinearity test is conducted on the landslide influencing factors obtained in the study area, and the frequency ratio of the landslide influencing factors is calculated based on the landslide catalog, including: The variance inflation factor (VIF) test method was used to evaluate the multicollinearity among the landslide influencing factors. The frequency ratio method is used to quantitatively describe the correlation between landslide influencing factors and historical landslides.

4. The method according to any one of claims 1 to 3, characterized in that: The CNN module includes an input layer, a convolution layer, a pooling layer, a fully connected layer and an output layer. The training set is input into the CNN module, and the output of the fully connected layer is used as the extracted high-level features.

Images