Multi-pixel collaborative landslide susceptibility evaluation method based on UPerNet network

Abstract

The invention relates to a multi-pixel collaborative landslide susceptibility evaluation method based on a UPerNet network, and the method comprises the steps: firstly extracting landslide characteristic factors related to a landslide according to remote sensing image data, regional survey data and landslide sensitivity influence factors, and constructing a landslide disaster database in combination with historical landslide data and non-landslide data; dividing the landslide disaster database into a training set and a test set, performing normalization processing on the data, and performing online data enhancement processing; constructing a UPerNet model, regarding all evaluation samples as single-band images, inputting the single-band images into the constructed model, calculating the probability of landslide occurrence in each pixel, and carrying out training and parameter tuning on the model; and finally, verifying the model by using the test set landslide characteristic factors, and carrying out landslide susceptibility prediction based on the constructed model. Therefore, the model can more accurately predict the susceptibility of the regional landslide by utilizing the latest machine learning and statistical model and combining the collaborative analysis of the adjacent pixels.

Description

Technical Field

[0001] The invention relates to the field of geological disaster risk assessment, and in particular to a multi-pixel collaborative landslide susceptibility assessment method based on a UPerNet network. Background Art

[0002] Landslide susceptibility analysis refers to the analysis of geological environment, topography, meteorological conditions, etc. to assess the possibility and degree of danger of landslides. By correlating some of the main hazard factors that cause landslides with historical landslide events, a functional relationship between the hazard factor variable and the dependent variable of "landslide event occurrence" is constructed to predict areas where landslides may occur in the future. Landslides are a complex natural phenomenon, in which many hazard factors may play a role in the occurrence of landslide movement. These hazard factors include geology, topography, hydrology, climate, vegetation cover, soil properties and human activities. This interaction of multiple hazard factors leads to the complexity of landslide susceptibility prediction. Because not only a large number of input variables need to be considered, but also the analysis techniques involved are usually complex and time-consuming. The complexity and difficulty of landslide susceptibility research methods lies in the need to deeply understand the interrelationships and mechanisms of action between these multiple factors in order to more accurately predict the occurrence of landslides. In addition, since landslide susceptibility is affected by multiple factors, it is necessary to use multiple data sources and technical means, such as remote sensing, geographic information systems, geological surveys, etc., to obtain and integrate these data. In terms of data processing and analysis, advanced technologies such as statistical models and machine learning are needed to solve complex nonlinear relationships and pattern recognition problems between factors.

[0003] Since the 1990s, with the gradual development of algorithm technology, remote sensing technology and global positioning system, 3S technology has become an important tool for landslide susceptibility prediction. They not only replace traditional methods, but also provide more effective models for landslide susceptibility prediction, thus filling the gaps and blind spots in previous studies. However, from the perspective of geographic information science, there are still some problems in the current research on landslide susceptibility prediction. Due to the particularity of landslide hazards, the determination of landslide locations sometimes requires expert knowledge or field investigations, and the number of samples in landslide interpretation databases is usually limited and small. This situation leads to data-driven learning models being largely limited by the small sample characteristics of landslides. In the past research on susceptibility prediction, the edge characteristics and spatial shape characteristics of landslides were not effectively utilized. There is a complex nonlinear relationship between landslides and influencing factors, and simple network structures often cannot make full use of feature information. There is still much room for improvement in the relevant research on landslide susceptibility evaluation. Summary of the invention

[0004] The present invention provides a multi-pixel collaborative landslide susceptibility evaluation method based on a UPerNet network. The method proposes a new method of neighboring pixel collaboration based on a unified UPerNet network (UPerNet), and aims to utilize the advantages of UPerNet technology in feature extraction, sequence modeling and graphic data processing. By considering the information of multiple pixels, landslide-prone areas can be more accurately identified, and the possibility of misjudgment and missed judgment can be reduced, so as to better carry out landslide prevention and control work.

[0005] This application is implemented through the following technical solutions:

[0006] A multi-pixel collaborative landslide susceptibility assessment method based on UPerNet network, the method comprising the following steps:

[0007] S1. Extract the landslide characteristic factors associated with landslides based on remote sensing image data, regional survey data and landslide sensitivity influencing factors, and build a landslide hazard database based on historical landslide data and non-landslide data;

[0008] S2, dividing the landslide hazard database into a training set and a test set, normalizing the data, and performing online data enhancement processing on the landslide hazard database through an online data expansion module;

[0009] S3, construct the UPerNet model, treat all evaluation samples as single-band images, input them into the constructed model, calculate the probability of landslide occurrence in each pixel, train the UPerNet model based on the training set in the landslide hazard database, and tune the model parameters;

[0010] S4. Verify the UPerNet model based on the landslide characteristic factors of the test set in the landslide hazard database, and predict the landslide susceptibility based on the constructed UPerNet model.

[0011] In particular, the landslide hazard database refers to the data set required for building the UPerNet model.

[0012] In particular, the dataset includes regional survey data, remote sensing image data and landslide susceptibility influencing factors.

[0013] In particular, the landslide sensitivity influencing factors include ten factors, namely, elevation, slope, slope aspect, terrain undulation, plane curvature, profile curvature, structural line, river, rainfall, and road, to construct a landslide disaster susceptibility evaluation index system.

[0014] In particular, the landslide hazard database is processed by ArcGIS software and the data is uniformly processed into csv and npy files for storage.

[0015] In particular, the normalization process refers to uniformly sampling all the evaluation factors used to a spatial resolution consistent with the DEM (30×30 meters), converting the interpretation results into raster data consistent with the spatial resolution of the evaluation factors, and marking landslides as 1 and non-landslides as 0.

[0016] In particular, the online data enhancement is also called in-situ data enhancement or real-time data enhancement.

[0017] In particular, the steps of online data enhancement include ① starting to train the model; ② extracting a small amount of data from the training data set according to specified or random rules; ③ sending the original small batch data to the data amplifier, and the data amplifier performs geometric transformation of the data according to the specified rules; ④ the data enhancement module inputs the enhanced small batch data into the training model and trains it; ⑤ repeating steps ② to ④ until the specified number of training times is completed.

[0018] In particular, in each online data enhancement, the "new" data entering the model is very similar to the original data. The data enhancement is performed from two dimensions, namely, geometric shape transformation and image color transformation, through the online data enhancement module, so as to achieve gradual improvement of small batch data.

[0019] In particular, the geometric transformation of the data refers to flipping, rotating, cropping, deforming, scaling and other operations on the image data.

[0020] In particular, the UPerNet model is based on a feature pyramid network (FPN), and uses a bottom-up process to sequentially extract image features, including small-scale detailed feature information and large-scale generalized feature information.

[0021] In particular, the UPerNet model connects high-level features representing general information and low-level features representing detailed information through a top-down path with bottom-up sidelobes, so that features at all scales can contain rich information.

[0022] In particular, in step S4, all evaluation samples are regarded as single-band images, which are input into the established model in turn to calculate the probability of landslide occurrence in each pixel. According to the equal spacing method, the calculated landslide susceptibility index value is divided into five levels to achieve regional landslide susceptibility prediction.

[0023] Specifically, the five levels are extremely low landslide susceptibility area (0.0-0.2], low landslide susceptibility area (0.2-0.4], medium landslide susceptibility area (0.4-0.6], high landslide susceptibility area (0.6-0.8], and extremely high landslide susceptibility area (0.8-1.0].

[0024] The present invention proposes a new strategy for phase pixel collaboration using the UPerNet deep learning model to effectively utilize features such as landslide edges and spatial shapes in landslide susceptibility prediction. The invention effectively combines the advantages of deep learning in image processing, sequence modeling and data processing. Compared with the prior art, the present invention has the following beneficial effects:

[0025] (1) We designed an online data expansion module that successfully expanded the sample size of the dataset to three times its original size. This expansion significantly improved the accuracy of the prediction model, and compared with the model before data augmentation, UperNet improved PA, OA, Recall, and F1-score by 26%, 2%, 8%, and 18%, respectively.

[0026] (2) The UPerNet model provides the best performance in magnetic susceptibility zoning prediction. The extremely high landslide-prone areas predicted by UPerNet contain 62.4% of the landslides. This prediction result is significantly improved compared with previous deep learning models because the regional features and spatial structure of image data can be better extracted through convolution.

[0027] (3) Using the latest machine learning and statistical models, combined with neighboring pixel collaborative analysis, the model can more accurately predict landslide susceptibility. These predictions not only identify flat areas such as rivers, but also distinguish five levels of landslide-prone areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The drawings described herein are used to provide a further understanding of the embodiments of the present application, constitute a part of the present application, and do not constitute a limitation on the embodiments of the present invention.

[0029] Figure 1 It is a flow chart of a multi-pixel collaborative landslide susceptibility assessment method based on UPerNet network according to an embodiment of the present invention;

[0030] Figure 2 A schematic diagram of the area range selected for the embodiment of the present invention;

[0031] Figure 3 A spatial distribution map of landslide sample points in a landslide disaster sample collection library in an embodiment of the present invention;

[0032] Figure 4 In an embodiment of the present invention, a production process of a prediction data set is described;

[0033] Figure 5 This is a flowchart of online data enhancement according to an embodiment of the present invention;

[0034] Figure 6 This is a schematic diagram of the structure of a UPerNet network in an embodiment of the present invention;

[0035] Figure 7 A landslide susceptibility index map based on UPerNet in an embodiment of the present invention;

[0036] Figure 8 This is a landslide susceptibility zoning map based on UPerNet in an embodiment of the present invention. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical solutions and advantages of the present application clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0038] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0039] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments of the present invention can be combined with each other. It should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same and similar parts between the various embodiments can be referred to each other.

[0040] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0041] In the description of the present invention, it should be noted that the terms "upper", "lower", "inside", "outside", etc. indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, or are the orientations or positional relationships in which the inventive product is conventionally placed when in use, or are the orientations or positional relationships conventionally understood by those skilled in the art. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention.

[0042] Landslide susceptibility prediction usually involves a comprehensive analysis of terrain and other factors that may have spatial distribution patterns. Traditional prediction methods often only focus on the information of a single pixel without considering the spatial correlation and mutual influence between pixels. To solve this problem, this embodiment proposes a multi-pixel collaborative landslide susceptibility evaluation method based on the UPerNet network. Figure 1 As shown, the method mainly includes the following steps:

[0043] S1. Based on remote sensing image data, regional survey data and landslide sensitivity influencing factors, extract landslide characteristic factors related to landslides, combine historical landslide data with non-landslide data, and build a landslide disaster database;

[0044] In this embodiment, Figure 2 As shown in Figure 2, taking the Wenchuan landslide area as an example, it is located in the northwest of the Sichuan Basin and the southeast of the Aba Tibetan and Qiang Autonomous Prefecture, with an area of ​​4084 km 2 , build a risk assessment model.

[0045] In this landslide hazard database, the factors affecting landslide sensitivity include elevation, slope, slope aspect, terrain undulation, plane curvature, profile curvature, structural line, river, rainfall, and road. ArcGIS and its integrated Spatial Analyst toolbox are used for calculation and processing. The data sources of landslide susceptibility prediction factors are shown in Table 1.

[0046] Table 1. Data sources of landslide susceptibility prediction factors.

[0047] name type Spatial resolution use source Gaofen-1 image Raster Full color 2 meters Landslide Visual Interpretation Sichuan Data and Application Center of High Resolution Earth Observation System Sichuan Province Geological Hazard Survey Data point - Landslide Interpretation Reference Sichuan Provincial Department of Natural Resources Basic geographic data Line, surface 1:50000 Extract factors such as roads and rivers National Basic Geographic Information Center Digital Elevation Model Raster 30 meters Extract factors such as slope and aspect Geospatial Data Cloud Geological data Line, surface 1:200000 Extract construction line factors Sichuan Geological Survey Rainfall station data point - Extracting multi-year average rainfall factors China Meteorological Administration

[0048] A detailed, objective and accurate landslide disaster interpretation database is the basis for subsequent scientific research on landslide disasters, such as the spatial distribution pattern and distribution characteristics of landslide disasters. Combined with the detailed landslide disaster investigation point data of Wenchuan County and historical high-resolution remote sensing images. On this basis, through field investigation and on-site verification, the remote sensing interpretation of landslide disasters in Wenchuan County in the past 16 years was completed. The interpretation was mainly based on the exposure of landslides in images, vegetation destruction and accumulation body identification, and a total of 5,944 landslide disasters were interpreted.

[0049] The landslide hazard database is processed based on the open source geospatial data processing library GDAL, and the historical landslide points are marked as "1". At the same time, combined with elevation data and satellite maps, non-landslide samples are marked as "0" to obtain the spatial distribution set of landslide sample points, such as Figure 3 shown.

[0050] S2, dividing the landslide hazard database into a training set and a test set, normalizing the data, and performing online data enhancement processing on the landslide hazard database through an online data expansion module;

[0051] The landslide sample data consists of landslide attribute data (landslide or non-landslide data) and influencing factor attributes that affect the occurrence of landslides. The ultimate goal of the landslide susceptibility prediction model is to predict landslide attributes, with influencing factor attributes as input parameters. Since many factors in this embodiment come from DEM (digital elevation model) data, all influencing factors used are uniformly sampled to a spatial resolution of 30×30m, consistent with DEM. Based on the interpretation results, this embodiment converts these results into raster data with the same spatial resolution as the influencing factors (single pixel size is 30×30 m). Here, landslide data is marked as 1 and non-landslide data is marked as 0.

[0052] In order to effectively utilize the landslide edge and spatial shape features, a moving block of 64 × 64 pixels across the entire study area was introduced to construct the prediction dataset, as shown in Figure 4 As shown in Figure 2. This results in a training set of 924 samples and a test set of 396 samples. In addition, 25% of the data in the training set is divided into a validation set to select the training period and avoid overfitting.

[0053] Online data enhancement, also known as in-situ data enhancement or real-time data enhancement, has the following main steps: ① Start training the model; ② Extract a small amount of data from the training data set according to specified or random rules; ③ Send the original small batch data to the data amplifier, which performs geometric transformation according to the specified rules; ④ The data enhancement module inputs the enhanced small batch data into the training model and trains it; ⑤ Repeat steps ② to ④ until the specified number of training times is completed. See the online data enhancement process for details. Figure 5 .

[0054] Each time the data is enhanced online, the "new" data that enters the model is very similar to the original data. The online data enhancement module is used to perform data enhancement in two dimensions: geometric shape transformation and image color transformation, to achieve a gradual improvement of small batches of data.

[0055] The geometric transformation of data generally includes flipping, rotating, cropping, deforming, scaling and other operations on image data. For image semantic segmentation, the transformation method of flipping and rotating the image can effectively enhance the data. The flipping operation is a method similar to the folding of a mirror; the rotation operation refers to a clockwise or counterclockwise rotation. It should be noted that it is best to rotate in the range of 90 to 180 degrees, otherwise there will be problems with inconsistent image sample scales. In addition, the enhanced image after flipping and rotation differs from the original image only in direction and angle, and their image sizes are the same.

[0056] As an embodiment, random rotation and flipping methods are used to enhance the data of the geometric shape of landslide samples, and the number of data set samples is expanded to three times the original number of samples. Finally, 2772 training samples and 1188 test samples are obtained. While geometrically transforming the training images, the labels are also transformed accordingly, otherwise the labels will not correspond to the transformed images.

[0057] S3, build the UPerNet model, treat all evaluation samples as single-band images, input them into the established model, calculate the probability of landslide occurrence in each pixel, use the landslide hazard database to train the UPerNet model, and tune the model parameters;

[0058] UPerNet is based on the Feature Pyramid Network (FPN) and uses a bottom-up process to extract image features in sequence, including small-scale detailed feature information and large-scale generalized feature information.

[0059] As an example, the backbone network used is the classic ResNet50. UPerNet uses a top-down path to connect high-level features representing general information with low-level features representing detailed information using bottom-up side lobes, so that features at all scales can contain rich information. The network architecture is as follows Figure 6 As shown in Figure 2, since 10 impact factor layers were used in this sensitivity prediction study, the input features correspond to 10 channels.

[0060] S4. Verify the UPerNet model using the landslide characteristic factors of the test set and make landslide susceptibility prediction based on the constructed UPerNet model;

[0061] The above enhanced training set is input into the UPerNet model, and the best results are obtained by optimizing the model parameters.

[0062] All prediction samples are regarded as single-band images, and they are input into the established model in turn to calculate the probability of landslide occurrence in each pixel, and the landslide susceptibility index map is obtained, such as Figure 7 shown.

[0063] According to the equal spacing method, the calculated index values ​​are divided into five levels: extremely low landslide susceptibility zone (0-0.2], low landslide susceptibility zone (0.2-0.4], medium landslide susceptibility zone (0.4-0.6], high landslide susceptibility zone (0.6-0.8], and extremely high landslide susceptibility zone (0.8-1]. The landslide susceptibility zoning map is shown in the figure. Figure 8 shown.

[0064] The above specific implementation methods further illustrate the purpose, technical solutions and beneficial effects of the present application in detail. It should be understood that the above are only specific implementation methods of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A multi-pixel collaborative landslide susceptibility assessment method based on UPerNet network, characterized in that: The following steps are involved: S1. Extract the landslide characteristic factors associated with landslides based on remote sensing image data, regional survey data and landslide sensitivity influencing factors, and build a landslide hazard database based on historical landslide data and non-landslide data; S2, dividing the landslide hazard database into a training set and a test set, normalizing the data, and performing online data enhancement processing on the landslide hazard database through an online data expansion module; S3, construct the UPerNet model, treat all evaluation samples as single-band images, input them into the constructed model, calculate the probability of landslide occurrence in each pixel, train the UPerNet model based on the training set in the landslide hazard database, and tune the model parameters; S4. Verify the UPerNet model based on the landslide characteristic factors of the test set in the landslide hazard database, and predict the landslide susceptibility based on the constructed UPerNet model.

2. The multi-pixel collaborative landslide susceptibility assessment method based on UPerNet network according to claim 1 is characterized in that: In S1, the landslide hazard database refers to the data set required for building the UPerNet model, which includes regional survey data, remote sensing image data and landslide sensitivity influencing factors. After being processed by ArcGIS software, the data of the landslide hazard database is uniformly processed into csv and npy files for storage.

3. The multi-pixel collaborative landslide susceptibility assessment method based on UPerNet network according to claim 2 is characterized in that: The landslide sensitivity influencing factors include ten influencing factors, namely, elevation, slope, slope direction, terrain undulation, plane curvature, profile curvature, structural line, river, rainfall, and road, and a landslide disaster susceptibility evaluation index system is constructed.

4. The multi-pixel collaborative landslide susceptibility assessment method based on UPerNet network according to claim 1 is characterized in that: In S2, the normalization process refers to uniformly sampling all the evaluation factors used to a spatial resolution consistent with the DEM, converting the interpretation results into raster data consistent with the spatial resolution of the evaluation factors, and marking landslides as 1 and non-landslides as 0.

5. The multi-pixel collaborative landslide susceptibility assessment method based on UPerNet network according to claim 4 is characterized in that: The spatial resolution is 30×30 meters, and the pixel size of a single pixel in the raster data is 30×30 meters.

6. The multi-pixel collaborative landslide susceptibility assessment method based on UPerNet network according to claim 1 is characterized in that: In S2, the online data enhancement includes the following steps: ①Start training the model; ② Extract a small amount of data from the training data set according to specified or random rules; ③ Send the original small batch data to the data amplifier, which performs geometric transformation of the data according to the specified rules; ④The data enhancement module inputs the enhanced small batch data into the training model and performs training; ⑤Repeat steps ② to ④ until you complete the specified number of training sessions.

7. The multi-pixel collaborative landslide susceptibility assessment method based on UPerNet network according to claim 6 is characterized in that: The geometric transformation of the data refers to flipping, rotating, cropping, deforming, scaling and other operations on the image data.

8. The multi-pixel collaborative landslide susceptibility assessment method based on UPerNet network according to claim 1 is characterized in that: In S3, the UPerNet model is based on the Feature Pyramid Network (FPN), and uses a bottom-up process to extract image features in sequence, including small-scale detailed feature information and large-scale generalized feature information.

9. The multi-pixel collaborative landslide susceptibility assessment method based on UPerNet network according to claim 1 is characterized in that: In S4, all prediction samples are regarded as single-band images, which are input into the constructed model in sequence to calculate the probability of landslide occurrence in each pixel. According to the equal spacing method, the calculated landslide susceptibility index values ​​are divided to achieve regional landslide susceptibility prediction.

10. The multi-pixel collaborative landslide susceptibility assessment method based on UPerNet network according to claim 9, characterized in that: The calculated landslide susceptibility index values ​​are divided into five levels: extremely low landslide susceptibility area (0.0-0.2], low landslide susceptibility area (0.2-0.4], medium landslide susceptibility area (0.4-0.6], high landslide susceptibility area (0.6-0.8], and extremely high landslide susceptibility area (0.8-1.0].

Images