Debris flow dynamic reserve estimation method based on deep learning and time sequence InSAR

Abstract

The invention provides a debris flow dynamic reserve estimation method based on deep learning and time sequence InSAR. The method comprises the following steps: collecting environmental background data of a research area according to literature compilation and field investigation; performing slope object marking on the research area according to the environmental background data to obtain a marking graph; performing slope material source identification on the annotation graph through a semantic segmentation model to obtain an identification result; geometric decomposition is carried out on the deformation of the SAR satellite through a multi-track time sequence analysis method to obtain a deformation component; and calculating the debris flow dynamic reserves through the deformation component. According to the method, two-dimensional deformation components in the vertical direction and the normal direction are extracted through a semantic segmentation model and an SAR geometric imaging principle, a debris flow dynamic reserve calculation equation is constructed, and quantitative evaluation of the debris flow dynamic reserve is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of natural disaster prevention, and particularly to a method for estimating the mobile storage volume of debris flow by deep learning and temporal InSAR. Background Art

[0002] Debris flow is one of the most destructive natural disasters affecting mountain landscape evolution. Debris flow can rapidly transport a large amount of material in a short time, causing significant damage to vulnerable human settlements and infrastructure. The formation of debris flow is mainly affected by three factors: steep terrain, heavy rainfall, and abundant material sources. The material sources of debris flow mainly come from riverbed sediments, adjacent landslides, gullies, and slope erosion. Generally, the larger the volume of the material source and the more the total amount of scoured material generated, the more serious the potential debris flow disaster.

[0003] Traditional debris flow material source assessment methods, such as morphological feature investigation, stability discrimination, and proportional statistics methods, are difficult to meet the requirements of debris flow prevention and control engineering design in complex mountainous areas. In addition, not all loose materials can be directly supplied to debris flow, resulting in a significant difference between the assessment results and the actual situation. Due to the complexity of assessing DFDR, most existing studies rely on qualitative or semi-empirical and semi-quantitative methods, which can only estimate local material sources and are difficult to accurately reflect the distribution of mobile storage volume within the basin. Therefore, it is very necessary to design a method for estimating the mobile storage volume of debris flow by deep learning and temporal InSAR. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for estimating the mobile storage volume of debris flow by deep learning and temporal InSAR, which extracts two-dimensional deformation components in the vertical direction and the normal direction through a semantic segmentation model and the SAR geometric imaging principle, and constructs a calculation equation for the mobile storage volume of debris flow to achieve quantitative assessment of the mobile storage volume of debris flow.

[0005] To achieve the above purpose, the present invention provides the following solution:

[0006] A method for estimating the mobile storage volume of debris flow by deep learning and temporal InSAR, comprising the following steps:

[0007] Collect environmental background data of the study area according to literature compilation and field investigation;

[0008] Mark slope objects in the study area according to the environmental background data to obtain an annotation map;

[0009] Identify slope material sources from the annotation map through a semantic segmentation model to obtain an identification result;

[0010] The deformation of the SAR satellite is geometrically decomposed by the multi-track time series analysis method to obtain deformation components; the deformation components include: the LOS direction component, the downhill direction component, and the normal deformation component.

[0011] The debris flow reserve is calculated from the deformation components.

[0012] Optionally, the semantic segmentation model is obtained by introducing a hybrid vision transformer into the standard U-Net architecture; the hybrid vision transformer consists of an overlapping patch merging module, an efficient self-attention module, and a hybrid feed-forward network connected in sequence.

[0013] Optionally, the expression of the efficient self-attention module includes:

[0014]

[0015] where Q, K, and V are the query, key, and value of the slope source characteristics respectively, N and C are both characterization factors, R is the scaling factor, O(·) represents the complexity of the model, Reshape(·) is the reshaping operation, Linear(·) is the fully connected operation, d head is the scaling factor of the attention mechanism, is the key vector after the reshaping operation.

[0016] Optionally, the LOS direction component includes: the north-south component, the east-west component, and the up-down component, and the expression of the LOS direction component is: v los = v n sinθsinα - v e sinθcosα + v u cosθ; where v los is the deformation velocity in the LOS direction, θ is the satellite incident angle, α is the azimuth angle, v n , v e and v u are the sub-components of the north-south component, the east-west component, and the up-down component respectively.

[0017] Optionally, the expression of the deformation component is: where a and c are the deformation projection coefficients along the downhill direction and the normal direction respectively, v los is the deformation velocity in the LOS direction, v slope and v normal are the deformation velocities in the downhill direction and the normal direction respectively, β and are the average slope and slope direction of the slope source respectively, and θ is the satellite incident angle.

[0018] Optionally, the calculation formula for the debris flow reserve is: V = (A slope ·∑|v normal | + A catchment ·∑|vu |)·T; wherein, V is the volume of the DFDR, A slope is the area of the ramp source material, v normal is the deformation velocity in the normal direction, A catchment is the area excluding the ramp source material, T is the time, v u is the sub-component of the up-and-down component in the LOS direction component.

[0019] According to the specific embodiments provided by the present invention, the following technical effects are disclosed: The method for estimating the debris flow reserve of deep learning and temporal InSAR provided by the present invention includes: collecting environmental background data of the study area according to literature compilation and field surveys; marking slope objects in the study area based on the environmental background data to obtain a marked map; identifying slope object sources from the marked map through a semantic segmentation model to obtain an identification result; geometrically decomposing the deformation of the SAR satellite through a multi-track temporal analysis method to obtain deformation components; and calculating the debris flow reserve through the deformation components. This method extracts two-dimensional deformation components in the vertical and normal directions through a semantic segmentation model and the SAR geometric imaging principle, and constructs an equation for calculating the debris flow reserve, realizing the quantitative evaluation of the debris flow reserve. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0021] Figure 1 is the flow chart of the method for estimating the debris flow reserve of the present invention;

[0022] Figure 2 is the architecture diagram of the MVT-UNet model of the embodiment of the present invention;

[0023] Figure 3 is the SAR geometric imaging diagram of the embodiment of the present invention;

[0024] Figure 4 is the optical image diagram of the embodiment of the present invention;

[0025] Figure 5 is the NDVI diagram of the embodiment of the present invention;

[0026] Figure 6 is the marked diagram of the embodiment of the present invention;

[0027] Figure 7 is the spatial distribution diagram of the slope source material of the embodiment of the present invention;

[0028] Figure 8 It is a comparison chart of the validation set loss curve for the embodiments of the present invention;

[0029] Figure 9 It is a comparison chart for identifying slope material sources in the embodiments of the present invention;

[0030] Figure 10 It is a variable speed rate chart of the ascending orbit for the embodiments of the present invention;

[0031] Figure 11 It is a variable speed rate chart of the descending orbit for the embodiments of the present invention;

[0032] Figure 12 It is a two-dimensional deformation field chart for the embodiments of the present invention;

[0033] Figure 13 It is a schematic diagram of the surface normal deformation speed for the embodiments of the present invention;

[0034] Figure 14 It is a spatial distribution chart of the reserve of debris flow for the embodiments of the present invention;

[0035] Figure 15 It is a relationship chart of the reserve of debris flow, the drainage area, and the area elevation integral for the embodiments of the present invention;

[0036] Figure 16 It is a relationship chart of the reserve of debris flow, the slope, and the elevation for the embodiments of the present invention. Detailed implementation manners

[0037] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0038] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific implementation manners.

[0039] As Figure 1 shown, the present invention provides a method for estimating the reserve of debris flow using deep learning and temporal InSAR, including the following steps:

[0040] Step 100: Collect environmental background data of the study area according to literature compilation and field surveys;

[0041] Step 200: Mark slope objects in the study area according to the environmental background data to obtain a labeled map;

[0042] Step 300: Perform slope source identification on the annotated map through a semantic segmentation model to obtain the identification result;

[0043] Specifically, as Figure 2 shown, the semantic segmentation model (MVT-UNet) is obtained by introducing a hybrid vision transformer into the standard U-Net architecture; the hybrid vision transformer consists of an overlapping patch merging module, an efficient self-attention module, and a hybrid feed-forward network connected in sequence.

[0044] Furthermore, MVT-UNet enhances the extraction of slope source information by adding an efficient self-attention mechanism, which can significantly reduce the data calculation cost of the network and improve the network operation speed. The efficient self-attention module adds a scaling factor R to the original attention mechanism, thereby reducing the calculation cost of each attention mechanism. The expression of this module is as follows:

[0045]

[0046]

[0047] where Q, K, and V are the query, key, and value of the slope source features respectively, N and C are both characterization factors, R is the scaling factor, O(·) represents the complexity of the model, Reshape(·) is the reshaping operation, Linear(·) is the fully connected operation, d head is the scaling factor of the attention mechanism, is the key vector after the reshaping operation. This module first converts the feature data of N×C into feature data through the Reshape operation, and then passes through a fully connected layer to convert into

[0048] It should be noted that after the image is cropped, it enters the overlappatch merging process in the Mix-vision-Transformer module. The image information passes through two-dimensional convolution to obtain the spatial features and texture information related to the source, and parameters such as the corresponding batch size and feature channels are obtained. After overlap patch merging, efficient selfattn is used to calculate the parameters, so as to capture the different position relationships in the sequence and the long-distance dependence relationships and importance in the sequence, so as to improve the model performance and expression ability. Then, the Mix feed-forward network (Mix-FFN) is used to perform linear transformation and processing on the features at each position.

[0049] Step 400: Geometrically decompose the deformation of the SAR satellite through multi-track time series analysis to obtain deformation components; the deformation components include: LOS direction component, downhill direction component, and normal deformation component;

[0050] Specifically, the LOS direction component includes: north-south component (N-S), east-west component (E-W), and up-down component (U-D), and the expression of the LOS direction component is:

[0051] c los = v n sinθsinα - v e sinθcosα + v u cosθ;

[0052] Among them, v los is the deformation velocity in the LOS direction, with the unit of mm / year, θ is the satellite incident angle, α is the azimuth angle, v n , v e and v u are the sub-components in the N-S, E-W, and U-D directions respectively.

[0053] Due to the orbital structure of the SAR satellite, the characteristic of the SAR satellite is that the deformation component in the north-south direction can be ignored. Therefore, the data of the ascending and descending orbits of the Sentinel-1A satellite are merged, and the expression of the LOS direction component is expanded into the following expression:

[0054]

[0055] Among them, v u in this formula represents the dynamic change of the thickness of the basin sediment, and the superscript asc indicates that the data is for the ascending orbit, and the superscript des indicates that the data is for the descending orbit.

[0056] However, for the slope material source, the expanded expression cannot accurately represent the change in its thickness. Due to the action of gravity, the slope source material generally moves downward along the sliding surface. For each target point on the slope material source, a deviation coordinate system is established by combining the slope direction with the conventional three-dimensional coordinate system and the SAR range-Doppler coordinate system. As Figure 3 shown, the migration coordinate system of the slope material source is defined as the downhill, vertical, and normal directions, where the downward, vertical, and upward movements are considered positive in their respective directions, V prep represents the deformation rate perpendicular to the downhill direction, H (aspect) is the distance of the satellite along the azimuth direction, P (i,j) represents a certain point on the slope material source, Q (i,j)Represents any point other than the slope material source. Gravity mainly drives the slope source material downward along the slip surface, and the deformation in the downhill direction is more significant than that in the vertical direction. Therefore, it is assumed that the slope deformation mainly occurs along the slope and the normal direction.

[0057] Specifically, the expressions for the deformation components are as follows:

[0058]

[0059] where a and c are the deformation projection coefficients along the downhill direction and the normal direction respectively, and v los is the deformation velocity in the LOS direction, v slope and v normal are the deformation velocities in the downhill direction and the normal direction respectively, and v normal characterizes the dynamic change of the thickness of the slope material source, β and are the average slope and slope aspect of the slope material source respectively, and θ is the satellite incident angle.

[0060] Step 500: Calculate the debris flow reserve through the deformation components.

[0061] Specifically, the calculation formula for the debris flow reserve is:

[0062] V = (A slope ·∑|v normal | + A catchment ·∑|v u |)·T;

[0063] where V is the volume of the debris flow reserve (DFDR), A slope is the area of the slope source material, v normal is the deformation velocity in the normal direction, A catchment is the area not including the slope source material, T is the time, and v u is the sub-component of the up-down component in the LOS direction component.

[0064] As Figures 4 to 6 shown, in the embodiment of the present invention, a certain county in a certain province is used as the study area, and optical remote sensing images and the vegetation normalized difference index (NDVI) are collected, and a marked map of the slope material source in the study area is obtained through visual interpretation. The MVT-UNet model is used to identify the slope material source in the study area, and the results are compared and verified by selecting three other conventional models, as Figure 8 and Figure 9 shown. The ascending and descending orbit images covering the study area are processed using the time-series InSAR technology, and the annual average deformation rates of the ascending orbit and the descending orbit as shown in Figure 10 and Figure 11 are obtained respectively. Based on the SAR geometric imaging principle, the results as shown in Figure 12 andFigure 13 The two-dimensional deformation field of the study area shown, where the maximum subsidence velocity in the upper reaches is -102.2 mm / year, and the maximum uplift velocity in the upper reaches is 61.6 mm / year. Finally, the debris flow storage of each pixel in the study area is calculated, and the spatial distribution results are as Figures 14 to 16 shown. Figure 16 It shows the statistical distribution of the dynamic storage of each pixel in the study area with respect to slope and elevation. The results show that approximately 71.09% of the DFDR is concentrated in the slope range of 20° - 45°, with a peak at 40 - 45°. 61.59% of the DFDR, as the main supply area of debris flow, is located at an altitude of 1200 - 3000 m, which is consistent with the original results, verifying the correctness of the method of the present invention.

[0065] The beneficial effects of the present invention are as follows:

[0066] 1) A hybrid vision transformer is introduced into the standard U-Net architecture, enhancing the feature response of slope material sources;

[0067] 2) The two-dimensional deformation components in the vertical and normal directions in the annotation map are extracted through the SAR geometric imaging principle, enabling more accurate characterization of surface deformation.

[0068] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. For the same and similar parts among the various embodiments, reference can be made to each other.

[0069] Specific examples are applied in the present invention to elaborate on the principle and implementation manner of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea; at the same time, for those of ordinary skill in the art, based on the idea of the present invention, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. A method for estimating the reserve of debris flow using deep learning and time-series InSAR, characterized in that, It includes the following steps: Collect the environmental background data of the study area according to literature compilation and field investigation; Mark the slope objects in the study area according to the environmental background data to obtain a marked map; Identify the slope object sources of the marked map through a semantic segmentation model to obtain an identification result; Geometrically decompose the deformation of the SAR satellite through a multi-track time series analysis method to obtain deformation components; The deformation components include: LOS direction component, downhill direction component, and normal deformation component; Calculate the mudslide reserve through the deformation components.

2. The method for estimating the debris flow storage volume of deep learning and temporal InSAR according to claim 1, wherein The semantic segmentation model is obtained by introducing a hybrid vision transformer into the standard U-Net architecture; the hybrid vision transformer is composed of an overlapping patch merging module, an efficient self-attention module, and a hybrid feed-forward network connected in sequence.

3. The method for estimating the debris flow storage volume by deep learning and temporal InSAR according to claim 2, wherein The expression of the efficient self-attention module includes: Among them, Q, K, and V are the query, key, and value of the slope source characteristics respectively, N and C are both characterization factors, R is a scaling factor, O(·) represents the complexity of the model, Reshape(·) is a reshaping operation, Linear(·) is a fully connected operation, and d head is the scaling factor of the attention mechanism, is the key vector after the reshaping operation.

4. The method for estimating the debris flow reserve by deep learning and temporal InSAR according to claim 1, characterized in that The LOS direction component includes: north-south component, east-west component and up-down component, and the expression of the LOS direction component is: v los = v n sinθsinα - v e sinθcosα + v u cosθ; where, v los is the deformation velocity in the LOS direction, θ is the satellite incident angle, α is the azimuth angle, and v n , v e and v u are the sub-components of the north-south component, east-west component and up-down component respectively.

5. The method for estimating the debris flow storage volume of deep learning and temporal InSAR according to claim 1, characterized in that, The expression of the deformation component is as follows: where a and c are the deformation projection coefficients along the downhill direction and the normal direction respectively, v los is the deformation velocity in the LOS direction, v slope and v normal are the deformation velocities in the downhill direction and the normal direction respectively, β and are the average slope and aspect of the slope material source respectively, and θ is the satellite incident angle.

6. The method for estimating the reserve of debris flow using deep learning and temporal InSAR according to claim 1, characterized in that The calculation formula for the reserve of debris flow is: V = (A slope ·∑|v normal | + A catchment ·∑|v u |)·T; Among them, V is the volume of the DFDR, A slope is the area of the ramp source material, v normal is the deformation velocity in the normal direction, A catchment is the area excluding the ramp source material, T is the time, v u is the sub-component of the up and down components in the LOS direction component.

Images