Method for evaluating landslide-debris flow disaster chain based on graph neural network

Abstract

The invention discloses a landslide-debris flow disaster chain assessment method based on a graph neural network, and the method comprises the steps: generating a composite ground feature unit through the watershed segmentation of a composite curvature field and hydrological analysis, abstracting the composite ground feature unit as a geographic node, optimizing a multi-source environment factor through mutual information screening, constructing a directed weighted graph in combination with spatial similarity, and carrying out the reconstruction of a landslide-debris flow disaster chain. An improved GraphSAGE model which introduces neighbor weight, self-loop and residual error information and cancels random sampling is adopted, a sample set adaptive to a less-data area is constructed through a unit splitting-feature matching-terrain fitting mechanism, and model training and risk prediction are completed. The method solves the problems of single disaster assessment, strong sample dependence and high missed judgment rate in the traditional technology, has excellent performance indexes, effectively reduces the missed judgment risk of a high-risk area, and provides accurate and efficient technical support for disaster chain risk assessment of a small watershed in a small-data mountainous area.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of geological disaster detection, and particularly relates to a landslide-debris flow disaster chain evaluation method based on a graph neural network. BACKGROUND

[0002] Landslides and debris flows are common geological disasters in mountainous areas, and there is a close correlation between the two in the process of occurrence. In some areas, loose deposits generated after landslides often become an important source of debris flow, and under the action of environmental factors such as precipitation and topography, these loose sources are easily transported by water flow, thereby inducing debris flow and forming a landslide-debris flow disaster chain. This type of disaster chain will have an impact on the ecological environment, infrastructure and personnel safety in the region, so it is of practical significance to carry out risk assessment on landslide-debris flow disaster chains.

[0003] At present, there has been some research on risk prediction technology for landslides and debris flows, but there are still many deficiencies. Firstly, existing technologies are mostly focused on risk evaluation of a single disaster, and there is no systematic quantitative evaluation framework for landslide-debris flow disaster chains, making it difficult to fully depict the causal transmission relationship between landslides and debris flows. Secondly, traditional evaluation methods rely too much on high-quality training samples, and in areas where the number of samples is insufficient or the quality is not high, the model prediction performance may decrease. Thirdly, the spatial feature expression ability of most methods is limited, and they cannot effectively capture the potential correlation between similar geographical environments in distant areas, making it difficult for the evaluation results to cover all risk characteristics in the region. Fourthly, existing methods are often limited by the computational complexity of modeling the entire map, and it is difficult to achieve efficient operation when carrying out risk evaluation in large-scale regions (such as cross-city, cross-province basins), limiting the popularization and application of the technology in practical scenarios. SUMMARY

[0004] In order to solve the problems existing in the prior art, the present application provides a landslide-debris flow disaster chain evaluation method based on a graph neural network, which aims to provide a systematic technical path for disaster chain risk evaluation by constructing a risk evaluation system for landslide-debris flow disaster chains.

[0005] The technical solution adopted by the present application is as follows: In a first aspect, the present application provides a landslide-debris flow disaster chain evaluation method based on a graph neural network, comprising the following steps: Step 100: generating a composite ground feature unit by composite curvature field watershed segmentation and hydrological analysis, and abstracting the composite ground feature unit as a geographic node; Step 200: constructing node features corresponding to the geographic node, and verifying the independence of each environmental factor in the node features using a mutual information method; Step 300: a directed weighted graph is constructed according to the environmental factors and the spatial similarity, and the directed weighted graph is used to reflect the spatial transmission relationship of the landslide-debris flow disaster chain; Step 400: an improved GraphSAGE model is designed, and information in the directed weighted graph is transmitted and node representation is optimized by the improved GraphSAGE model; Step 500: the improved GraphSAGE model is trained and verified, and risk prediction and spatial mapping of the landslide-debris flow disaster chain are completed.

[0006] With reference to the first aspect, in a first implementation manner of the first aspect, the process of generating the composite ground feature unit in the step 100 comprises: first obtaining a slope unit through composite curvature field watershed segmentation, and then introducing a gully buffer zone on the basis of the slope unit, and further forming the composite ground feature unit.

[0007] With reference to the first implementation manner of the first aspect, in a second implementation manner of the first aspect, after the composite ground feature unit is formed in the step 100, the operation of quantifying the source area features and the flow-through area features corresponding to the composite ground feature unit is further included.

[0008] With reference to the first aspect, in a third implementation manner of the first aspect, the process of constructing the node features in the step 200 comprises: collecting multi-source heterogeneous environmental factors, and taking the multi-source heterogeneous environmental factors as initial input factors of the node features; and the step of verifying the independence of each environmental factor by using the mutual information method specifically comprises: calculating the correlation degree between each initial input factor by using a normalized mutual information method.

[0009] With reference to the third implementation manner of the first aspect, in a fourth implementation manner of the first aspect, after the correlation degree between each initial input factor is calculated in the step 200, the operation of removing a strongly dependent factor according to the correlation degree is further included, and the strongly dependent factor is an initial input factor with a correlation degree exceeding a preset threshold.

[0010] With reference to the first aspect, in a fifth implementation manner of the first aspect, the process of designing the improved GraphSAGE model in the step 400 comprises: introducing a neighbor weight mechanism, retaining self-loop information, retaining residual information, and canceling a random sampling operation for the GraphSAGE model.

[0011] With reference to the fifth implementation manner of the first aspect, in a sixth implementation manner of the first aspect, the step of introducing the neighbor weight mechanism specifically comprises: according to the influence degree of neighbor nodes of each geographical node on the geographical node, differentiating weights of different neighbor nodes.

[0012] In combination with the fifth implementation manner of the first aspect, the present application provides a seventh implementation manner of the first aspect, and the reserved self-loop information is specifically that: when the improved GraphSAGE model performs information transmission, the feature information of each geographical node itself is reserved, and the information contribution of the geographical node to itself is not excluded; and the reserved residual information is specifically that: when the improved GraphSAGE model performs node representation optimization, the node representation result of the previous round is connected with the node representation result of the current round in residual, and the historical representation information is reserved.

[0013] In combination with the first aspect, the present application provides an eighth implementation manner of the first aspect, and before the information transmission and node representation optimization of the improved GraphSAGE model in the step 400, the operation of constructing a disaster chain sample set is further included, and the construction of the disaster chain sample set adopts a hierarchical sampling strategy, and positive and negative samples are counted according to environmental units to ensure balanced sample distribution.

[0014] In combination with the eighth implementation manner of the first aspect, the present application provides a ninth implementation manner of the first aspect, and in the process of constructing the disaster chain sample set, the operation of splitting the disaster chain label is further included, and the data diversity of the disaster chain sample set is enhanced through the label splitting.

[0015] The present application has the following beneficial effects: (1) The present application constructs a risk assessment system for landslide-debris flow disaster chain, abstracts disasters and environmental factors into a graph structure, realizes modeling and causal relationship quantification of landslide-debris flow disaster chain, makes up for the limitation of single disaster research in the prior art, and provides a systematic technical path for disaster chain risk assessment; (2) The present application uses an improved GraphSAGE model and a transfer learning mechanism, reduces the dependence on high-quality training samples, still maintains good model performance in a sample-scarce area, improves the generalization ability of the model in a regional expansion scenario, and is helpful to expand the application scope of the technology; (3) The present application enhances the spatial feature expression ability by constructing a directed and weighted graph according to environmental factors and spatial similarity, can effectively capture the potential correlation between distant and geographically similar areas, and makes the risk assessment result more comprehensive and more in line with the actual disaster distribution characteristics; (4) The present application improves the GraphSAGE model, introduces a neighbor weight mechanism, reserves self-loop and residual information, and cancels random sampling, optimizes the information transmission and node representation process, reduces the computational complexity of whole-graph modeling, is conducive to efficient risk assessment in a large-scale area, and improves the practical application value of the technology; (5) The present application performs redundancy screening and optimization on multi-source heterogeneous environmental factors through the mutual information method, quantifies factor independence and eliminates strong dependent factors, thereby improving the robustness of input data; at the same time, the hierarchical sampling strategy is used to construct a disaster chain sample set and support label splitting, so as to ensure balanced sample distribution and enhance data diversity, provide a more reliable data basis for model training, and further improve the stability of risk prediction results. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is a schematic diagram of the information transmission of the GraphSAGE model improved in the embodiment of the present application; Figure 2 is a schematic diagram of the node representation optimization of the GraphSAGE model improved in the embodiment of the present application; Figure 3 is a schematic diagram of risk prediction mapping in the embodiment of the present application; Figure 4 is a schematic diagram of risk assessment of the target area in the embodiment by the random forest method in the prior art; Figure 5 is a schematic diagram of risk assessment of the target area in the embodiment by the support vector machine method in the prior art; Figure 6 is a schematic diagram of risk assessment of the target area in the embodiment by the multilayer perception machine method in the prior art; Figure 7 is a schematic diagram of risk assessment of the target area in the embodiment by the method in the embodiment of the present application. DETAILED DESCRIPTION

[0017] The present application will be further explained in conjunction with the accompanying drawings and specific embodiments.

[0018] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all the embodiments. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations.

[0019] Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without making creative efforts belong to the scope of protection of the present application. EMBODIMENTS

[0020] The scenario to be solved by the embodiment is risk assessment of landslide-debris flow disaster chain in small mountainous watershed. The terrain of such area is large in relief, and gullies are densely distributed. Landslides are easy to supply gullies with loose deposits and then induce debris flow. However, the existing assessment steps can be further refined in terms of adapting to the terrain characteristics, environmental factor specificity and model information transmission pertinence of small watershed, so as to better fit the disaster chain occurrence regularity of small watershed.

[0021] The embodiment discloses a landslide-debris flow disaster chain assessment method based on a graph neural network, comprising the following steps: Step 100: According to the terrain characteristics of small mountainous watershed, first, the compound curvature field watershed segmentation is used to obtain the slope unit adapting to the slope gradient and slope direction characteristics of the region, and then the gully buffer zone covering the main confluence area is introduced based on the slope unit through hydrological analysis in combination with the gully strike and confluence path of small watershed, to generate a compound ground object unit that can completely contain the potential landslide occurrence area and potential debris flow circulation area in small watershed, and then each compound ground object unit is abstracted as a geographic node with geographic spatial attributes; Step 200: Collecting multi-source heterogeneous environmental factors related to landslide-debris flow disaster chain in small mountainous watershed, taking these environmental factors as the initial input of node features corresponding to the geographic node, and then using mutual information method to calculate the correlation degree between each initial input factor, verifying the independence of different environmental factors in small watershed scenario, and selecting the factors more targeted to the disaster chain of the region; Step 300: According to the environmental factors selected in step 200, the disaster influence relationship between each geographic node in small mountainous watershed is analyzed due to factors such as terrain and source, and the weight of the influence relationship between each geographic node is determined in combination with the geographical environmental similarity of different regions in small watershed, and then a directed weighted graph reflecting the spatial transmission relationship of landslide-debris flow disaster chain in small watershed is constructed; Step 400: According to the characteristics of the number of geographic nodes and the correlation density between nodes in small mountainous watershed, an improved GraphSAGE model is designed, which introduces a neighbor weight mechanism based on the disaster influence degree of small watershed in the model, retains the feature information and historical representation information of each geographic node, and cancels random sampling to ensure the information capture of key associated nodes in small watershed, and the improved GraphSAGE model is used for efficient transmission and node representation optimization of the information in the directed weighted graph; Step 500: Collect historical landslide-debris flow disaster chain data in the mountainous small watershed, construct a disaster chain sample set adapted to the region, train the improved GraphSAGE model using the sample set, adjust the model parameters during training in combination with the disaster occurrence regularity of the small watershed, and then test the prediction effect of the model through independent verification data. After passing the test, the model is used to complete the risk prediction and spatial mapping of landslide-debris flow disaster chains in the mountainous small watershed.

[0022] Further, the present embodiment is directed to the characteristics of small watershed terrain, which is narrow and has significant spatial differences in slope gradient, i.e., the gentle slope section and steep slope section are distributed alternately, the steep slope section is prone to landslides, and the gentle slope section is prone to accumulation of loose materials, etc. The specific operations of obtaining slope units through compound curvature field watershed segmentation, introducing a gully buffer zone to form a compound ground feature unit, and quantifying the characteristics of the source area and flow area are optimized, and the scheme content is as follows: Obtaining slope units through compound curvature field watershed segmentation 1. Composition of compound curvature field The compound curvature field is a terrain feature field that combines planar curvature and profile curvature, which is used to accurately identify the boundary profile and internal uniformity of the slope in the small watershed: Planar curvature: reflects the bending shape of the small watershed slope in the horizontal direction, the positive value corresponds to the ridge line, which is the natural boundary of the slope unit, can block the transport of the source across the unit, and the negative value corresponds to the valley line, which is the transition zone between the slope unit and the gully, and is the key path for the source to enter the gully; Profile curvature: reflects the bending shape of the small watershed slope in the vertical slope direction, the negative value corresponds to the convex slope, which is the transition zone from the slope section to the steep slope section in the small watershed, and the loose material is prone to instability and sliding, which is a potential landslide area, and the positive value corresponds to the concave slope, which is the bottom of the gentle slope section or steep slope section in the small watershed, and is prone to accumulation of loose materials, which is a source reserve area; Calculation of compound curvature field combined with the slope gradient characteristics of the small watershed: According to the slope gradient classification of the small watershed, where the gentle slope section has a slope < 15°, the medium slope section has a slope of 15° ≤ slope ≤ 35°, and the steep slope section has a slope > 35°, different weights are assigned to the planar curvature and profile curvature of different slope sections.

[0023] The profile curvature weight of the medium slope section and the steep slope section is increased to 0.5, i.e., the planar curvature weight is 0.5. The planar curvature weight of the gentle slope section (source accumulation area) is increased to 0.6, i.e., the profile curvature weight is 0.4, to ensure that the segmentation logic is preferentially adapted to the topographic features of the potential landslide area.

[0024] 2. Steps for obtaining slope units based on compound curvature field watershed segmentation (1) Preprocessing of small watershed DEM data To obtain high-precision digital elevation model data for a small watershed, depression filling is first performed to eliminate local depressions in the DEM, thus avoiding distortion in subsequent slope and curvature calculations. Then, based on the DEM, slope data for the entire small watershed is extracted, and slope is graded into gentle, medium, and steep sections to form a slope grading map of the small watershed.

[0025] (2) Calculation of composite curvature field Based on the preprocessed DEM and slope grading map, terrain analysis tools are used to calculate the planar curvature and profile curvature of each grid cell: For grids in medium and steep slope sections: Composite curvature value = 0.5 × absolute value of planar curvature + 0.5 × absolute value of profile curvature; For the grid on a gentle slope: Composite curvature value = 0.6 × absolute value of planar curvature + 0.4 × absolute value of profile curvature; Through the above calculations, a grid map of the composite curvature field of the small watershed is obtained. In the grid map, the high curvature value region with a curvature value greater than 2.0 corresponds to the slope boundary, and the low curvature value region with a curvature value less than 0.8 corresponds to the homogeneous region inside the slope.

[0026] (3) Watershed segmentation and slope unit generation Step 1: Generate a composite curvature field gradient map, which involves calculating the gradient of the composite curvature field using spatial analysis tools. Regions with high gradient values ​​correspond to the slope boundary, while regions with low gradient values ​​correspond to the interior of the slope. Step 2: Screening and segmenting seed points. Seed points are only screened in the low gradient value area (composite curvature value < 0.8) of the slope and steep slope in the small watershed. Areas with no landslide potential in the gentle slope are excluded to ensure that the seed points correspond to the core area of ​​the slope of the potential landslide area. Step 3: Perform watershed segmentation, using the gradient map as the segmentation boundary and seed points as the core of the unit, and complete the segmentation using terrain analysis tools to obtain the initial slope unit raster map; Step 4: Post-processing of the units. Initial slope units with an area of ​​<0.02km² within the small watershed are merged to avoid computational redundancy caused by overly fine units. Initial slope units with a slope of <10° are deleted. Finally, a slope unit vector map that adapts to the slope characteristics of the small watershed is obtained.

[0027] Regarding the definition of the channel buffer, data acquisition and generation logic 1. Definition of channel buffer zone The buffer zone of a gully within a small watershed refers to an area of ​​a certain width extending to the slopes on both sides, centered on the natural gullies (including seasonal gullies) within the small watershed. This area is the transition zone for landslide material to enter the gully within the slope unit, and is also the outer area of ​​the core channel for debris flow. Its width needs to be dynamically adjusted in conjunction with the slope gradient of the small watershed to avoid insufficient regional adaptability caused by a fixed width.

[0028] 2. Data acquisition of channel buffer zone Basic data sources: small watershed 1:10,000 DEM data, high-resolution remote sensing image, small watershed field hydrological survey data, which contains channel trend, perennial flow / seasonal flow attributes, etc. Channel extraction: based on DEM data, the initial river network is calculated by hydrological analysis tool with flow accumulation greater than 2000, and then the high-resolution remote sensing image and field survey data are superimposed to correct the initial river network, that is, to delete artificial drainage channels and supplement seasonal gullies not recognized by DEM, and then the actual channel vector data of the small watershed is obtained. Buffer zone width determination: combined with small watershed slope gradient classification, the width is set, the channel buffer zone width of steep slope section with slope greater than 35° is 10-20m, the material source transport distance of steep slope section is short, and the buffer zone does not need to be too wide; the medium slope section with slope between 15° and 35° is 20-30m, and the gentle slope section with slope less than 15° is 30-40m, the material source transport distance of gentle slope section is long, and the buffer zone coverage range needs to be expanded.

[0029] 3. Generation of channel buffer zone Based on the corrected small watershed channel vector data, the channel buffer zone vector map is generated by the buffer analysis function of the spatial analysis tool according to the above dynamic width, and then the small watershed boundary vector is superimposed to cut off the buffer zone part beyond the small watershed range, and finally the small watershed channel buffer zone data is obtained.

[0030] III. Quantitative description of characteristics of material source area and flow area 1. Quantification of material source area characteristics Taking slope unit as the basic unit, combined with the slope gradient characteristics of small watershed, the material source quantity potential and material source stability are quantified from two dimensions: Material source quantity potential: through slope unit grading assignment, steep slope section is assigned 3, medium slope section is assigned 2, and gentle slope section is assigned 1. Combined with the relative thickness of loose deposits in the unit, that is, based on the vegetation coverage degree of remote sensing image: unit with vegetation coverage degree less than 30% is assigned 3, 30%-60% is assigned 2, and more than 60% is assigned 1, and the product of the two is taken as the material source quantity potential index. The higher the index, the more landslide material source the unit can provide; Material source stability: through the profile curvature type assignment of slope unit, convex slope is assigned 3, straight slope is assigned 2, and concave slope is assigned 1. The more negative the curvature value, the worse the material source stability, and the more prone to landslides and material source replenishment.

[0031] 2. Quantification of flow area characteristics Taking the channel buffer zone as the basic unit, the material source transport capacity and the smoothness of flow channel are quantified from two dimensions: Transportation capacity of the source: the accumulation of the corresponding confluence of the channel buffer zone is classified, and the accumulation of the confluence is > 5000, 3000-5000, and < 3000, which is assigned a value of 3, 2, and 1 respectively. The higher the accumulation of the confluence, the stronger the ability of the water flow to transport the source. Smoothness of the flow channel: the longitudinal slope of the channel buffer zone corresponding to the channel is classified, i.e. the longitudinal slope is > 5 ‰, 2 ‰-5 ‰, and < 2 ‰, which is assigned a value of 3, 2, and 1 respectively. The larger the longitudinal slope, the steeper the channel slope, and the smoother the flow of the debris flow, which is not easy to accumulate.

[0032] Four, the final formation of the composite land unit The slope unit with the slope gradient of the above-mentioned obtained small watershed is spatially superimposed and analyzed with the channel buffer zone with dynamic width. The area where the slope unit and the channel buffer zone overlap and the overlapping area accounts for ≥30% of the area of the slope unit is retained. The slope unit and the channel buffer zone in the overlapping area are merged to form a composite land unit containing the potential source area of the landslide (slope unit) and the flow passage area of the debris flow (channel buffer zone). Each composite land unit is attached with the above-mentioned quantified source area characteristic index and flow passage area characteristic index.

[0033] Further, the node characteristics constructed in the above step 200 and the independence of the environmental factors are verified, optimized and deepened. Through specific mutual information calculation, independence quantification and redundancy elimination logic, the problem of low model training efficiency and prediction accuracy interference caused by redundant multi-source factors is solved, and the composite land unit is connected.

[0034] The embodiment further clarifies the whole process of screening and optimization in view of the characteristics of small watershed multi-source environmental factor type and strong correlation between some factors. The specific steps are as follows: Step 201: Around the formation mechanism of the landslide-debris flow disaster chain in the small watershed, collect multi-source heterogeneous environmental factors corresponding to the composite land unit, and classify them according to the influence dimension: Topographic factors: average slope, relative height difference, profile curvature, and plane curvature of the composite land unit; Geological factors: rock-soil type, fault structure density, and weathering layer thickness level in the unit; Hydrological factors: multi-year average precipitation, confluence accumulation intensity, and groundwater depth level in the unit; Vegetation factors: vegetation coverage and vegetation type (tree / shrub / herb) in the unit; Human activity factors: road density and engineering activity disturbance range proportion in the unit.

[0035] Factor preprocessing: convert different dimension factors to the same numerical interval to eliminate the interference of dimension difference on mutual information calculation; Outlier processing: remove extreme values caused by data acquisition errors, and keep effective data that conforms to the terrain and geological regularities according to the distribution characteristics of the same factor in small watersheds. Factor gridding: convert all environmental factors into raster data with the same spatial resolution as the composite ground feature units, ensuring that each geographic node can accurately match the corresponding factor value.

[0036] Step 202: Calculate the correlation degree between any two environmental factors using normalized mutual information. Normalized mutual information quantifies the correlation degree as a value between 0 and 1, avoiding the distortion of original mutual information caused by different factor value ranges.

[0037] Mutual information matrix construction: take all collected environmental factors as matrix rows and columns, calculate the normalized mutual information value between each pair of factors, such as average slope and relative height difference, vegetation coverage and weathering layer thickness grade, and form a factor-factor mutual information matrix. Matrix element definition: let the matrix be M n×n , n is the total number of factors, and element M ij represents the normalized mutual information value between the i th factor and the j th factor. M ij =M ji , the diagonal element M ii =1 .

[0038] Step 203: Calculate the comprehensive independence index for each environmental factor to quantify its independence degree with all other factors. The formula logic is as follows: Average correlation degree of a single factor: calculate the mean value of all non-diagonal elements in the corresponding mutual information matrix row (or column) of the factor, denoted as NMI i , NMI i The larger the value, the stronger the overall correlation between the factor and other factors, and the weaker the independence. Comprehensive independence index: I i =1-NMI i I i The value range is 0-1, I i The closer to 1, the stronger the comprehensive independence of the factor; the closer to 0, the higher the redundancy of the factor.

[0039] Independence classification: According to the characteristics of the small watershed disaster chain influencing factors, the comprehensive independence index is divided into three levels I i : High independence: I i ≥0.7, the factor is weakly related to other factors, and the information is unique; Medium independence: 0.4≤ I i <0.7, the factor has weak association with some factors, but still has effective information); Low independence: I i <0.4, the factor has strong association with other multiple factors, and the redundant information accounts for a high proportion.

[0040] Step 204: Combined with mutual information matrix and comprehensive independence index, double determination of strong dependent factors: Rule 1: If the normalized mutual information value between two factors is M ij ≥ preset adaptive threshold, which is calibrated based on small watershed historical disaster data, to ensure that truly redundant factors are removed, then the two are determined as a strong dependent factor pair; Rule 2: In the strong dependent factor pair, the factor with smaller I i is preferentially removed, and the factor with stronger independence and more unique information is retained; Rule 3: If a single factor has M ij ≥ preset adaptive threshold, no matter its I i size, it is directly determined as a strong dependent factor and removed, and the factor has too much redundant information and limited model gain.

[0041] Iterative removal process: First round of removal: identify and remove strong dependent factors according to the above rules to obtain a preliminary screened factor set; Second round of verification: rebuild the mutual information matrix for the preliminary screened factor set, calculate the comprehensive independence index of each factor, and if there are still strong dependent factor pairs, repeat the removal operation; Termination condition: in the screened factor set, the M ij < preset adaptive threshold, and the comprehensive independence index of all factors I i ≥0.4, stop iteration.

[0042] Step 205: The adaptability verification of the screened factor set is matched with the complex ground object unit (geographic node) generated in step 100 to ensure that each geographic node can correspond to complete optimization factor data; at the same time, in combination with the key influence mechanism of the small watershed landslide-mudflow disaster chain, it is verified whether the screened factor set covers the core dimension of the disaster chain formation to avoid missing key information due to excessive elimination.

[0043] Through the above steps, a node feature set without redundancy, high independence and strong pertinence is finally constructed for each geographic node.

[0044] Further, for step 300, the embodiment provides further preferred limitations, and the specific optimization is as follows: Step 301: Based on the function attributes of the slope unit of the material source area and the channel buffer area of the flow area of the complex ground object unit, in combination with the conduction path of the landslide-mudflow disaster chain, the directed association premise between geographic nodes is determined to avoid meaningless spatial association.

[0045] The association direction rule is that the material source area node points to the spatial adjacent flow area node to reflect the conduction direction of material source supply to mudflow flow; the flow area node points to the downstream flow area node to reflect the conduction direction of flow continuation; the flow area node is not allowed to point to the material source area node, and the downstream flow area node is not allowed to point to the upstream flow area node, which conforms to the irreversible conduction characteristics of the disaster chain.

[0046] The spatial adjacency determination is based on the vector boundary of the complex ground object unit. If the boundaries of two geographic nodes overlap or the distance does not exceed the preset threshold value based on the small watershed channel density calibration, such as 500 meters, it is determined to be spatially adjacent and has the basis for association.

[0047] Step 302: Based on the material source area factor subset, the flow area factor subset and the common factor subset screened in step 200, the disaster influence intensity of each node on the associated node is quantified, and the weight size directly reflects the contribution degree of the factor to the conduction of the disaster chain.

[0048] The factor weight distribution follows a professional priority orientation. The influence weight of the source area node is based on the distribution proportion of loose deposits, average slope, and rock-soil shear strength as core factors, with weight proportions of 30%, 25%, and 20% respectively. Vegetation coverage and fault structure density are auxiliary factors, with weight proportions of 15% and 10% respectively. The higher the factor value, such as a high proportion of loose deposits or a steep slope, the greater the influence weight of the source area node on the downstream flow area node. The influence weight of the flow area node is based on the channel longitudinal gradient and the convergence cumulative intensity as core factors, with weight proportions of 35% and 30% respectively. Channel width and rainstorm intensity level are auxiliary factors, with weight proportions of 20% and 15% respectively. The higher the factor value, such as a large longitudinal gradient or strong convergence, the greater the influence weight of the flow area node on the downstream flow area node.

[0049] In the calculation of the comprehensive influence weight of the node, each factor value in the optimized factor set of each geographic node is standardized in the range of 0-1 to eliminate dimensional differences. Then, the weighted sum is calculated according to the above weight distribution to obtain the comprehensive influence weight of the node, which is in the range of 0-1 and serves as a core component of the subsequent edge weight.

[0050] Step 303: Spatial similarity quantification focuses on the similarity of disaster chain transmission potential rather than simply geographical distance. It combines the environmental attributes of composite ground units with the similarity calculation logic defined by disaster chain links. The similarity evaluation index selects common factors strongly related to disaster chains, namely fault structure density and multi-year average precipitation, as well as exclusive factors for corresponding functional areas. For source area nodes, rock-soil types are added, and for flow area nodes, riverbed material composition is added.

[0051] The similarity calculation method is to calculate the Euclidean distance of two associated nodes, such as source area A and flow area B, and flow area B and flow area C, in each similarity index. Then, weights are assigned according to the importance of the index to the disaster chain, with structure density accounting for 30%, precipitation accounting for 30%, and exclusive factors accounting for 40%. The weighted sum of the index difference is obtained. The spatial similarity coefficient is 1 minus the normalized index difference, with a value range of 0-1. The closer the coefficient is to 1, the more similar the disaster chain influence conditions of the two nodes, and the stronger the correlation of the transmission potential.

[0052] At the same time, the spatial similarity coefficient is distance decay corrected. If the distance between associated nodes is less than 500 meters, the correction coefficient is 1, i.e. no decay. If the distance exceeds 500 meters, the correction coefficient decreases with distance gradient, such as 0.8 for 500-1000 meters and 0.6 for 1000-1500 meters, ensuring that the influence of spatial distance on transmission efficiency conforms to reality, i.e. the transmission potential decreases with distance.

[0053] Step 304: Based on the results of the above steps, a complete directed weighted graph is constructed, and each element of the graph corresponds to the professional logic of the disaster chain.

[0054] The node definition abstracts the geographical node after step 100 as the vertex of the graph, and each vertex is attached with node type, source area or flow area, comprehensive influence weight, optimization factor set data, spatial coordinates and other attributes, and the function positioning of the node in the disaster chain is clear.

[0055] The direction of the edge follows the association direction rule of step 301, and only the directed edges from the source area to the flow area and from the upstream flow area to the downstream flow area are drawn, and the arrow direction of the edge directly corresponds to the conduction path of the disaster chain.

[0056] The edge weight is the product of the node comprehensive influence weight and the modified spatial similarity coefficient, and the calculation formula is edge weight W ij The geometric mean of the comprehensive influence weight of node i and the comprehensive influence weight of node j is multiplied by the modified spatial similarity coefficient of node i and j. The geometric mean of the comprehensive influence weight of two nodes is taken to ensure the balance of two-way influence, and the similarity coefficient is multiplied to reflect the enhancement effect of similar environmental conditions on the conduction intensity.

[0057] When simplifying the graph structure, the directed edges with edge weight less than 0.1 are removed, and the core associated edges with weight not less than 0.1 are retained to avoid redundancy of the graph structure.

[0058] Step 305: Verify whether the graph structure fits the actual spatial association of the landslide-mudflow disaster chain to ensure no logical deviation. When verifying the conduction path, superimpose the historical landslide-mudflow disaster path graph of the small watershed to check whether the directed edges from the source area to the flow area and then to the downstream flow area in the graph are consistent with the actual disaster spread path. If there is deviation, such as a historical disaster path not reflected in the graph, adjust the spatial threshold or similarity index of node association.

[0059] When verifying the weight rationality, compare the edge weight distribution of high-risk and low-risk disaster areas to ensure that the node associated edge weight in the high-risk disaster area is significantly higher than that in the low-risk area, such as the average edge weight in the high-risk area not less than 0.4 and the low-risk area not more than 0.2, which reflects the positive correlation between weight and disaster risk. When verifying the function area association, check whether the source area node is associated with at least one flow area node to ensure that the source can flow into the flow area, and whether the flow area nodes form a continuous association chain from upstream to downstream to ensure that the debris flow can flow completely, avoiding isolated nodes without associated edges. If there are isolated nodes, re-determine their spatial adjacency.

[0060] Through the above refinement steps, the directed and weighted graph constructed not only has the characteristics of mathematical graph structure, but also deeply integrates the functional attributes of complex ground object units and the conduction rules of landslide-mudflow disaster chain, providing a structured input that fits the professional scenario for improving the information transmission of the GraphSAGE model in step 400.

[0061] Further, with reference to Figure 1 and Figure 2 , the following preferred solutions are proposed for step 400.

[0062] Step 401: Explicitly improve the input adaptation logic of the GraphSAGE model, taking the directed weighted graph constructed in step 300 as the core input carrier, and taking the optimized factor set of the geographic node as the initial feature vector of the model. At the same time, import the node type source area or flow area attribute, node comprehensive influence weight and edge weight data, ensure the depth matching of the model input and the spatial correlation characteristics of the landslide-debris flow disaster chain, the node function attribute, and provide data basis for subsequent targeted improvement.

[0063] Step 402: Cancel the original random sampling mechanism of the GraphSAGE model, and adopt a deterministic neighbor selection strategy based on the conduction intensity of the disaster chain.

[0064] For each target node in the directed weighted graph, first filter the effective neighbor nodes according to the directionality of the edge, and only retain the downstream flow area neighbors of the source area nodes and the downstream flow area neighbors of the flow area nodes, excluding the reverse nodes that have no disaster conduction significance.

[0065] Then sort the effective neighbor nodes according to the edge weight from large to small, and dynamically determine the number of neighbor nodes according to the node type, select the top N neighbors for the source area nodes, and select the top M neighbors for the flow area nodes, where the values of N and M are based on the calibration of the small watershed disaster chain influence range, to ensure that the selected neighbors are all core related nodes that significantly contribute to the disaster risk conduction of the target node, and to avoid missing key conduction information or interference from irrelevant nodes caused by random sampling.

[0066] Step 403: Introduce a neighbor weight mechanism that fits the disaster chain influence mechanism, and convert the edge weight calculated in step 300 into the aggregated weight of the neighbor nodes. For each neighbor node of the target node, its weight value directly adopts the weight of the corresponding edge, and the larger the edge weight, the stronger the disaster conduction influence of the neighbor node on the target node, and the higher the contribution ratio in information aggregation.

[0067] At the same time, adjust the weight distribution combined with the neighbor node type, if the neighbor is a source area node, superimpose a correction coefficient corresponding to the loose deposit distribution proportion on the basis of the weight, to highlight the enhancement effect of the source supply capacity on the weight; if the neighbor is a flow area node, superimpose a correction coefficient corresponding to the accumulation intensity, to strengthen the influence of the flow conduction capacity on the weight. In the aggregation process, a weighted average aggregation function is used to weight and sum the feature vectors of each neighbor node according to the corresponding weight, to obtain the neighbor aggregated features, and to ensure that the aggregation result accurately reflects the differentiated disaster conduction contribution of different neighbors to the target node.

[0068] Step 404: Retain the self-loop information of the model and optimize its integration logic, highlighting the fundamental impact of the geographical node's own properties on the disaster chain. Add a self-loop edge to each geographical node, and the self-loop weight is determined according to the node type and the comprehensive impact weight. The self-loop weight of the source area node is set to 1.2 times its comprehensive impact weight, highlighting the core role of the source area's own stability in landslide initiation. The self-loop weight of the flow-through area node is set to 1.1 times its comprehensive impact weight, emphasizing the fundamental support of the flow-through area's own topographic and hydrological conditions on debris flow transport. During information aggregation, multiply the target node's own initial feature vector by the self-loop weight to obtain the self-loop feature vector, and process it in parallel with the neighbor aggregated feature vector to avoid representation bias caused by the model ignoring the node's own properties.

[0069] Step 405: Introduce residual information connection mechanism to solve the problem of information attenuation in deep training, while adhering to the continuity feature of disaster chain transmission. After each layer of aggregation and feature conversion in the model, construct a residual connection path to add the output feature vector of the current layer to the input feature vector of the previous layer element-wise.

[0070] To adapt to the information retention needs of different links of the disaster chain, set the residual weight coefficient, which is higher for source area nodes than for flow-through area nodes, ensuring that the initial topographic and geological features of source area nodes and historical transmission information of flow-through area nodes are not weakened in deep networks. Through residual connection, node representation contains both current layer aggregated neighbor information and historical level key features, achieving cross-layer continuous transmission of disaster chain information.

[0071] Step 406: Design a multi-stage aggregation and feature conversion structure to gradually optimize node representation effects.

[0072] The model is set to K layers of aggregation structure, with K value determined according to the transmission level of the small watershed disaster chain, ensuring coverage of the complete chain from source initiation to flow-through spread.

[0073] The first layer of aggregation only integrates the target node's direct neighbor information and self-loop information, outputting a first-order representation vector. Each subsequent layer of aggregation is based on the current layer representation vector, combined with the corresponding layer representation vector of the neighbor nodes, repeating the neighbor weight aggregation, self-loop information integration, and residual connection process to achieve progressive integration of multi-stage disaster transmission information. After each layer of aggregation, a linear conversion layer and an activation function are connected to perform dimension conversion and non-linear mapping of aggregated features, enhancing the model's ability to capture complex correlation features of the disaster chain while avoiding computational redundancy caused by feature dimension expansion.

[0074] Step 407: Perform the full flow operation of information transmission and node characterization optimization, start model calculation according to the disaster conduction order of the directed weighted graph, first perform feature aggregation and characterization update on all source area nodes, then process upstream flow area nodes and downstream flow area nodes in turn, to ensure that the information transmission direction is consistent with the actual disaster chain transmission path.

[0075] After the operation of each layer is completed, the output characterization vector is standardized to ensure that the characterization vectors of each node are in the same numerical interval, eliminating the influence of dimension difference on subsequent calculation. After K-layer aggregation and optimization, the final characterization vector of each geographic node is obtained, which integrates its own attributes, multi-order neighbor disaster conduction information, and cross-layer residual information, fully describing the risk characteristics of the node in the landslide-debris flow disaster chain.

[0076] Step 408: Verify the adaptability of the node characterization result, compare the final characterization vectors of disaster nodes and non-disaster nodes to test whether there is a significant distinction between them, and ensure that the characterization result can effectively reflect the disaster risk difference of the nodes; at the same time, analyze the dimension characteristics of the characterization vector combined with the node type, verify whether the characterization vectors of the source area nodes and the flow area nodes present feature differences matching their own functional attributes, if the distinction is insufficient or the feature deviation is large, adjust the number of neighbor selection, the self-loop weight coefficient or the residual weight coefficient, and execute the characterization optimization process again until the result meets the risk assessment requirements of the landslide-debris flow disaster chain, providing high-quality node characterization data for model training and risk prediction in step 500.

[0077] Further, referring to Figure 3 , for the detailed definition of step 500, in order to realize the construction of few-data regional samples and model training based on unit splitting feature matching topography fitting, the following steps are taken: Step 501: Split the complex ground feature units of the existing disaster area, and extract fine-grained unit-level basic samples.

[0078] Multiple regional data of landslide-debris flow disasters that have occurred at home and abroad need to be collected, which can be complete regions or only small range data around the disaster point.

[0079] According to the complex ground feature unit segmentation logic in step 100, each data is split into the smallest granularity of source area units and flow area units, the area of the source area unit is not more than 0.05 square kilometers, and the area of the flow area unit is not more than 0.1 square kilometers. Ensure that each fine-grained unit only contains a single topographic feature and a single disaster chain link, such as a steep slope source unit only corresponding to a landslide initiation link and a narrow ditch flow unit only corresponding to a debris flow transport link. For each fine-grained unit, three types of core information are extracted simultaneously.

[0080] The first type is the topographic feature of the unit itself, including slope, rock-soil type, relative height difference, and channel longitudinal gradient, which is consistent with the optimization factor set in step 200.

[0081] The second type is the correlation coefficient between the unit and adjacent units, which is based on the edge weight calculation logic in step 300, and the correlation strength and direction data between the unit and upstream and downstream units are retained.

[0082] The third type is the disaster label of the unit, indicating whether the unit has experienced a disaster and the link of the disaster chain, such as landslide in the source unit and debris flow in the flow unit. Through the above operations, a unit-level basic sample library containing thousands of fine units is formed.

[0083] Step 502: Establish a unit feature topographic type mapping index to prepare for target area fitting. For each fine unit in the unit-level basic sample library, perform double classification and labeling according to topographic dimensions and disaster chain links.

[0084] The topographic dimensions are divided into 27 topographic types according to slope levels, rock-soil stability, and convergence conditions. The disaster chain links are divided into 4 types according to source initiation, source supply, flow transport, and flow deposition.

[0085] Based on the classification results, a mapping index table is constructed, with each row corresponding to a fine unit and the columns containing topographic type codes, disaster link codes, unit feature vectors, adjacent unit correlation coefficients, and disaster labels. This index table supports quick retrieval and matching of samples according to topographic types and disaster links, solving the unit screening efficiency problem in subsequent target area terrain adaptation.

[0086] Step 503: Analyze the target area terrain partition to determine the target template for unit fitting.

[0087] The target area is divided into topographic zones according to the composite ground units generated in step 100. First, the source area and flow area are split into large units according to the functional division in step 100, and then each large unit is matched with the 27 topographic types in step 502 according to its slope, rock-soil, and convergence three-dimensional characteristics, forming the target area terrain partition template.

[0088] For example, the topographic features of a large unit in a source area of the target area are steep slope greater than 35 degrees, low rock-soil stability, and medium convergence condition, corresponding to a specific topographic type code; the features of a large unit in a flow area are medium slope 15-35 degrees, medium rock-soil stability, and strong convergence condition, corresponding to another topographic type code. At the same time, the correlation between units in the target area is recorded, such as the adjacency of source area A and flow area B, and the upstream and downstream connection of flow area B and flow area C, forming a target area unit correlation topology graph.

[0089] Step 504: Generate target region exclusive sample set by terrain matching and associated adaptive fitting. Based on the target region's terrain zoning template and associated topological graph, filter and combine the fine-grained units from the unit-level basic sample library.

[0090] In the single-unit matching link, for each minimum granularity small unit of the target region split by step 100, retrieve the fine-grained units with the same code from the sample library according to its terrain type code. If the number of completely matched units is not less than 5, randomly select 3 to 5 as candidate units. If there is no completely matched unit, select the units that match the two core features of slope and rock-soil body and have similar confluence conditions, adjust the confluence-related feature values through linear interpolation, such as adjusting the sample unit confluence intensity 0.8 to the target unit required confluence intensity 0.6, to ensure feature adaptation.

[0091] In the associated relationship fitting link, refer to the target region unit associated topological graph, and supplement the correlation coefficients of the adjacent units of the matched candidate units from the sample library. If target unit A is adjacent to B, and the candidate unit of A comes from unit X in the sample library, and the candidate unit of B comes from unit Y in the sample library, calculate the historical correlation coefficient of units X and Y. Then, combined with the actual distance between A and B in the target region, calibrate and correct the correlation coefficient of A and B according to the spatial adjacency threshold of step 301, to ensure that the correlation of the fitted units conforms to the conduction law of the actual terrain of the target region.

[0092] In the combination verification link, combine the fitted units according to the associated topological graph of the target region to form the target region fitting sample set. Check whether the combined sample set covers all terrain types of the target region, such as steep slope source area and gentle slope flow-through area. At the same time, confirm that there is no logical contradiction in the correlation of each unit, such as no reverse correlation of downstream unit pointing to upstream unit. If there is a lack or contradiction, repeat the single-unit matching and associated relationship fitting steps to supplement and adjust.

[0093] Step 505: Split the disaster chain tags of the fitted sample set to enhance data diversity and adaptability. For the target region fitting sample set generated by fitting, according to the unit function and correlation of step 503, split the traditional binary tags of disaster occurrence or non-occurrence into two-dimensional tags of link intensity.

[0094] Link tags label the disaster chain link to which the unit belongs according to the source start, source recharge, flow transport, and flow deposition, such as labeling the fitted source area unit as source start and the flow-through area unit as flow transport.

[0095] The intensity label refers to the disaster intensity of the candidate unit in the sample library, such as landslide scale, debris flow flow, and combines the topographic features of the target area unit, such as steep slope unit intensity is higher than that of gentle slope unit, and labels three intensities of strong, medium and weak. For example, the target area steep slope source unit is labeled as strong source start, and the gentle slope flow unit is labeled as weak flow transport. Through label splitting, the fitted sample set can cover the characteristics of different links and intensities of the disaster chain even if the number of units is limited, and the learning ability of the model to the details of the disaster chain of the target area is improved.

[0096] Step 506: Based on the phased model training of the fitted sample set, the adaptability of the target area is strengthened.

[0097] The first stage carries out unit-level pre-training, taking a single unit in the fitted sample set as the training unit, and inputting the unit feature vector and link label. The improved GraphSAGE model designed in step 400 is used to learn the mapping relationship from the topographic features of the unit to the disaster link, and the feature capture ability of the model for a single unit is optimized. When the unit-level prediction accuracy is stable at more than 75%, the pre-training is stopped.

[0098] The second stage carries out correlation-level fine-tuning, introduces the unit correlation index in the fitted sample set, and constructs a unit correlation subgraph according to the correlation topology graph of the target area. The pre-trained model parameters are used as the initial value, and the model is trained using the correlation subgraph to learn the rules of correlation between units to disaster transmission, such as strong start of the source unit plus high correlation index corresponding to strong transport of the flow unit. The residual connection and neighbor weight mechanism of step 400 are retained during fine-tuning to ensure that the correlation information does not decay. When the F1 score of the correlation-level prediction is stable at more than 70%, the fine-tuning is stopped.

[0099] The third stage carries out regional-level verification, randomly selects 20% of the units from the fitted sample set to form a verification sub-region, simulates the complete topography of the target area, and uses the fine-tuned model to predict the disaster chain risk of the verification sub-region. Compare the difference between the prediction results and the labels of the fitted sample set. If the difference rate exceeds 15%, return to step 504 to adjust the unit fitting logic, such as supplementing candidate units of specific topographic types, until the difference rate does not exceed 15%, to ensure that the model adapts to the overall topographic features of the target area.

[0100] Step 507: Perform risk prediction and spatial mapping of the target area, consistent with the sample fitting logic. Split the composite land unit of the target area according to the terrain zoning template generated in step 100. Input the trained improved GraphSAGE model, and the model outputs the disaster chain risk index of each unit according to the process of single-unit feature extraction, inter-unit correlation calculation, and multi-link risk judgment. The index integrates the comprehensive value of link label and intensity label. Divide the units into four risk levels: very high, high, medium, and low. Generate a risk spatial map based on the unit correlation topology graph of the target area. Label the unit function type source area or flow area and the core disaster-causing factor during mapping, such as the strong activation of the very high-risk steep slope source unit, to ensure that the risk map not only reflects the unit-level risk but also demonstrates the source-flow-conduction correlation of the disaster chain, providing precise spatial guidance for subsequent disaster prevention deployment.

[0101] Step 508: Dynamically update the unit sample library and model parameters to continuously optimize prediction effect. Regularly collect new environmental data in the target area, such as annual vegetation changes and rainfall data, and new disaster unit data in the surrounding area. Split the new disaster units into fine units according to step 501 and supplement them to the unit-level basic sample library. At the same time, repeat steps 503 to 506 based on the new data to fine-tune the fitting sample set and model parameters of the target area, ensuring that the model can continuously adapt to environmental changes in the target area and maintain the accuracy of risk prediction in the long term.

[0102] This embodiment provides an application example: landslide-debris flow disaster chain risk assessment in a small watershed in a mountainous area Select a small watershed in a mountainous area in southwest China as the evaluation area. The area is about 80 square kilometers, with mainly steep slopes, dense gullies, and scarce historical disaster data, with only 3 small-scale landslide-debris flow events recorded, which meets the evaluation scenario of few data regions.

[0103] Based on the technical process of this application, the watershed is first split into 1260 composite land units according to step 100, 12 core environmental factors are selected through step 200, a directed weighted graph is constructed in step 300, and the improved GraphSAGE model in step 400 is trained. The sample set is constructed according to the unit splitting-feature matching-terrain fitting mechanism in step 500, and the fine unit data of 5 different terrain regions are fused to generate the fitting.

[0104] At the same time, refer to Figures 4-7 , use traditional machine learning models MLP, RF, and SVM as a comparison, and conduct parallel testing under the same sample set and evaluation indicators. The performance indicators are compared as follows: Accuracy: 0.895 for the method of this application, 0.823 for RF, 0.801 for SVM, and 0.786 for MLP; Precision: 0.858 for the method of the present application, 0.812 for RF, 0.793 for SVM, 0.765 for MLP; Recall: 0.885 for the method of the present application, 0.784 for RF, 0.752 for SVM, 0.731 for MLP; F1 score: 0.820 for the method of the present application, 0.728 for RF, 0.705 for SVM, 0.689 for MLP; AUC value: 0.854 for the method of the present application, 0.796 for RF, 0.778 for SVM, 0.753 for MLP; Kappa coefficient: 0.816 for the method of the present application, 0.735 for RF, 0.712 for SVM, 0.690 for MLP.

[0105] The actual evaluation results show that a total of 42 high-risk units are identified in the small watershed. Among them, the method of the present application only misses 2 high-risk units, while RF misses 7, SVM misses 9, and MLP misses 11. The missed high-risk units are mostly in the complex terrain of steep slope source area and the junction of the channel buffer zone. The method of the present application effectively reduces the risk of missing judgment in such areas by strengthening the capture of unit correlation characteristics and fitting the disaster chain links, providing more accurate spatial guidance for the deployment of disaster prevention and monitoring in the watershed.

[0106] The present application is not limited to the above-mentioned optional embodiments, and anyone can derive other various forms of products under the inspiration of the present application. The above specific embodiments should not be understood as limiting the protection scope of the present application, and the protection scope of the present application should be defined by the claims, and the specification can be used to explain the claims.

Claims

1. A method for assessing landslide-debris flow hazard chains based on graph neural networks, characterized in that, Includes the following steps: Step 100: Generate composite land feature units through composite curvature field watershed segmentation and hydrological analysis, and abstract the composite land feature units as geographic nodes; Step 200: Construct the node features corresponding to the geographic node, and use the mutual information method to verify the independence of each environmental factor in the node features; Step 300: Construct a directed weighted graph based on the environmental factors and spatial similarity. The directed weighted graph is used to reflect the spatial transmission relationship of the landslide-debris flow disaster chain. Step 400: Design an improved GraphSAGE model, and use the improved GraphSAGE model to transfer information and optimize node representation in the directed weighted graph; Step 500: Train and validate the improved GraphSAGE model to complete the risk prediction and spatial mapping of the landslide-debris flow hazard chain.

2. The method for assessing landslide-debris flow hazard chains based on graph neural networks according to claim 1, characterized in that, The process of generating composite land feature units in step 100 includes: firstly, obtaining slope units by dividing the watershed through composite curvature field, and then introducing gully buffer zones on the basis of the slope units to form the composite land feature units.

3. The method for assessing landslide-debris flow hazard chains based on graph neural networks according to claim 2, characterized in that, After forming the composite land cover unit in step 100, the method further includes quantifying the source area characteristics and circulation area characteristics corresponding to the composite land cover unit.

4. The method for assessing landslide-debris flow hazard chains based on graph neural networks according to claim 1, characterized in that, The process of constructing node features in step 200 includes: collecting multi-source heterogeneous environmental factors and using the multi-source heterogeneous environmental factors as initial input factors for the node features; the verification of the independence of each environmental factor using the mutual information method specifically involves using the normalized mutual information method to calculate the degree of correlation between each initial input factor.

5. The method for assessing landslide-debris flow hazard chains based on graph neural networks according to claim 4, characterized in that, After calculating the correlation degree between each initial input factor in step 200, the method further includes removing strongly dependent factors based on the correlation degree. The strongly dependent factors are initial input factors whose correlation degree exceeds a preset threshold.

6. The method for assessing landslide-debris flow hazard chains based on graph neural networks according to claim 1, characterized in that, The process of designing the improved GraphSAGE model in step 400 includes: introducing a neighbor weighting mechanism to the GraphSAGE model, preserving self-loop information, preserving residual information, and canceling random sampling operations.

7. The method for assessing landslide-debris flow hazard chains based on graph neural networks according to claim 6, characterized in that, The introduced neighbor weighting mechanism specifically involves assigning differentiated weights to different neighbor nodes based on the degree of influence of each geographical node's neighbor nodes on that geographical node.

8. The method for assessing landslide-debris flow hazard chains based on graph neural networks according to claim 6, characterized in that, The retention of self-loop information specifically means that when the improved GraphSAGE model transmits information, it retains the feature information of each geographic node and does not exclude the information contribution of the geographic node to itself; the retention of residual information specifically means that when the improved GraphSAGE model optimizes node representation, it performs residual connection between the node representation results of the previous round and the node representation results of the current round to retain historical representation information.

9. The method for assessing landslide-debris flow hazard chains based on graph neural networks according to claim 1, characterized in that, Before performing information transfer and node representation optimization on the improved GraphSAGE model in step 400, the method also includes constructing a disaster chain sample set. The construction of the disaster chain sample set adopts a hierarchical sampling strategy, and positive and negative samples are counted according to environmental units to ensure a balanced sample distribution.

10. The method for assessing landslide-debris flow hazard chains based on graph neural networks according to claim 9, characterized in that, The process of constructing the disaster chain sample set also includes splitting the disaster chain labels, which enhances the data diversity of the disaster chain sample set.

Images