Landslide hidden danger point identification method based on aerospace remote sensing fusion neural network

Abstract

The invention relates to the technical field of landslide geological disaster hidden danger point identification, in particular to a landslide hidden danger point identification method based on an air-space remote sensing fusion neural network, which comprises the following steps: acquiring satellite SAR data, a digital elevation model, a vector boundary map and a precision orbit ephemeris of a research area; carrying out interference processing on the InSAR data, and carrying out deformation rate inversion, terrain residual error correction and atmospheric delay phase modeling and correction; preprocessing an original high-resolution remote sensing image, interpreting a high-resolution image, and outputting spatial distribution of suspected landslide hidden danger points in a research area; planning an aerial survey route of the unmanned aerial vehicle, and performing distortion correction, feature matching, dense point cloud generation and orthoimage generation processing; performing multi-dimensional feature analysis, and performing comprehensive interpretation in combination with an expert knowledge base; according to the method, the key effect of the multi-source remote sensing fusion technology in landslide hidden danger monitoring is displayed from data acquisition to hidden danger recognition.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of landslide geological disaster hidden point identification, and particularly relates to a landslide hidden point identification method based on space remote sensing fusion neural network. BACKGROUND

[0002] Global climate change intensifies, extreme weather events occur frequently, and human activities expand, leading to an increasing frequency and scale of landslides, which seriously threaten human life and property safety and infrastructure. China's mountainous and hilly areas account for about 65% of the country's land area, with complex geological conditions and frequent tectonic activities. Landslides, landslides, mudslides and other sudden geological disasters are widespread and difficult to prevent, making China one of the countries with the most severe geological disasters and the most threatened population in the world. Therefore, accurately identifying landslide hidden points and taking effective prevention and control measures in advance are crucial to reducing disaster losses. With the rapid development of remote sensing technology, interferometric synthetic aperture radar (InSAR), high-resolution optical remote sensing, deep learning analysis of remote sensing images, and unmanned aerial vehicle oblique photography have brought new opportunities for landslide hidden point identification.

[0003] Small Baseline Subset Interferometric Synthetic Aperture Radar (SBAS-InSAR) is a time series deformation inversion method based on statistical characteristics of differential interferometric phase. Its core theory is to divide long time series SAR images into multiple small baseline sets that meet the spatial baseline threshold through baseline optimization criteria, and to construct a networked interference combination model based on the least squares principle. However, the time series deformation results of spaceborne InSAR are often restricted by geometric distortion, space-time decorrelation, long revisit period, and atmosphere, making the results have certain limitations.

[0004] High-resolution remote sensing technology, with its high spatial resolution, can improve the accuracy of landslide hidden point identification by combining with geological maps and topographic map data for comprehensive analysis in areas with sparse vegetation cover, large terrain slope, and land use changes. Although high-resolution remote sensing can identify landslide hidden points with high accuracy, it requires a large amount of time for manual visual interpretation of the entire study area, and cannot reflect deep information and surface change patterns, making it difficult to identify potential landslide hidden points.

[0005] As a new remote sensing method, the unmanned aerial vehicle oblique photography technology has the characteristics of flexible operation, efficient identification, low cost and high resolution in local landslide monitoring and risk assessment. The unmanned aerial vehicle can fly at low altitude and use high-resolution cameras to monitor key areas in detail, and can obtain centimeter-level resolution images to accurately reflect the local terrain and surface features. Although the unmanned aerial vehicle photogrammetry has high accuracy, it needs manual field operation and is difficult to achieve high-frequency measurement, and is poor in identifying landslide hazards formed in a short period. SUMMARY

[0006] The purpose of the present application is to provide a landslide hazard point identification method based on space remote sensing fusion neural network, which aims to solve the problem that the existing landslide hazard point identification technology is difficult to achieve high-frequency measurement and is poor in identifying landslide hazards formed in a short period.

[0007] To achieve the above purpose, the present application provides a landslide hazard point identification method based on space remote sensing fusion neural network, comprising the following steps:

[0008] Obtain S+1 scene Sentinel-1A satellite SAR data, digital elevation model, vector boundary map and precise orbit ephemeris of the study area;

[0009] Interference processing is performed on the InSAR data, and deformation rate inversion, terrain residual correction, and atmospheric delay phase modeling and correction are performed to obtain the deformation rate results of the study area and mark the key deformation area;

[0010] The original high-resolution remote sensing image is preprocessed, the improved DeepLabV3+ neural network is used to interpret the high-resolution image, and the spatial distribution of suspected landslide hazard points in the study area is output;

[0011] The spatial distribution of the landslide hazard points is planned for unmanned aerial vehicle survey route, high-precision images are obtained, and distortion correction, feature matching, dense point cloud generation and orthographic image generation processing are performed;

[0012] Fusion SAR deformation data, high-resolution remote sensing interpretation results and unmanned aerial vehicle survey data are analyzed for multi-dimensional feature analysis, and expert knowledge base is combined for comprehensive interpretation;

[0013] Verify and confirm the spatial position of the landslide hazard points, and construct a demonstration point oblique photography three-dimensional model.

[0014] Among them, "S+1" in the Sentinel-1A satellite SAR data represents multi-scene time series data containing reference images, which is used to construct a time-space interference baseline.

[0015] The InSAR data is subjected to interference processing, including radiation scaling and polarization filtering preprocessing of the obtained SAR data; satellite orbit error correction based on precise orbit ephemeris; determination of the spatial baseline parameters and time interval of adjacent images through baseline estimation; generation of an original interferogram and coherence calculation, and screening of high-coherence pixels.

[0016] The terrain residual correction includes calculating a terrain phase component by using a digital elevation model; separating the terrain phase and the deformation phase by a phase unwrapping technique; and removing the terrain phase component from the total phase to obtain a pure deformation phase.

[0017] The atmospheric delay phase modeling and correction includes establishing an empirical model of the atmospheric delay phase and geographical coordinates and time parameters; estimating the atmospheric delay phase of each pixel by a space-time interpolation method; and removing the atmospheric delay phase from the deformation phase to obtain a corrected deformation rate result.

[0018] The improved DeepLabV3+ neural network includes the following optimizations: introducing an attention mechanism module in the encoder part to enhance feature extraction capability; adding a multi-scale feature fusion layer in the decoder part to optimize edge detail recognition; designing a specific loss function for landslide hidden danger point features to improve small target detection accuracy.

[0019] The planning of the unmanned aerial vehicle survey route includes delimiting an unmanned aerial vehicle flight survey area based on key deformation areas and suspected hidden danger points; and adopting a grid or spiral flight route design to ensure that the key area coverage rate and image overlap rate meet the requirements of three-dimensional reconstruction.

[0020] This invention discloses a method for identifying landslide hazard points based on a space-air remote sensing fusion neural network. The method acquires S+1 scenes of Sentinel-1A satellite SAR data, a digital elevation model, a vector boundary map, and precise orbital ephemeris for the study area. Interferometric processing is performed on the InSAR data, followed by deformation rate inversion, terrain residual correction, and atmospheric delay phase modeling and correction to obtain the deformation rate results for the study area and mark key deformation zones. The original high-resolution remote sensing image is preprocessed, and an improved DeepLabV3+ neural network is used to interpret the high-resolution image, outputting the landslide hazard points for the study area. The spatial distribution of potential landslide sites is analyzed. A drone aerial survey route is planned based on this spatial distribution to acquire high-precision images, which are then processed for distortion correction, feature matching, dense point cloud generation, and orthophoto generation. SAR deformation data, high-resolution remote sensing interpretation results, and drone aerial survey data are integrated for multi-dimensional feature analysis, combined with an expert knowledge base for comprehensive interpretation. The spatial location of the potential landslide sites is verified and confirmed, and a three-dimensional model of the demonstration site is constructed using oblique photography. This method integrates high-resolution optical images from the Gaofen-2 satellite, single-look complex (SLC) data from the Sentinel-1A satellite, and a three-dimensional digital model generated by drone oblique photography. Ground deformation results were acquired using SBAS-InSAR technology, and key deformation areas were marked. Then, based on landslide morphological features revealed by optical imagery, a DeeplabV3+ semantic segmentation algorithm was used as the baseline network. An attention module was introduced into the residuals, and an adaptive multi-scale feature fusion module, ASFE-Wlock, was designed to fuse features at different scales in the backbone network. This was verified by combining geological expert experience with neural network theory to initially identify suspected landslide hazard points. Subsequently, UAV technology was used to conduct on-site verification of the hazard points. Through detailed analysis of landslide signs and topography, the total number of potential landslide hazard points in the area was finally determined. This solves the problem that existing landslide hazard identification technologies are difficult to implement high-frequency measurements and have poor identification of landslide hazards formed in a short period. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is an overall block diagram of the landslide hazard identification method based on the fusion of aerospace remote sensing and neural network provided in this invention.

[0023] Figure 2 This is a flowchart of the method for obtaining surface deformation provided by the present invention.

[0024] Figure 3 This is an improved DeeplabV3+ structure diagram provided by the present invention.

[0025] Figure 4 This is a structural diagram of ECA-Net provided in this invention.

[0026] Figure 5 This is a structural diagram of the ASFE-Wlock provided in this invention.

[0027] Figure 6 This is a map showing the distribution of landslide hazard points confirmed by drones, provided by an embodiment of the present invention.

[0028] Figure 7 This is a flowchart of the landslide hazard identification method based on the fusion neural network of space-air remote sensing provided in this invention. Detailed Implementation

[0029] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0030] Please see Figures 1 to 7 This invention provides a method for identifying landslide hazard points based on a fusion neural network of space-air remote sensing, comprising the following steps;

[0031] S1 acquires S+1 scene Sentinel-1A satellite SAR data, digital elevation model, vector boundary map and precise orbital ephemeris of the study area;

[0032] In this embodiment of the invention, "S+1" in the Sentinel-1A satellite SAR data represents multiple time-series data including a reference image, used to construct a time-space interferometric baseline.

[0033] S2 performs interferometric processing on InSAR data and performs deformation rate inversion, topographic residual correction, and atmospheric delay phase modeling and correction to obtain deformation rate results for the study area and mark key deformation areas.

[0034] In this embodiment of the invention, the interferometric processing of InSAR data includes radiometric calibration and polarization filtering preprocessing of the acquired SAR data; satellite orbit error correction based on precise orbital ephemeris; determination of spatial baseline parameters and time intervals of adjacent images through baseline estimation; generation of original interferograms and coherence calculation, and selection of high coherence pixels for subsequent deformation analysis.

[0035] The terrain residual correction includes calculating the terrain phase component using a digital elevation model; separating the terrain phase from the deformation phase using phase unwrapping technology; and removing the terrain phase component from the total phase to obtain the pure deformation phase.

[0036] The atmospheric delay phase modeling and correction includes establishing an empirical model of atmospheric delay phase and geographic coordinates and time parameters; estimating the atmospheric delay phase of each pixel using spatiotemporal interpolation methods; and subtracting the atmospheric delay phase from the deformation phase to obtain the corrected deformation rate result.

[0037] S3 preprocesses the original high-resolution remote sensing images and uses an improved DeepLabV3+ neural network to interpret the high-resolution images, outputting the spatial distribution of suspected landslide hazard points in the study area.

[0038] In this embodiment of the invention, the improved DeepLabV3+ neural network includes the following optimizations: an attention mechanism module is introduced in the encoder to enhance feature extraction capability; a multi-scale feature fusion layer is added in the decoder to optimize edge detail recognition; and a specific loss function is designed for landslide hazard point features to improve small target detection accuracy.

[0039] S4 plans the UAV aerial survey route for the spatial distribution of the landslide hazard points, acquires high-precision images, and performs distortion correction, feature matching, dense point cloud generation, and orthophoto generation processing.

[0040] In this embodiment of the invention, the UAV aerial survey route includes delineating the UAV flight survey area based on the key deformation areas marked in step four and the suspected hidden danger points output in step six; a gridded or spiral route design is adopted to ensure that the coverage rate and image overlap rate of key areas meet the requirements of 3D reconstruction. UAV image processing involves image orientation and regional network adjustment using bundle adjustment; high-density point cloud data is generated using the PMVS algorithm; and orthophoto geometric correction is performed based on ground control points.

[0041] S5 integrates SAR deformation data, high-resolution remote sensing interpretation results, and UAV aerial survey data to perform multi-dimensional feature analysis, and combines expert knowledge base for comprehensive interpretation.

[0042] In this embodiment of the invention, multidimensional feature analysis includes temporal deformation rate and gradient features in SAR deformation data; land cover type, surface texture and crack features in high-resolution remote sensing interpretation results; three-dimensional terrain morphology, slope and free surface features in UAV aerial survey data; and the construction of landslide hazard point discrimination feature vectors through feature fusion algorithms.

[0043] The comprehensive interpretation, which combines an expert knowledge base, includes establishing a prior knowledge database containing geological structures, hydrological conditions, and historical disaster records; designing a rule-based inference engine to perform logical verification of multi-source features; and using a visual interactive interface to assist experts in manually reviewing suspected potential hazards.

[0044] S6 verified and confirmed the spatial location of the landslide hazard point and constructed a 3D model of the demonstration point using oblique photography.

[0045] In this embodiment of the invention, verifying and confirming the spatial location of landslide hazard points includes obtaining the actual coordinates of the hazard points through on-site reconnaissance; verifying consistency by comparing the InSAR deformation time series with the UAV three-dimensional deformation monitoring results; and using statistical significance tests to eliminate misjudgments caused by random errors.

[0046] To better understand this technical solution, the following embodiments are provided for further explanation:

[0047] Example:

[0048] This invention proposes a landslide hazard identification method based on a space-air remote sensing fusion neural network, integrating high-resolution optical imagery from the Gaofen-2 satellite, SLC data from the Sentinel-1A satellite, and a 3D digital model obtained through UAV oblique photogrammetry. The core idea is to obtain the cumulative deformation of the study area using SBAS-InSAR technology to identify high-deformation areas requiring focused attention. Based on this, and combined with the description of the external characteristics of landslide hazard points from the National Geological Disaster Management Agency, an improved DeeplabV3+ neural network interpretation method is used to further confirm suspected landslide hazard points. To verify the identification results, this study deploys UAVs at suspected landslide hazard points to conduct high-precision oblique photogrammetry, observing the external characteristics of the hazard points in detail to determine their precise locations. Simultaneously, a refined 3D model of a landslide hazard demonstration point is created to provide data support and reference for subsequent monitoring and research. The specific method flow is as follows: Figure 1 As shown, this method, from data acquisition to hazard identification, embodies systematicity and scientific rigor, fully demonstrating the crucial role of multi-source remote sensing fusion technology in landslide hazard monitoring.

[0049] First, deformation monitoring and analysis of the study area were conducted using SBAS-InSAR deformation monitoring technology.

[0050] This technology generates a large number of differential interferograms using the original SLC imagery, effectively reducing or eliminating phase errors, atmospheric errors, and terrain errors in subsequent processing.

[0051] Acquire S+1 SLC images of the same target region, spaced 12 days apart, with the representative times for each image being t0, t1, ... t2. S The main image represents time t.A Other times t Q This means that from the image, image registration to the master image, M interferograms can be generated. M satisfies the following condition:

[0052]

[0053] Establish a pixel interferometric phase coordinate system with azimuth x as the horizontal coordinate and distance s as the vertical coordinate. Assume the i-th differential interferogram is derived from the main image t. A With image t Q The SLC image corresponding to time t is generated, and the interference phase at (x,s) at that moment can be expressed as:

[0054]

[0055] In the formula:

[0056]

[0057] Where d(t) A ,x,s) and d(t Q (x,s) represents the cumulative deformation of the master and slave images at time t0 in the LOS direction; Indicates the deformation phase; This represents the residual topographic phase in differential interferometry; Indicates the delayed atmospheric phase; This represents the noise phase of the original data.

[0058] When using the Delaunay MCF algorithm for phase unwrapping, it is well-suited for unwrapping low-coherence regions because it only needs to process the portion with coherence greater than a threshold, without considering all pixels in the image. After all differential interferograms have undergone correct phase unwrapping, and the unwrapped phases are corrected to a high-coherence pixel with a stable deformation trend or known deformation information, a system of equations can be constructed from S SAR images and M equations (the number of differential interferograms). This system of equations can be expressed in matrix form as follows:

[0059]

[0060] In the above formula, matrix B is an M×S dimensional approximate correlation matrix, which can be directly obtained from a known differential interferogram. S+1 SAR images are typically divided into several short baseline sets. To obtain a unique solution, singular value decomposition (SVD) is used to decompose matrix B, find its inverse, and then calculate the vector... The least-normal squared solution is obtained, and finally the ground deformation is calculated by integrating the cumulative deformation rate of the S-scene image over each time period. The specific workflow of SBAS-InSAR is as follows: Figure 2 As shown:

[0061] Input S+1 original SAR images, import precise orbit data and study area vector map to crop out the specific study area;

[0062] Optimal time and spatial baseline thresholds are set, and interferograms are obtained through differential interferometry, supplemented by digital elevation models to eliminate ground phase.

[0063] The residual phase is obtained by unwrapping the previous step using Delaunay MCF, and the nonlinear phase and linear phase are accumulated to obtain the complete surface deformation phase;

[0064] The deformation is estimated by finding a unique solution using singular value decomposition.

[0065] Spatial low-pass filtering and temporal high-pass filtering are applied to the residual phase after removing linear deformation, thereby separating the atmospheric phase; after removing topographic, atmospheric, and noise phase images, the average displacement rate and DEM correction coefficient of the study area are estimated.

[0066] By geocoding the estimated average displacement rate and integrating the time series data of the study area, the deformation results can be obtained.

[0067] Next, the criteria for identifying suspected landslide hazard points are clarified. In InSAR deformation results, suspected landslide hazard points are usually clustered and show significant differences from the surrounding environment. Their main characteristics are significant contrasts between InSAR high-deformation points and surrounding areas in terms of shape, color, shadow, gullies, cracks, and surface vegetation growth. Through the interpretation and analysis of geological landslide hazards using high-resolution remote sensing imagery, combined with processed high-resolution remote sensing images and surface deformation maps, a neural network method can effectively identify suspected landslide hazard points. The improved DeeplabV3+ network proposed in this invention (e.g., Figure 3 As shown, based on the classic DeeplabV3+ architecture, this paper addresses semantic segmentation tasks in complex scenes by optimizing the backbone network feature extraction process, introducing an efficient channel attention mechanism (ECA) with residual connections, and then introducing an adaptive feature fusion module (ASFE-Wlock) to fuse the first three stages of the backbone network before performing attention mechanism feature fusion with the first stage. Finally, a boundary guidance module (Bgm) is introduced to enhance feature extraction capabilities. The network takes remote sensing imagery as input and outputs pixel-level semantic segmentation results, achieving accurate differentiation between foreground targets and background. The overall architecture consists of four main modules: Encoder, Backbone, ASPP module, and Decoder. Through multi-scale feature extraction, context enhancement, and cross-level feature fusion, it solves the shortcomings of traditional segmentation models in recognizing small targets, regions with blurred boundaries, and complex scenes.

[0068] Then, an attention mechanism module is proposed. The input image is first subjected to lightweight convolution for preliminary feature extraction, generating a feature map with a resolution of 1 / 4 of the original image (256 channels), completing the initial encoding of basic edge and texture information. The backbone network is built based on the classic convolution module, generating a three-level feature map through cascaded multi-scale atrous convolution. An ECA-Net module is introduced at the ASPP input. The ECA-Net attention mechanism module focuses on the channel weights that are more beneficial to the landslide hazard segmentation results, strengthens key semantic features, and suppresses redundant background noise. Its structure is as follows: Figure 4 As shown.

[0069] Subsequently, the adaptive attention-based multi-scale feature fusion module ASFE-Wlock was proposed. This study uses the ASFF (Adaptive Spatial Feature Fusion) method to fuse feature maps from Stage 1, Stage 2, and Stage 3. The specific process is as follows: Let X... n Let n represent the feature map of stage n. Using stage 1 as a baseline, the feature maps of stage 2 and stage 3 are first adjusted to have the same number of channels as stage 1 through a 1×1 convolution. Then, the feature maps of stage 2 and stage 3 are upsampled to align their spatial dimensions with stage 1, thus unifying the multi-scale features in terms of channel number and spatial dimensions. Finally, an adaptive feature fusion mechanism is used to weight and fuse the vectors of the three stage feature maps at spatial location (a, b), where the fusion weights are adaptively learned by the neural network. The mathematical expression is as follows:

[0070]

[0071] In the formula Y ab It is the fused feature vector, α ab ,β ab γ ab This represents three importance weights, and α ab +β ab +γ ab =1.

[0072] The result of adaptively fusing the three stage modules is called the adaptive module. Then, an iterative attention feature fusion module is introduced. This module fuses the adaptive module's feature map with the feature map from stage 1. This module is implemented based on the multi-scale feature attention module (MS-CAM). Furthermore, an adaptive attention multi-scale feature fusion module, ASFE-Wlock, is proposed, with the structure shown in the figure. Figure 5 As shown. The attention calculation formula for local features is K(X), and the feature extraction method is point convolution, with the following mathematical formula:

[0073] K(X)=B(PWConv2(λ(B(PWConv1(X)))))

[0074] In the formula, PWConv1 represents a 1×1 point convolution, which reduces the original number of channels to 1 / r; B represents a normalization layer; λ represents the ReLU activation function; and PWConv2 restores the number of channels to r. The attention calculation formula for global features is L(X), and the mathematical formula is as follows:

[0075] L(X)=B(PWConv2(λ(B(PWConv1(GCP(X))))))

[0076] The outputs of K(X) and L(X) are combined with the X weights obtained through the activation function. The weighted sum of X and Y is then used to obtain the final feature fusion map.

[0077] Finally, landslide hazard points were confirmed using UAV oblique photogrammetry. UAV oblique photogrammetry technology, with its mobility, efficiency, speed, and high precision, has significant advantages in rapidly acquiring high-resolution images of small areas and mountainous regions. This study conducted UAV oblique photogrammetry surveys on 22 suspected landslide hazard points selected in the study area. The results showed that 19 suspected hazard points were verified, with an identification accuracy rate of 86.36%. It is evident that the distribution characteristics of landslide hazard points are significantly influenced by topographical conditions. The identification results are shown in the image below. Figure 6 As shown.

[0078] The above-disclosed embodiments are merely preferred embodiments of the landslide hazard identification method based on a fusion neural network of space-air remote sensing according to the present invention. Of course, they should not be construed as limiting the scope of the present invention. Those skilled in the art can understand that all or part of the processes of the above embodiments can be implemented, and equivalent changes made in accordance with the claims of the present invention are still within the scope of the invention.

Claims

1. A method for identifying landslide hazard points based on a fusion neural network of space-air remote sensing, characterized in that, Includes the following steps; Acquire S+1 scene Sentinel-1A satellite SAR data, digital elevation model, vector boundary map and precise orbit ephemeris of the study area; Interferometric processing was performed on InSAR data, followed by deformation rate inversion, topographic residual correction, and atmospheric delay phase modeling and correction to obtain deformation rate results for the study area and mark key deformation zones. The original high-resolution remote sensing images were preprocessed, and the improved DeepLabV3+ neural network was used to interpret the high-resolution images to output the spatial distribution of suspected landslide hazard points in the study area. Plan UAV aerial survey routes for the spatial distribution of the landslide hazard points, acquire high-precision images, and perform distortion correction, feature matching, dense point cloud generation, and orthophoto generation processing. Multidimensional feature analysis is performed by integrating SAR deformation data, high-resolution remote sensing interpretation results and UAV aerial survey data, and comprehensive interpretation is carried out by combining expert knowledge base; Verify and confirm the spatial location of landslide hazard points, and construct a 3D model of the demonstration point using oblique photography.

2. The landslide hazard identification method based on a fusion neural network of space-air remote sensing as described in claim 1, characterized in that... ； In the Sentinel-1A satellite SAR data, "S+1" indicates multiple time-series data including a reference image, used to construct a time-space interferometric baseline.

3. The landslide hazard identification method based on a fusion neural network of space-air remote sensing as described in claim 1, characterized in that; The interferometric processing of InSAR data includes radiometric calibration and polarization filtering preprocessing of the acquired SAR data; satellite orbit error correction based on precise orbit ephemeris; determination of spatial baseline parameters and time intervals of adjacent images through baseline estimation; generation of original interferograms and coherence calculation, and selection of high coherence pixels.

4. The landslide hazard identification method based on a fusion neural network of space-air remote sensing as described in claim 1, Its characteristics are: The terrain residual correction includes calculating the terrain phase component using a digital elevation model; and separating the terrain phase from the deformation phase using phase unwrapping technology. Remove the terrain phase component from the total phase to obtain the pure deformation phase.

5. The landslide hazard identification method based on a fusion neural network of space-air remote sensing as described in claim 1, characterized in that; The atmospheric delay phase modeling and correction includes establishing an empirical model of atmospheric delay phase and geographic coordinates and time parameters; estimating the atmospheric delay phase of each pixel using spatiotemporal interpolation methods; and subtracting the atmospheric delay phase from the deformation phase to obtain the corrected deformation rate result.

6. The landslide hazard identification method based on a fusion neural network of space-air remote sensing as described in claim 1, Its characteristics are: The improved DeepLabV3+ neural network includes the following optimizations: an attention mechanism module is introduced in the encoder to enhance feature extraction capabilities; a multi-scale feature fusion layer is added in the decoder to optimize edge detail recognition; and a specific loss function is designed for landslide hazard point features to improve small target detection accuracy.

7. The landslide hazard identification method based on a fusion neural network of space-air remote sensing as described in claim 1, Its characteristics are: The planned UAV aerial survey route includes delineating the UAV flight survey area based on key deformation areas and suspected potential hazard points; and adopting a gridded or spiral route design to ensure that the coverage rate and image overlap rate of key areas meet the requirements of 3D reconstruction.

Images