A geological disaster early warning method based on InSAR and deep learning

Abstract

This invention discloses a geological disaster early warning method based on InSAR and deep learning, belonging to the field of disaster early warning technology. The method includes: collecting geological disaster monitoring data, including L-SAR ascending and descending trajectory data, DEM data, optical image data, and disaster-induced data; performing SBAS-InSAR calculation on the L-SAR ascending and descending trajectory data to generate ascending and descending trajectory line-of-sight deformation variables; performing temporal interpolation registration and spatial homologous pixel registration on the ascending and descending trajectory line-of-sight deformation variables to generate registered line-of-sight deformation variables; calculating a two-dimensional deformation field and slope aspect deformation variables based on the registered line-of-sight deformation variables and DEM data; rasterizing the slope aspect deformation variables to obtain a slope aspect deformation map, which is then input into a MaskRCNN model improved based on the Involution operator and deformable pooling module to generate deformation anomaly segmentation results. This invention improves the identification of the true deformation direction of the slope and the linkage capability of multi-source induced information, enhancing the spatial matching and timeliness of dynamic geological disaster early warning.

Description

Technical Field

[0001] This invention relates to the field of disaster early warning technology, specifically to a geological disaster early warning method based on InSAR and deep learning. Background Technology

[0002] InSAR remote sensing monitoring can acquire information on minute surface deformations using synthetic aperture radar imagery, while deep learning image processing can automatically identify abnormal areas in remote sensing images. The combination of the two can be used to investigate potential geological hazards in mountainous areas, mining areas, reservoir banks, and transportation corridors.

[0003] Currently, in the process of identifying geological hazards based on radar interferometry, the deformation information obtained from a single observation track is mainly limited by the radar line of sight, making it difficult to fully reflect the deformation changes of the slope along the actual sliding direction, resulting in a spatial deviation between the identification boundary and the actual location of the landslide hazard.

[0004] Secondly, in the process of geological disaster early warning and judgment, topographic relief, image texture edges, abnormal surface deformation and rainfall triggering factors are usually analyzed separately. The correspondence between spatial scale and time scale is not continuous enough. When static hazard conditions are not linked with triggering conditions such as short-term rainfall and cumulative rainfall, the early warning judgment is prone to lag behind external induced changes, affecting the reliability of risk level classification. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a geological disaster early warning method based on InSAR and deep learning.

[0006] A geological disaster early warning method based on InSAR and deep learning includes: Collect geological disaster monitoring data, including L-SAR ascending and descending orbit data, DEM data, optical image data and disaster-induced data, perform SBAS-InSAR calculation on L-SAR ascending and descending orbit data, and generate ascending and descending orbit line-of-sight deformation variables; Temporal interpolation registration and spatial same-name pixel registration are performed on the line-of-sight deformation variables of the lifting rail to generate registered line-of-sight deformation variables. Based on the registered line-of-sight deformation variables and DEM data, the two-dimensional deformation field and slope deformation variables are calculated. The aspect deformation is rasterized to obtain an aspect deformation map, which is then input into the MaskRCNN model improved based on the Involution operator and deformable pooling module to generate deformation anomaly segmentation results. Based on the results of deformation anomaly zoning, DEM data, optical image data, and disaster-induced data, dynamic early warning results for geological disasters are generated.

[0007] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention utilizes the correspondence between the elevation and descent rail line-of-sight deformation variables, the registration line-of-sight deformation variables, the two-dimensional deformation field, and the slope aspect deformation variables. This allows the deformation information of the elevation and descent rails to jointly reflect the deformation characteristics of the slope in the east-west and vertical directions, and ensures that the slope aspect deformation variables are consistent with the slope sliding direction. This enhances the ability of the slope aspect deformation variable map to express the spatial orientation of landslide hazards, thereby improving the matching between the deformation anomaly zone segmentation results and the actual hazard area. Furthermore, this invention also enables the deformation anomaly intensity value, optical disturbance value, and susceptibility level to participate in the generation of dynamic early warning results for geological disasters by linking the deformation anomaly segmentation results, DEM data, optical image data, and disaster-induced data with the inducing factor value. This reduces the early warning deviation caused by judging based solely on a single type of deformation or static terrain conditions, and ensures that the susceptibility assessment and rainfall-triggered changes remain continuously correlated. In summary, this invention improves the ability to identify the true deformation direction of slopes and link multi-source induced information, thereby enhancing the spatial matching and timeliness of dynamic early warning of geological disasters. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0009] Figure 1 The flowchart illustrates a geological disaster early warning method based on InSAR and deep learning provided by this invention. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0011] Please see Figure 1 As shown in the figure, this embodiment discloses a geological disaster early warning method based on InSAR and deep learning, the method including: S11: Collect geological disaster monitoring data, including L-SAR ascending and descending orbit data, DEM data, optical image data and disaster-induced data, perform SBAS-InSAR calculation on L-SAR ascending and descending orbit data, and generate ascending and descending orbit line-of-sight deformation. Specifically, the L-SAR ascending and descending orbit data includes ascending orbit SLC imagery, descending orbit SLC imagery, precise orbit recording, ascending orbit imaging time, descending orbit imaging time, and pixel position; The DEM data includes elevation values, slope values, aspect angle, DEM spatial resolution values, and DEM grid size; The optical image data includes texture values, edge values, optical spatial resolution values, and optical grid size; The disaster-induced data includes previous effective rainfall, short-term heavy rainfall, cumulative rainfall, and rainfall collection time.

[0012] The ascending-orbit SLC imagery and descending-orbit SLC imagery were acquired by L-band synthetic aperture radar satellites. The precise orbit record is read from the satellite orbit determination file, including the satellite's three-dimensional orbital position and orbital time identifier corresponding to each ascent and descent imaging moment; The cell location is obtained by converting the row and column numbers of the ascending and descending SLC images to a unified map coordinate system through distance Doppler geocoding. The cell location includes the cell row and column coordinates, horizontal coordinates, and raster number. DEM data is read from the digital elevation model and registered to a unified map coordinate system according to the pixel location; Optical image data is read from optical remote sensing images or UAV optical images; Disaster-induced data are obtained from rain gauges, meteorological grid data, or geological disaster monitoring platforms.

[0013] In this embodiment, the character Indicates an upgrade indicator. Indicates a drop-track indicator. , Indicates the image sequence number. Indicates the cell position number. Indicates the reference imaging time sequence number.

[0014] Specifically, the steps for generating the line-of-sight deformation of the ascending and descending rails are as follows: S111: Based on the ascending SLC image, descending SLC image, ascending imaging time, descending imaging time, and precision track record, calculate the ascending time baseline value, ascending space baseline value, descending time baseline value, and descending space baseline value respectively, and generate the ascending and descending baseline value record. Specifically, based on the first SLC images and the first Up-track SLC images, read the up-track imaging time corresponding to two up-track SLC images. , The time interval between the two ascent imaging moments is used as the ascent time baseline value. The calculation formula is: ; Two-dimensional horizontal coordinates were read based on the central pixel location of the study area. And read the elevation value corresponding to the same central pixel location based on the DEM data. The two-dimensional horizontal coordinates and elevation values ​​are written into the same three-dimensional coordinate system to obtain the three-dimensional position of the central surface target point. , is represented as: ; It should be noted that the two-dimensional horizontal coordinates and elevation values ​​are converted to a three-dimensional geocentric coordinate system consistent with the satellite's three-dimensional orbital position in the precise orbit record before being written into the three-dimensional coordinate system.

[0015] Read the first from the precision track record Satellite 3D orbit position corresponding to the rising orbit SLC image and the Satellite 3D orbit position corresponding to the rising orbit SLC image ; Based on the The three-dimensional orbital position of the satellite and the three-dimensional position of the central surface target point corresponding to the rising orbit SLC image are used to calculate the unit vector of the radar line-of-sight direction of the rising orbit main image. , is represented as: ; In the formula, For vector magnitude operations; Based on the Satellite 3D orbit position corresponding to the rising orbit SLC image and the Satellite 3D orbit position corresponding to the rising orbit SLC image Calculate the baseline vector of the ascending orbit. , is represented as: ; The vertical baseline vector of the ascending orbit is obtained by subtracting its projection component onto the unit vector in the line-of-sight direction of the ascending orbit main image radar from the ascending orbit baseline vector. , is represented as: ; In the formula, This is a vector dot product operation.

[0016] Normalize the vertical baseline vector of the ascending orbit to obtain the unit vector of the vertical baseline direction of the ascending orbit. , is represented as: ; Projecting the baseline vector of the ascending orbit onto the unit vector in the direction perpendicular to the ascending orbit baseline yields the spatial baseline value of the ascending orbit. , is represented as: ; Based on the SLC imagery and the first De-orbiting SLC images, and read the de-orbiting imaging times corresponding to two de-orbiting SLC images. , The time interval between the two descent imaging moments is used as the descent time baseline value. The calculation formula is: ; Read the first from the precision track record Satellite 3D orbit position corresponding to the SLC image of the descent orbit and the Satellite 3D orbit position corresponding to the SLC image of the descent orbit Following the calculation method for the ascent orbit baseline value, the unit vector of the descent orbit main image radar line-of-sight direction is obtained sequentially. , Descending orbit baseline vector Vertical baseline vector of the descending orbit Unit vector of the vertical baseline direction of the descending orbit , is represented as: ; ; ; ; Projecting the descending orbit baseline vector onto the unit vector in the direction perpendicular to the descending orbit baseline yields the descending orbit spatial baseline value. , is represented as: ; Write the image number, ascent time baseline value, ascent space baseline value, descent time baseline value, and descent space baseline value into the same record item to generate an ascent / descent baseline value record.

[0017] S112: Filter the small baseline image pairs of ascending and descending orbits by recording the baseline values ​​of ascending and descending orbits; Specifically, read the lifting time baseline value from the lifting rail baseline value record. and the baseline value of the ascending orbit space And read the preset ascent time baseline threshold. and the baseline threshold of the ascent space When the same uptracked image pair satisfies the following formula, the uptracked image pair is written into the uptracked small baseline image pair: ; ; Read the lowering time baseline value from the lowering rail baseline value record. and the baseline value of the lower orbit space And read the preset descent time baseline threshold. and the lowering orbit space baseline threshold When the same down-track image pair satisfies the following formula, the down-track image pair is written into the down-track small baseline image pair: ; ; The time baseline threshold for orbit ascent, the space baseline threshold for orbit ascent, the time baseline threshold for orbit descent, and the space baseline threshold for orbit descent were determined through experimental data. In this embodiment, the time baseline threshold for orbit ascent and the time baseline threshold for orbit descent are both set to 90 days, and the space baseline threshold for orbit ascent and the space baseline threshold for orbit descent are both set to 1000m.

[0018] S113: Based on the precise orbit records, perform differential interferometry calculations on the ascending orbit small baseline image pairs and the descending orbit small baseline image pairs respectively to generate ascending orbit differential interferometry records and descending orbit differential interferometry records.

[0019] Specifically, based on the first of the small baseline image pairs in the rising orbit... SLC images and the first SLC imagery will be the first Using the SLC imagery of the rising orbit as the main image, the first... SLC imagery with rising orbit is used as supplementary imagery; The auxiliary image is registered so that the pixel positions of the auxiliary image are mapped to the corresponding pixel positions of the primary image. The registered uptracked auxiliary image is then processed using complex numerical values. Represented as: ; In the formula, The complex image interpolation function performs bicubic convolution interpolation on the real and imaginary parts of the complex values ​​of the auxiliary image, and combines the interpolated real and imaginary parts into the registered complex values ​​of the auxiliary image.

[0020] Read the first Jingsheng Track SLC image in the first The complex value corresponding to each pixel position The complex values ​​of the main image and the registered auxiliary image are multiplied by conjugate to obtain the ascending orbit interferometry complex values. The phase value of the ascending orbit interference is extracted from the complex value of the ascending orbit interference. , is represented as: ; ; In the formula, It is a complex argument function.

[0021] Read the first according to the cell position The horizontal coordinates of each pixel position And read the elevation value corresponding to the same pixel location based on the DEM data. By writing the horizontal coordinates and elevation values ​​into the same three-dimensional coordinate system, the three-dimensional location of the target point on the ground is constructed. By writing the same horizontal coordinates and reference zero elevation into the same three-dimensional coordinate system, the three-dimensional position of the reference surface target point is constructed. , is represented as: ; ; Read the first from the precision track record Satellite 3D orbit position corresponding to the rising orbit SLC image and the Satellite 3D orbit position corresponding to the rising orbit SLC image Calculate the line-of-sight distance of the main image radar for the rising orbit. Line-of-sight radar for elevated orbital imaging The main image reference radar line-of-sight distance for the orbital ascent And the line-of-sight distance of the image reference radar for the orbital-assisted imagery , is represented as: ; ; ; ; Calculate the ground phase value of the ascent orbit based on the line-of-sight distance of the main ascent orbit image reference radar and the line-of-sight distance of the auxiliary ascent orbit image reference radar. The terrain phase value for the ascent is calculated based on the line-of-sight distances of the primary ascent radar, the secondary ascent radar, the primary ascent radar reference radar, and the secondary ascent radar reference radar. , is represented as: ; ; In the formula, This refers to the L-band radar wavelength.

[0022] Subtracting the flat-land phase value and the terrain phase value from the ascending orbit interferometric phase value yields the ascending orbit differential interferometric phase value. , is represented as: ; In the formula, This is a phase wrapping function used to restrict the phase value to a certain value. Within the range; Goldstein filtering is applied to the ascending orbit differential interferometry phase values ​​to obtain the filtered ascending orbit differential interferometry phase values. , is represented as: ; In the formula, For Fourier transform, This is the inverse Fourier transform. The imaginary unit, This is the Goldstein filter index, which was determined through experimental data.

[0023] The first of the low baseline image pairs SLC imagery and the first Using the SLC image with reduced orbit, the complex values ​​of the reduced orbit interferometry are obtained sequentially according to the calculation method of the phase value of the differential interferometry for the rising orbit described above. Phase value of orbital deflection interference , Descending orbit level phase value , Terrain phase value of the de-orbiting orbit , differential interferometry phase value and the filtered down-orbit differential interferometric phase value ; Write the image pair number, pixel position, and filtered differential interferometric phase value of the ascending orbit small baseline image into the same record item to generate an ascending orbit differential interferometric record; write the image pair number, pixel position, and filtered differential interferometric phase value of the descending orbit small baseline image into the same record item to generate a descending orbit differential interferometric record.

[0024] S114: Perform phase unwrapping and time-series inversion on the ascending and descending differential interferometric records respectively to obtain the ascending and descending line-of-sight deformation variables, and generate ascending and descending line-of-sight deformation variables according to the pixel positions.

[0025] Specifically, minimum-cost flow phase unwrapping is performed on the filtered ascending-orbit differential interferometry phase values ​​in the ascending-orbit differential interferometry record to obtain the ascending-orbit unwrapped phase values. The filtered reduced-orbit differential interferometric phase values ​​in the reduced-orbit differential interferometric record are subjected to minimum-cost flow phase unwrapping to obtain the reduced-orbit unwrapped phase values. .

[0026] The deformation increment along the line of sight of the ascending orbit is calculated using the unwrapping phase value of the ascending orbit. The deformation increment along the line of sight of the reduced orbit is calculated by using the phase value of the unwrapped phase value. , is represented as: ; ; Arrange the deformation increments along the line of sight for each ascending orbit small baseline image according to the ascending orbit imaging time to construct a set of equations for ascending orbit small baseline deformation increments; perform least squares solution on the set of equations to obtain the deformation rate vectors of adjacent ascending orbit time periods. , is represented as: ; ; In the formula, For the first The deformation increment vector of the ascending orbit line of sight corresponding to each pixel position. This is the orbital timing design matrix constructed from the orbital small baseline images corresponding to the orbital imaging times.

[0027] The deformation rate vector of adjacent time periods of the ascending orbit and the imaging time interval of adjacent time periods of the ascending orbit are accumulated to obtain the deformation of the line of sight of the ascending orbit. , is represented as: ; In the formula, The first element in the deformation rate vector of adjacent time periods of the ascending orbit One deformation rate value, For the first The interval between adjacent imaging times of each ascending orbit.

[0028] The deformation along the line of sight during orbit reduction is calculated from the deformation increment along the line of sight during orbit reduction, the orbit reduction timing design matrix, and the time interval between adjacent imaging moments during orbit reduction, and is expressed as follows: ; ; ; Pair the ascending and descending line-of-sight deformation variables corresponding to the same pixel position to generate ascending and descending line-of-sight deformation variables.

[0029] S12: Perform temporal interpolation registration and spatial same-name pixel registration on the line-of-sight deformation variables of the lifting rail to generate registered line-of-sight deformation variables. Calculate the two-dimensional deformation field and slope deformation variables based on the registered line-of-sight deformation variables and DEM data. Specifically, the steps for generating the registration line-of-sight deformation are as follows: S121: Arrange the time series of ascending orbit deformation according to the ascending orbit line-of-sight deformation and the ascending orbit imaging time, and arrange the time series of descending orbit deformation according to the descending orbit line-of-sight deformation and the descending orbit imaging time.

[0030] Specifically, the ascent and descent line-of-sight deformation variables are read according to pixel position; the ascent line-of-sight deformation variables corresponding to the same pixel position are arranged in ascending order according to the ascent imaging time to obtain the ascent deformation time series. , is represented as: ; In the formula, This represents the number of times the image was taken during the ascent.

[0031] Arrange the reduced-orbit line-of-sight deformations corresponding to the same pixel location in ascending order according to the reduced-orbit imaging time to obtain the reduced-orbit deformation time series. , is represented as: ; In the formula, This represents the number of moments for down-orbit imaging.

[0032] S122: Merge the ascending orbit imaging time and descending orbit imaging time in chronological order to generate a reference imaging time table; Specifically, the system reads the ascending and descending orbit imaging times, retains the repeated imaging times once, and arranges them in ascending order to generate a reference imaging time table. , is represented as: ; In the formula, For ascending order sorting operations, This is the union operation.

[0033] S123: Input the ascending orbit deformation time series, descending orbit deformation time series and reference imaging time table into the spline interpolation function to obtain the time registration line-of-sight deformation.

[0034] Specifically, regarding the first time in the benchmark imaging timetable Reference imaging time Read adjacent ascending orbit imaging times and their corresponding ascending orbit line-of-sight deformation variables from the ascending orbit deformation time series, and combine the adjacent ascending orbit imaging times, their corresponding ascending orbit line-of-sight deformation variables, and the first ascending orbit deformation time. Inputting a cubic spline interpolation function at each reference imaging time yields the orbital time registration deformation. , is represented as: ; In the formula, To and Adjacent and less than or equal to The moment of orbital ascent imaging, , , and These are the cubic spline interpolation coefficients calculated from the time series of the ascending orbit deformation; It should be noted that reference imaging times in the reference imaging timetable that are outside the coverage of the ascending or descending orbit deformation time series are not included in the spline interpolation calculation for the corresponding orbit.

[0035] Read the adjacent descent imaging times and their corresponding descent line-of-sight deformations from the descent deformation time series, and obtain the descent time registration deformations in the same way. , is represented as: ; In the formula, To and Adjacent and less than or equal to The moment of descent imaging, , , and These are the cubic spline interpolation coefficients calculated from the time series of the deflection deformation.

[0036] Write the time registration deformation variables for ascending and descending orbits corresponding to the same pixel position and the same reference imaging time into the same record to obtain the time registration line-of-sight deformation variables.

[0037] S124: Based on the pixel position, perform same-name pixel registration on the time registration line-of-sight deformation variable to generate the registration line-of-sight deformation variable.

[0038] Specifically, using the raster coordinates of the DEM data as a unified spatial reference, the temporal registration line-of-sight deformation variables are mapped to the corresponding raster coordinates of the DEM data according to their pixel positions. When the spatial resolution of the temporal registration line-of-sight deformation variables is inconsistent with that of the DEM data, bilinear resampling is performed according to the DEM spatial resolution value. Taking the ascending orbit temporal registration deformation variables as an example, the resampled ascending orbit temporal registration deformation variables... Represented as: ; In the formula, , , and For the positions of four adjacent pixels around the floating-point position corresponding to the DEM raster coordinates, and These represent the horizontal and vertical proportions of the floating-point position relative to the adjacent pixel positions, respectively.

[0039] The resampled down-track time registration shape variables are obtained using the same bilinear resampling method; the resampled up-track time registration shape variables and the resampled down-track time registration shape variables are written according to the positions of the same pixel to generate the registration line direction shape variables.

[0040] Specifically, the steps for calculating the two-dimensional deformation field and aspect deformation are as follows: S125: Based on the registration line-of-sight deformation variables, separate the ascending track projection deformation variables and descending track projection deformation variables to generate ascending and descending track projection deformation records.

[0041] Specifically, the resampled up-track time registration deformation variables and resampled down-track time registration deformation variables corresponding to the same pixel position and the same reference imaging time are read from the registration line-of-sight deformation variables; the resampled up-track time registration deformation variables are written into the up-track projection deformation variables. Write the resampled down-orbit time registration shape variables into the down-orbit projection shape variables. , is represented as: ; ; Write the pixel position, reference imaging time, ascending orbit projection deformation, and descending orbit projection deformation into the same record item to generate an ascending orbit projection deformation record.

[0042] S126: Calculate the radar incident angle and radar observation direction values ​​using precise orbit recording and pixel positions to generate a radar observation geometric record.

[0043] Specifically, the first pixel is read based on its position. The horizontal coordinates of each pixel position Elevation values ​​corresponding to the same pixel location are read from the DEM data. By writing the horizontal coordinates and elevation values ​​into the same three-dimensional coordinate system, the three-dimensional location of the target point on the ground is constructed. , is represented as: ; Read the first from the precision track record The three-dimensional orbital position of the ascending satellite at each reference imaging time and the three-dimensional orbital position of the de-orbiting satellite ; It should be noted that when the reference imaging time does not coincide with the corresponding ascent or descent imaging time, time interpolation is performed on the satellite's three-dimensional orbital position at adjacent orbital times based on the orbital time markers in the precise orbital record, to obtain... or .

[0044] Based on the three-dimensional orbital position of the ascending satellite and the three-dimensional position of the target point on the ground Calculate the first The pixel position is at the The observation direction value of the lifting radar at each reference imaging time Based on the three-dimensional orbital position of the de-orbiting satellite and the three-dimensional position of the target point on the ground Calculate the first The pixel position is at the The radar observation direction value corresponding to each reference imaging time. , is represented as: ; ; Construct the first [pixel location] based on the local vertical direction of the map coordinate system where the pixel is located. The local vertical unit vector corresponding to each pixel position In a local coordinate system composed of east-west, north-south, and vertical directions, the local vertical unit vector Values Based on the direction value observed by the ascending radar. and local vertical unit vector Calculate the first The pixel position is at the The incident angle of the lifting radar at each reference imaging moment ; Based on the direction value observed by the de-orbiting radar and local vertical unit vector Calculate the first The pixel position is at the The angle of incidence of the de-orbiting radar at each reference imaging moment , is represented as: ; ; Based on the map coordinate system where the pixel location is located, construct the first... The east-west unit vector corresponding to each pixel position North-South Unit Vector and local vertical unit vector Based on the direction value observed by the rising-orbit radar Respectively with the east-west unit vector North-South Unit Vector and local vertical unit vector The dot product result yields the east-west projection coefficient in the observation direction value of the ascending radar. North-South Projection Coefficient and vertical projection coefficient , is represented as: ; ; ; Direction value observed by the de-orbiting radar Respectively with the east-west unit vector North-South Unit Vector and local vertical unit vector The dot product result yields the east-west projection coefficient in the observation direction value of the down-orbit radar. North-South Projection Coefficient and vertical projection coefficient , is represented as: ; ; ; Write the observation direction value of the ascending radar, the observation direction value of the descending radar, the incident angle of the ascending radar, the incident angle of the descending radar, the east-west projection coefficient, the north-south projection coefficient, and the vertical projection coefficient into the same record item to generate a radar observation geometry record.

[0045] S127: Substitute the elevation and descent rail projection deformation records and radar observation geometric records into the preset elevation and descent rail projection equations to solve for the east-west deformation components and vertical deformation components.

[0046] Specifically, the projection deformation of the ascending and descending orbits is read from the projection deformation records of the ascending and descending orbits, and the east-west projection coefficient and vertical projection coefficient are read from the radar observation geometry records. Since the ascending and descending orbit projection deformations are less sensitive to the north-south deformation component, this embodiment limits the two-dimensional deformation field to the east-west deformation component. and vertical deformation component Substitute the relevant data into the system of two linear projection equations: ; In the formula, For the projection deformation of the ascending orbit, For the reduced orbital projection deformation.

[0047] Calculate the determinant of the projection equations based on the coefficients of the system of two linear projection equations. , is represented as: ; When the absolute value of the determinant of the projection equation is greater than a preset threshold, the system of two linear projection equations is solved to obtain the east-west deformation components. and vertical deformation component , is represented as: ; ; When the absolute value of the determinant of the projection equation is less than or equal to the preset determinant threshold of the projection equation, the corresponding pixel position is written into the invalid pixel position record of the projection, and the east-west deformation component and the vertical deformation component are not written into the pixel position. It should be noted that the preset projection equation determinant threshold is set based on statistical analysis of historical data.

[0048] S128: Arrange the east-west deformation components and vertical deformation components according to the pixel position to construct a two-dimensional deformation field. Then, project the two-dimensional deformation field onto the downward slope direction according to the aspect angle and slope value to obtain the two-dimensional deformation field and aspect deformation.

[0049] Specifically, the east-west and vertical deformation components corresponding to the same pixel location and the same reference imaging time are written into a two-dimensional vector, and arranged according to the pixel location to obtain a two-dimensional deformation field. , is represented as: ; Read the slope angle corresponding to the same cell location from the DEM data. and slope value The aspect angle and slope values ​​are uniformly converted to radians before being used in the calculation; The projection components of the slope sliding direction in the east-west and vertical directions are constructed based on the slope aspect angle and slope value. , , is represented as: ; ; Since the two-dimensional deformation field does not include the north-south deformation component, this embodiment projects the slope's downward direction onto an east-west vertical section and calculates the aspect deformation based on this section. , is represented as: ; Will and After substituting, we get: ; The pixel location, reference imaging time, two-dimensional deformation field, and aspect deformation are written into the same record to obtain the two-dimensional deformation field and aspect deformation.

[0050] S13: Rasterize the aspect deformation image to obtain the aspect deformation map and input it into the MaskRCNN model improved based on the Involution operator and deformable pooling module to generate deformation anomaly segmentation results; Specifically, the steps to obtain the aspect deformation diagram are as follows: S131: Generate a slope deformation pixel record based on the slope aspect deformation and pixel position; Specifically, read the first The pixel position is at the Slope deformation at each reference imaging time And read the row and column coordinates of the cell corresponding to the same cell position. and grid number The pixel location, pixel row and column coordinates, raster number, reference imaging time, and aspect deformation are written into the same record item to generate an aspect deformation pixel record. , is represented as: ; S132: Create raster cells according to the pixel positions, and write the aspect deformation into the raster cells to generate an aspect deformation raster map; Specifically, the DEM raster size is used as the raster size for the aspect deformation raster map, and raster cells are established based on the cell row and column coordinates; the aspect deformation variables in the aspect deformation cell records are written into the raster cells corresponding to the cell row and column coordinates, generating the first... Raster map of slope deformation at each reference imaging time , is represented as: ; When multiple slope aspect variables exist within the same raster cell, the raster number corresponding to the slope aspect variable is read. Aspect deformation The mean of multiple slope aspect variables is written in, and it is represented as follows: ; S133: Normalize the values ​​of the slope aspect deformation raster map to obtain the slope aspect deformation map; Specifically, read the first The slope aspect deformation raster map corresponding to each reference imaging time is used to extract the maximum slope aspect deformation value of the effective raster cells. and minimum value of slope aspect deformation , is represented as: ; ; Based on the maximum and minimum values ​​of the aspect deformation, the aspect deformation raster map is normalized and mapped to obtain the aspect deformation map. , is represented as: ; In the formula, The normalized zero constant of the slope aspect deformation diagram is determined through experimental data.

[0051] Specifically, the Mask R-CNN model improved based on the Involution operator and deformable pooling module includes a backbone feature extraction network, a feature pyramid network, a candidate region generation network, a deformable pooling module, a classification and bounding box regression branch, and a mask segmentation branch. The Involution operator is set in the backbone feature extraction network, and the deformable pooling module is set in the candidate region feature extraction stage.

[0052] Specifically, the steps for generating the Mask R-CNN model improved based on the Involution operator and deformable pooling module are as follows: a1: Read the historical slope deformation map, historical deformation anomaly area mask, historical deformation anomaly area bounding box, and historical anomaly category label, and use the historical slope deformation map as training input, and use the historical deformation anomaly area mask, historical deformation anomaly area bounding box, and historical anomaly category label as training output constraints to generate model training sample records.

[0053] a2: Input the historical aspect deformation map from the model training sample record into the backbone feature extraction network, and perform location-related feature aggregation in the preset feature layer of the backbone feature extraction network through the Involution operator to generate the historical aspect deformation feature map.

[0054] a3: Input the historical slope aspect deformation feature map into the feature pyramid network for multi-scale feature fusion to generate historical multi-scale slope aspect deformation feature map, and input the historical multi-scale slope aspect deformation feature map into the candidate region generation network to generate historical anomaly candidate regions, historical anomaly confidence and historical bounding box offset.

[0055] a4: Input the historical anomaly candidate region into the deformable pooling module, generate a sampling offset based on the local features corresponding to the initial sampling point within the historical anomaly candidate region, and perform position correction and pooling processing on the initial sampling point according to the sampling offset to generate historical anomaly candidate region features.

[0056] a5: Input the historical anomaly candidate region features into the classification and bounding box regression branch to generate historical anomaly category probability vectors and historical bounding box correction values; input the historical anomaly candidate region features into the mask segmentation branch to generate historical pixel-level mask probability maps.

[0057] a6: Calculate the classification loss based on the historical anomaly category probability vector and historical anomaly category identifier; calculate the bounding box regression loss based on the historical bounding box correction amount and the historical deformation anomaly region bounding box; calculate the mask segmentation loss based on the historical pixel-level mask probability map and the historical deformation anomaly region mask; and generate the model training loss by weighting the classification loss, bounding box regression loss, and mask segmentation loss, expressed as: ; In the formula, For model training loss, For classifying losses, For bounding box regression loss, For mask segmentation loss, For classification loss weights, Weights for bounding box regression loss. Weights for mask segmentation loss; It should be noted that the classification loss weights, bounding box regression loss weights, and mask segmentation loss weights are determined by the training error distribution corresponding to the historical aspect deformation map, the historical deformation anomaly area mask, and the historical deformation anomaly area bounding box.

[0058] a7: Update the network weights of the backbone feature extraction network, Involution operator, feature pyramid network, candidate region generation network, deformable pooling module, classification and bounding box regression branch and mask segmentation branch according to the model training loss. When the model training loss satisfies the training convergence condition, the trained Mask R-CNN model, improved based on the Involution operator and deformable pooling module, is obtained.

[0059] Specifically, the logic for generating the deformation anomaly segmentation results is as follows: S134: Input the aspect deformation map into the backbone feature extraction network, and extract the aspect deformation feature map through the Involution operator; Specifically, the aspect deformation diagram The input is processed by the backbone feature extraction network, which first performs a convolutional mapping on the aspect deformation map to obtain the input feature map. For the row coordinates in the input feature map Column coordinates Corresponding feature vector The weight matrix of the Involution operator is completed through training. , Generate the spatial weight kernel corresponding to this position. , is represented as: ; In the formula, It is a non-linear activation function. This is a shape rearrangement operation.

[0060] Based on spatial weight kernel For the input feature map, using row coordinates Column coordinates Centered on and with a neighborhood radius of The neighborhood features are aggregated based on location correlation to obtain the slope aspect deformation feature map. , is represented as: ; In the formula, For channel index, and This is the neighborhood offset.

[0061] S135: Input the slope deformation feature map into the feature pyramid network and the candidate region generation network to filter out abnormal candidate regions; Specifically, the aspect deformation feature map is input into the feature pyramid network, and multi-scale aspect deformation feature maps are generated according to the feature fusion method from high to low levels; for the first... Each scale, based on the slope aspect deformation feature map corresponding to that scale. Multi-scale slope aspect deformation characteristic map of the previous high-level scale , generate the first Multi-scale slope deformation characteristic map corresponding to each scale , is represented as: ; In the formula, This is a one-to-one convolution operation. This is for upsampling operations.

[0062] The multi-scale slope deformation feature map is input into the candidate region generation network, and the candidate region generation network performs [the following] on the [the following]... The feature vector corresponding to each anchor frame Calculate the confidence level of anomalies , is represented as: ; In the formula, and Generate weights and biases obtained from network training for candidate regions. This is the Sigmoid function.

[0063] Candidate region generation network for the first Calculate the bounding box offset for each anchor frame. , is represented as: ; in, , , and These are the horizontal offset, vertical offset, width offset, and height offset of the anchor frame center, respectively.

[0064] Anomaly candidate regions are generated based on anomaly confidence and bounding box offset; the anomaly confidence is compared with a preset anomaly confidence threshold, and when the anomaly confidence is greater than or equal to the anomaly confidence threshold, the corresponding anomaly candidate region is retained; the anomaly confidence threshold is determined through experimental data, and in this embodiment, the anomaly confidence threshold is set to 0.5.

[0065] S136: Input the abnormal candidate region into the deformable pooling module to obtain the abnormal candidate region features; Specifically, the bounding box position of the anomaly candidate region is read, and the initial sampling point corresponding to the anomaly candidate region is determined on the slope deformation feature map; Deformable pooling module according to the first Local eigenvectors of the initial sampling points in the slope deformation feature map Weights completed through training and bias Generate sampling offset , is represented as: ; The initial sampling points within the anomaly candidate region are corrected based on the sampling offset, and then weighted pooling is performed on the corrected sampling points to obtain the features of the anomaly candidate region. , is represented as: ; In the formula, For the features of the abnormal candidate region in the first... Pooling unit, channel index eigenvalues ​​at that location This represents the number of sampling points within the anomaly candidate region. For the first The pooling unit corresponds to the first... Pooling weights for each sampling point For the first The first pooling unit One initial sampling point.

[0066] S137: The abnormal candidate region features are processed by classification and bounding box regression branches and mask segmentation branches to generate deformation anomaly segmentation results including pixel-level masks and bounding boxes.

[0067] Specifically, the feature vector is obtained by flattening the features of the anomaly candidate region. The feature vector is input into the classification and bounding box regression branch, and the output is an anomaly class probability vector. and bounding box correction , is represented as: ; ; In the formula, and The weights and biases obtained from training the classification branch. and The weights and biases obtained from training the bounding box regression branch. It is a normalized exponential function.

[0068] Input the features of the abnormal candidate regions into the mask segmentation branch, and output a pixel-level mask probability map. , is represented as: ; In the formula, This refers to the convolution operation in the mask-based branching process.

[0069] Compare the pixel-level mask probability map with a preset mask probability threshold. Perform comparison to generate the first The pixel position is at the Pixel-level mask values ​​corresponding to each reference imaging time , is represented as: ; The mask probability threshold was determined through experimental data. In this embodiment, the mask probability threshold is set to 0.5.

[0070] The anomaly category identifier, bounding box, pixel-level mask, pixel position, and reference imaging time are written into the same result record to generate the deformation anomaly segmentation result.

[0071] S14: Generate dynamic early warning results for geological disasters based on the results of deformation anomaly zoning, DEM data, optical image data, and disaster-induced data; Specifically, the steps for generating dynamic early warning results for geological disasters are as follows: S141: Based on the pixel-level mask, bounding box, pixel position and reference imaging time in the deformation anomaly segmentation result, calculate the deformation anomaly area value and deformation anomaly boundary range value, and generate the deformation anomaly intensity value. Specifically, read the first The deformation anomaly region segmentation results are obtained at each reference imaging time, and pixel-level mask values ​​are extracted from the deformation anomaly region segmentation results. , bounding box, pixel position and reference imaging time; Based on the range of pixel locations covered by the bounding box, statistical analysis of the pixel-level mask is performed. Values The number of pixels, combined with the DEM raster size, is used to calculate the deformation anomaly area value, expressed as: ; In the formula, For the first The pixel position is at the The deformation anomaly area value corresponding to each reference imaging time. For the first The range of cell locations covered by the bounding box of a given cell location. For the first The pixel position is at the The pixel-level mask values ​​corresponding to each reference imaging time. This represents the area of ​​a DEM raster cell, which is calculated from the DEM raster size.

[0072] Based on the number of pixels covered by the bounding box and the area of ​​the DEM raster cells, the deformation anomaly boundary range value is calculated and expressed as: ; In the formula, For the first The pixel position is at the The deformation anomaly boundary range value corresponding to each reference imaging time. For the first The number of pixels covered by the bounding box of a given pixel location.

[0073] The deformation anomaly intensity value is calculated based on the area value and the boundary range value of the deformation anomaly, and is expressed as follows: ; In the formula, For the first The pixel position is at the The deformation anomaly intensity value corresponding to each reference imaging time. The zero-constant for preventing deformation anomaly intensity value is used to prevent zeroing.

[0074] S142: Calculate the optical disturbance value based on the texture value and edge value, and calculate the disaster susceptibility score through the deformation anomaly intensity value, slope value and optical disturbance value, and classify the susceptibility level according to the disaster susceptibility score.

[0075] Specifically, the texture values ​​are determined based on the optical spatial resolution and optical grid size. and edge values Mapping to pixel locations; when the spatial resolution or grid size of optical image data and DEM data are different, the texture values ​​and edge values ​​are resampled according to the DEM grid size so that the texture values, edge values ​​and slope values ​​correspond to the same pixel location; Texture values and edge values The normalized texture values ​​were obtained by processing the textures using the range normalization method. and normalized marginal values The normalized values ​​take values ​​in the range [0,1]. The optical perturbation value is calculated based on the normalized texture values ​​and the normalized edge values, and is expressed as follows: ; In the formula, For the first The optical perturbation value corresponding to each pixel position. For texture perturbation weights, The edge perturbation weights, texture perturbation weights, and edge perturbation weights are determined based on historical disaster samples and satisfy the following conditions: ; The slope values ​​were normalized using the range normalization method, resulting in normalized slope values ​​within the range [0,1]. The disaster susceptibility score is calculated using the deformation anomaly intensity value, the normalized slope value, and the optical disturbance value, and is expressed as: ; In the formula, For the first The pixel position is at the The disaster susceptibility score corresponding to each baseline imaging time. For the Sigmoid function, As the weight for the intensity of deformation anomalies, For slope weight, For optical perturbation weights, The disaster susceptibility score bias is obtained by training based on historical disaster samples, including the deformation anomaly intensity weight, slope weight, optical disturbance weight, and disaster susceptibility score bias.

[0076] The disaster susceptibility score is compared with a preset susceptibility level threshold to classify susceptibility levels, which are represented as follows: ; In the formula, For the first The pixel position is at the The susceptibility level corresponding to each baseline imaging time The threshold for the first susceptibility level. The second susceptibility level threshold is defined as the threshold for the first susceptibility level, which is determined using experimental data. Corresponding to low susceptibility level, Corresponding to the level of susceptibility, Corresponding to a high incidence level.

[0077] S143: Normalize and weight the previous effective rainfall, short-term heavy rainfall and cumulative rainfall to generate the inducing factor value; Specifically, the corresponding rainfall acquisition time is read based on the baseline imaging time, and the previous effective rainfall, short-term heavy rainfall, and cumulative rainfall are read based on the rainfall acquisition time; When the rainfall acquisition time is inconsistent with the reference imaging time, the rainfall acquisition time with the smallest time interval from the reference imaging time is read as the corresponding rainfall acquisition time.

[0078] The effective rainfall, short-duration heavy rainfall, and cumulative rainfall were normalized using the range normalization method to obtain the normalized value of the effective rainfall. Normalized value of short-term heavy rainfall and cumulative rainfall normalized value The normalized values ​​of the previous effective rainfall, short-term heavy rainfall and cumulative rainfall are in the range [0,1]. The inducing factor value is generated by weighting the normalized values ​​of previous effective rainfall, short-term heavy rainfall, and cumulative rainfall, and is expressed as follows: ; In the formula, For the first The evoked factor values ​​corresponding to each baseline imaging time. , and The rainfall-induced weights are obtained through training based on historical rainfall samples and historical disaster occurrence records, and satisfy the following conditions: .

[0079] S144: Calculate the warning level by using the susceptibility level and the inducing factor value, and write the warning level, pixel location, reference imaging time and rainfall acquisition time into the same result record to generate dynamic geological disaster warning results; Specifically, read the first The pixel position is at the The susceptibility level value corresponding to each baseline imaging time and read the first Induced factor values ​​corresponding to each baseline imaging time .

[0080] First, the susceptibility level is normalized to obtain the normalized susceptibility level value, which is expressed as: ; In the formula, For the first The pixel position is at the Normalized susceptibility level values ​​corresponding to each baseline imaging time.

[0081] The warning level value is obtained by coupling the normalized value of the susceptibility level and the value of the triggering factor. , is represented as: ; In the formula, For the first The pixel position is at the The warning level values ​​corresponding to each baseline imaging time. The susceptibility level weight is determined based on historical early warning samples, and... .

[0082] The warning level value is compared with the preset warning level threshold to obtain the warning level. , is represented as: ; In the formula, The threshold for the first warning level. The threshold for the second warning level is determined by experimental data; At that time, a low-risk warning is issued. At that time, a medium-risk warning was issued. At that time, a high-risk warning is issued.

[0083] By writing the warning level, pixel location, reference imaging time, and rainfall acquisition time into the same result record, a dynamic geological disaster warning result is generated, represented as follows: ; In the formula, For the first The pixel position is at the Dynamic early warning results of geological disasters corresponding to each baseline imaging time. For the first The rainfall acquisition time corresponding to each baseline imaging time.

[0084] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A geological disaster early warning method based on InSAR and deep learning, characterized in that, The method includes: Collect geological disaster monitoring data, including L-SAR ascending and descending orbit data, DEM data, optical image data and disaster-induced data, perform SBAS-InSAR calculation on L-SAR ascending and descending orbit data, and generate ascending and descending orbit line-of-sight deformation variables; Temporal interpolation registration and spatial same-name pixel registration are performed on the line-of-sight deformation variables of the lifting rail to generate registered line-of-sight deformation variables. Based on the registered line-of-sight deformation variables and DEM data, the two-dimensional deformation field and slope deformation variables are calculated. The aspect deformation is rasterized to obtain an aspect deformation map, which is then input into the MaskRCNN model improved based on the Involution operator and deformable pooling module to generate deformation anomaly segmentation results. Based on the results of deformation anomaly zoning, DEM data, optical image data, and disaster-induced data, dynamic early warning results for geological disasters are generated.

2. The geological disaster early warning method based on InSAR and deep learning according to claim 1, characterized in that, Specifically, the L-SAR ascent and descent orbit data includes ascent SLC imagery, descent SLC imagery, precise orbit recording, ascent imaging time, descent imaging time, and pixel location; the DEM data includes elevation values, slope values, aspect angle, DEM spatial resolution values, and DEM grid size; the optical image data includes texture values, edge values, optical spatial resolution values, and optical grid size; and the disaster-induced data includes previous effective rainfall, short-duration heavy rainfall, cumulative rainfall, and rainfall acquisition time.

3. The geological disaster early warning method based on InSAR and deep learning according to claim 2, characterized in that, The steps to generate the line-of-sight deformation of the ascending and descending rails are as follows: Based on the ascending SLC image, descending SLC image, ascending imaging time, descending imaging time, and precise track record, the ascending time baseline value, ascending space baseline value, descending time baseline value, and descending space baseline value are calculated respectively, and the ascending and descending baseline value records are generated. By recording the baseline values ​​of the ascending and descending rails, filter the ascending rail small baseline image pairs and the descending rail small baseline image pairs; Based on the precise orbit records, differential interferometry calculations were performed on the ascending orbit small baseline image pairs and the descending orbit small baseline image pairs to generate ascending orbit differential interferometry records and descending orbit differential interferometry records. Phase unwrapping and time-series inversion were performed on the ascending and descending differential interferometric records, respectively, to obtain the ascending and descending line-of-sight deformation variables. The ascending and descending line-of-sight deformation variables were then generated according to the pixel positions.

4. A geological disaster early warning method based on InSAR and deep learning according to claim 3, characterized in that, The steps for generating the registration line-of-sight deformation are as follows: Based on the deformation along the line of sight to the ascending orbit and the imaging time of the ascending orbit, the time series of the ascending orbit deformation is obtained; and based on the deformation along the line of sight to the descending orbit and the imaging time of the descending orbit, the time series of the descending orbit deformation is obtained. The times of ascending orbit imaging and descending orbit imaging are merged in chronological order to generate a reference imaging timetable; Input the time series of ascending orbit deformation, the time series of descending orbit deformation, and the reference imaging timetable into the spline interpolation function to obtain the time-registered line-of-sight deformation; Based on the pixel position, the time registration line-of-sight deformation variable is registered with the same pixel to generate the registration line-of-sight deformation variable.

5. A geological disaster early warning method based on InSAR and deep learning according to claim 4, characterized in that, The steps for calculating the two-dimensional deformation field and aspect deformation are as follows: Based on the registration line of sight deformation variables, separate the ascending track projection deformation variables and descending track projection deformation variables, and generate ascending and descending track projection deformation records; By using precise orbit recordings and pixel positions, the radar incident angle and radar observation direction values ​​are calculated to generate a radar observation geometric record. Substitute the projection deformation records of the ascending and descending orbits and the geometric records of radar observations into the preset projection equations of the ascending and descending orbits to solve for the east-west deformation components and the vertical deformation components. The east-west and vertical deformation components are arranged according to the pixel position to construct a two-dimensional deformation field. The two-dimensional deformation field is then projected onto the downward slope direction based on the aspect angle and slope value to obtain the two-dimensional deformation field and aspect deformation.

6. A geological disaster early warning method based on InSAR and deep learning according to claim 5, characterized in that, The steps to obtain the aspect deformation diagram are as follows: Based on the aspect deformation and pixel location, generate aspect deformation pixel records; Raster cells are created according to the pixel positions, and the aspect deformation is written into the raster cells to generate an aspect deformation raster map. Normalize the values ​​of the slope aspect deformation raster map to obtain the slope aspect deformation map.

7. A geological disaster early warning method based on InSAR and deep learning according to claim 6, characterized in that, The Mask R-CNN model, improved based on the Involution operator and deformable pooling module, includes a backbone feature extraction network, a feature pyramid network, a candidate region generation network, a deformable pooling module, a classification and bounding box regression branch, and a mask segmentation branch.

8. A geological disaster early warning method based on InSAR and deep learning according to claim 7, characterized in that, The training loss of the Mask R-CNN model improved based on the Involution operator and deformable pooling module is: ； In the formula, For model training loss, For classifying losses, For bounding box regression loss, For mask segmentation loss, For classification loss weights, Weights for bounding box regression loss. The weights are used for mask segmentation loss.

9. A geological disaster early warning method based on InSAR and deep learning according to claim 8, characterized in that, The logic for generating the deformation anomaly segmentation results is as follows: The aspect deformation map is input into the backbone feature extraction network, and the aspect deformation feature map is extracted by the Involution operator. The slope aspect deformation feature map is input into the feature pyramid network and the candidate region generation network to filter out abnormal candidate regions. Input the abnormal candidate region into the deformable pooling module to obtain the abnormal candidate region features; By processing the features of abnormal candidate regions through classification and bounding box regression branches and mask segmentation branches, deformation anomaly segmentation results including pixel-level masks and bounding boxes are generated.

10. A geological disaster early warning method based on InSAR and deep learning according to claim 9, characterized in that, The steps for generating dynamic early warning results for geological disasters are as follows: Based on the pixel-level mask, bounding box, pixel position and reference imaging time in the deformation anomaly segmentation results, the deformation anomaly area value and deformation anomaly boundary range value are statistically analyzed to generate the deformation anomaly intensity value. Optical disturbance values ​​are calculated based on texture and edge values, and disaster susceptibility scores are calculated using deformation anomaly intensity, slope, and optical disturbance values. Susceptibility levels are then determined based on the disaster susceptibility scores. The effective rainfall in the early stage, the short-term heavy rainfall, and the cumulative rainfall are normalized and weighted to generate the inducing factor value; The warning level is calculated by using the susceptibility level and the inducing factor value. The warning level, pixel location, reference imaging time and rainfall acquisition time are written into the same result record to generate dynamic early warning results for geological disasters.

Images