Deep Learning-Based Intelligent Recognition and Risk Warning Method and System for Map Patches

Abstract

The present invention discloses a method and system for intelligent identification and risk warning of map patches based on deep learning, which relates to the technical field of intelligent identification and risk warning of map patches. The present invention includes a map patch intelligent identification module, a change risk index evaluation module, a map patch correlation analysis module, and a risk warning module; the map patch intelligent identification module is used to intelligently identify the map patch objects and the boundaries of each map patch object in each three-dimensional remote sensing image to be analyzed; the change risk index evaluation module is used to evaluate the change risk index of each map patch object; the map patch correlation analysis module is used to find the associated map patch objects of the target map patch object; the risk warning module is used to analyze the risk warning levels of each map patch object. The present invention simplifies the processing complexity of map patch data to a certain extent based on map patch change parameters, and realizes global risk warning of map patch objects.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent identification and risk warning of map patches, and specifically to a method and system for intelligent identification and risk warning of map patches based on deep learning. Background Technique

[0002] The generation of map patches relies on satellite remote sensing technology. By regularly taking high-precision photos of the same area and performing overlay analysis on the image data in combination with the Geographic Information System (GIS), it is possible to accurately identify the areas where the surface coverage has changed (such as cultivated land being converted into construction land, vegetation degradation, etc.). These changed areas are outlined as independent units and marked as map patches on the topographic map. These map patches may represent changes in land use types, soil erosion, etc., and may thus trigger geological disaster risk warnings. Currently, it takes a long time to manually conduct on-site inspections of the soil type changes of each map patch regularly, and different map patches may be inspected by different investigators, resulting in the inability to perform correlation analysis on different map patch objects, which is not conducive to the accurate identification of map patch risks by investigators. At the same time, if there are new changes in the land or adjustments in its application during the investigation process, it is necessary to re-collect and investigate data, which not only increases the workload of processing but also has low timeliness. Summary of the Invention

[0003] The purpose of the present invention is to provide a method and system for intelligent identification and risk warning of map patches based on deep learning to solve the problems raised in the prior art.

[0004] To achieve the above purpose, the present invention provides the following technical solution: An intelligent map patch identification and risk warning system based on deep learning, the system includes a map patch intelligent identification module, a change risk index evaluation module, a map patch correlation analysis module, and a risk warning module;

[0005] The map patch intelligent identification module periodically takes pictures of the target area through satellite remote sensing technology, processes the multi-source data of the target area obtained at different times and performs 3D modeling to obtain a number of 3D remote sensing images, uses the watershed algorithm to perform raster segmentation on the 3D remote sensing images to obtain a number of 3D remote sensing images to be analyzed, and intelligently identifies the map patch objects and the boundaries of each map patch object in each 3D remote sensing image to be analyzed;

[0006] The change risk index evaluation module analyzes the boundary change data of each map patch object according to the boundary information of each map patch object, and evaluates the change risk index of each map patch object based on the abnormal boundary change data;

[0007] The patch correlation analysis module determines the real-time change types of each patch object based on the abnormal boundary change data of each patch object, collects the target patch object, and searches for the associated patch objects of the target patch object in combination with the coincidence of the abnormal trend between the target patch object and its adjacent patch objects;

[0008] The risk warning module is used to analyze the risk warning levels of each patch object and determine the risk warning times of each patch object according to the analysis results.

[0009] Furthermore, the patch intelligent recognition module includes a three-dimensional remote sensing image construction unit, a grid segmentation unit, and an intelligent recognition unit;

[0010] The three-dimensional remote sensing image construction unit acquires multi-source data of the target area at different times through satellite remote sensing technology. The multi-source data includes satellite remote sensing images, lidar point clouds, ground control points, and digital elevation models, and obtains three-dimensional remote sensing images of the target area at different times based on the multi-source data;

[0011] The grid segmentation unit uses the watershed algorithm to segment the three-dimensional remote sensing image into several three-dimensional remote sensing images to be analyzed;

[0012] The intelligent recognition unit takes the catchment basins in the three-dimensional remote sensing image as patch objects and the watersheds in the three-dimensional remote sensing image as the boundaries of the patch objects.

[0013] Furthermore, the change risk index evaluation module includes a boundary change data acquisition unit, a boundary change vector data analysis unit, and a change risk index evaluation unit;

[0014] The boundary change data acquisition unit is used to collect the acquisition time of the three-dimensional remote sensing image to be analyzed, and collect the acquisition time of the initial remote sensing image of the target area, sort the collected acquisition times in chronological order, and based on the sorting result, compare the three-dimensional remote sensing images obtained at two adjacent acquisition times, and based on the comparison result, analyze the abnormal boundary change data of each patch object and generate a boundary change data set for each patch object;

[0015] The boundary change vector data analysis unit determines whether each abnormal boundary change data stored in the boundary change data set of the selected patch object meets the screening conditions. The screening conditions are that the values of the change direction angle, change volume, and change scale of the patch object are all 0. If it meets, the corresponding abnormal boundary change data is removed from the boundary change data set of the selected patch object. If it does not meet, the corresponding abnormal boundary change data is retained in the boundary change data set of the selected patch object. After traversing all patch objects, an abnormal boundary change data set for each patch object is obtained;

[0016] The change risk index evaluation unit evaluates the real-time change risk index of each patch object according to the continuity of each abnormal boundary change data stored in the abnormal boundary change data set of each patch object in terms of time and the degree of abnormal movement in terms of time.

[0017] Furthermore, the specific method for the abnormal boundary change data acquisition unit to generate the abnormal boundary change data set of each patch object is as follows:

[0018] Number each three-dimensional remote sensing image to be analyzed in the order of the acquisition time, and the numbering result is: j = 1, 2, …, n; n represents the total number of three-dimensional remote sensing image maps to be analyzed;

[0019] Compare the three-dimensional remote sensing image to be analyzed 1 with the initial remote sensing image, and compare the three-dimensional remote sensing image j with the three-dimensional remote sensing image j - 1 to respectively obtain the change direction angle, change volume, and change scale corresponding to the patch object numbered i in the three-dimensional remote sensing images to be analyzed 1 and j, where i = 1, 2, …, m, representing the numbers corresponding to each patch object in the initial remote sensing image, and m represents the total number of patch objects;

[0020] Perform image registration processing on each three-dimensional remote sensing image to be analyzed with the initial remote sensing image, and randomly select an alignment point A based on the registration processing result j , taking the alignment point A j as the coordinate origin to construct a three-dimensional space coordinate system j and a three-dimensional space coordinate system p respectively;

[0021] Obtain the central coordinate U ip = (X ip , Y ip , Z ip ) of the patch object i in the three-dimensional space coordinate system p, calculate the distance value d ip between each boundary point of the patch object i in the three-dimensional space coordinate system j and the coordinate U ipj according to the distance formula between two points in space, and obtain the boundary point coordinate B ipj = (X ipj , Y ipj , Z ipj ) corresponding to the maximum value of d ipj . Calculate the included angle between the vector and the vector through the included angle formula of three-dimensional space vectors to obtain the change direction angle β ipj of the patch object i in the three-dimensional remote sensing image to be analyzed j, where b represents a constant greater than 0;

[0022] Calculate the difference between the area volume corresponding to the patch object i in the remotely sensed image to be analyzed and the area volume corresponding to the patch object i in the initial remotely sensed image, and obtain the change volume S of the patch object i in the remotely sensed image to be analyzed ipj ;

[0023] Denote the minimum value of d ipj as the change scale C of the patch object i in the remotely sensed image to be analyzed ipj ;

[0024] The boundary change data set M of the patch object i i ={(β ip1 , S ip1 , C ip1 ),…,(β ipj , S ipj , C ipj )}.

[0025] Furthermore, the specific method for the change risk index evaluation unit to evaluate the real-time change risk index of each patch object is as follows:

[0026] Let the abnormal boundary change data set of the patch object i be M´ i , calculate the difference S i between the change volume stored at the v + 1 position in M´ i and the change volume stored at the v position in M´ i(v+1→v) , calculate the ratio f i(v+1→v) between S i(v+1→v) and t, and obtain the abnormal degree set N i of the patch object i with respect to the change volume, N i ={f i(1+1→1) ,…, f i(v+1→v)}, calculate the first abnormal degree of the patch object i according to W i =P´ i / P i , P i represents the overall average value of N i , P´ i represents the overall standard deviation of N i , t represents the shooting interval time of the satellite remote sensing technology for the target area, v = 1, 2,…, V, indicating the numbering of each element stored in M´ i in the order from left to right, and V represents the total number of elements stored in M´ i ;

[0027] Calculate the second abnormal degree W´ i of the patch object i according to the calculation method of the first abnormal degree;

[0028] For M´i The transition direction angle stored in the v+1th position and M´ i The difference β between the transition direction angles stored at the vth position in i(v+1→v) Calculate, if β i(v+1→v) =0, then according to β i(v+1→v) The degree of discreteness in time is used to obtain the first influence coefficient h of the third degree of abnormality of the image spot object i. i(v+1→v) , if β i(v+1→v) ≠0, then for β i(v+1→v) The difference between t and t is calculated to obtain the second influence coefficient h' of the third abnormality degree of the image spot object i. i(v+1→v) , for h´ i(v+1→v) and 1+h i(v+1→v) The product f´´ i(v+1→v) Calculate and get the degree of change of the image spot object i with respect to the change direction angle N´´ i , N´´ i ={f´´ i(1+1→1) ,…,f´´ i(v+1→v)}, according to W´´ i =R´ i / R i Calculate the second abnormality degree of the image spot object i, R i Indicates N´´ i The overall mean, R´ i Indicates N´´ i The population standard deviation;

[0029] According to E i =g1×|W i |+g2×|W´ i |+g3×|W´´ i |Predict the change risk index of the patch object i at the current moment, where g1, g2, and g3 represent the weight coefficients corresponding to the change volume, change scale, and change direction angle, respectively.

[0030] Furthermore, the pattern association analysis module includes a change type judgment unit, a change coincidence analysis unit and a association analysis unit;

[0031] The change type judgment unit judges whether the change direction angle of the spot object i is within the range of [η,π-η]. If not, it indicates that the change type of the spot object i is soil erosion. If it is, when the change rate of the change scale and the change volume of the spot object i are the same, the change type of the spot object i is judged to be vegetation cover. When the change rate of the change scale and the change volume of the spot object i are different, the change type of the spot object i is judged to be land use type change, wherein η represents an angle and 0°≤η≤30°, π=180°;

[0032] The abnormality coincidence analysis unit collects patch objects whose change type is soil erosion, randomly selects a patch object from the collected patch objects, searches for adjacent patch objects of the selected patch object, wherein the adjacent patch objects refer to the patch objects whose shortest distance between the patch object to be analyzed and the selected patch object is within a set range and the line connecting any points on the two patch objects does not pass through other patch objects, and predicts the abnormality coincidence coefficient between the selected patch object and the adjacent patch objects based on the coincidence of the change direction angle and the change scale between the selected patch object and the adjacent patch objects;

[0033] When the coefficient of the difference overlap between the selected pattern object and the adjacent pattern object is greater than the selected threshold, the association analysis unit uses the adjacent pattern object as the associated pattern object of the selected pattern object; when the coefficient of the difference overlap between the selected pattern object and the adjacent pattern object is less than or equal to the selected threshold, the adjacent pattern object is not used as the associated pattern object of the selected pattern object, and all adjacent pattern objects of the selected pattern object are traversed to search for the associated pattern object of the selected pattern object.

[0034] Furthermore, the specific method for the said abnormality coincidence analysis unit to predict the abnormality coincidence coefficient between the selected pattern object and the adjacent pattern objects is:

[0035] The difference S between the change scale of the selected spot object at time L and the change scale of the selected spot object at time Lt L-t→L Calculate and calculate exp(S L-t→L ) and exp(S L-2×t→L-t ) L-2×t→L Calculate and get the abnormal trend of the selected spot object in the [L-2×t,L] time period. If s L-2×t→L ≥1, it means that the abnormal trend of the selected spot object in the [L-2×t,L] time period is positive. If s L-2×t→L ≥1, it means that the abnormal trend of the selected spot object in the [L-2×t,L] time period is positive. If 0<s L-2×t→L <1, it means that the abnormal trend of the selected spot object in the time period [L-2×t,L] is a reverse abnormal trend, where exp() represents an exponential function with e as the base and e=2.73, and L represents real time;

[0036] Judge whether the real-time change trend of the selected patch object and the adjacent patch object is the same during the time period [L0, L]. Calculate the first change image coefficient of the selected patch object and the adjacent patch object according to D1 = [H / (n - 1)]×exp(-F), where H represents the number of times that the real-time change trends of the selected patch object and the adjacent patch object are the same during the time period [L0, L], F represents the degree of dispersion of the same change trend in time, and L0 represents the initial shooting time of the target area;

[0037] Judge whether the terminal sides of the change direction angles of the selected patch object and the adjacent patch object intersect. If they do not intersect, the second change image coefficient D2 of the selected patch object and the adjacent patch object is 0. If they intersect, the second change image coefficient D2 of the selected patch object and the adjacent patch object is 1 - exp(-γ), where γ represents the included angle value corresponding to the intersection of the terminal sides of the change direction angles of the selected patch object and the adjacent patch object;

[0038] Predict the change coincidence coefficient between the selected patch object and the adjacent patch object according to K = λ1×D1 + λ2×D2, where both λ1 and λ2 represent proportionality coefficients and λ1 + λ2 = 1.

[0039] Further, the risk warning module includes a risk warning level analysis unit and a risk warning time determination unit;

[0040] The risk warning level analysis unit collects the total number w of the associated patch objects of the selected patch object, and determines the risk warning level of each patch object in combination with the shortest distance value between the selected patch object and each associated patch object;

[0041] The risk warning time determination unit determines the risk warning time of each patch object according to the risk warning level of each patch object. The risk warning time = L + (ε - δ´)×t, where ε represents the number corresponding to the risk warning level of the patch object.

[0042] Further, the specific method for the risk warning level analysis unit to determine the risk warning level of each patch object is as follows:

[0043] The shortest distance value between the selected patch object and each associated patch object is taken as the base number, and the angle between the terminal side of the change direction angle of the selected patch object and the terminal side of the change direction angle of each associated patch object is taken as the exponent to construct an exponential function, and the correlation coefficient between the selected patch object and each associated patch object is calculated. If the terminal side of the change direction angle of the selected patch object does not intersect with the terminal side of the change direction angle of each associated patch object, the angle is considered to be 0, and the mean value δ of the correlation coefficient is mapped to the range of 0 to 1 to obtain the mapping value δ´. δ´ is used as the risk warning coefficient of the selected patch object, and the risk warning level of the selected patch object is analyzed;

[0044] The risk warning coefficient δ´ of each patch object other than the selected patch object is 0;

[0045] If 0.8 ≤ δ´ ≤ 1, it is analyzed that the risk warning level of the selected patch object is a first-level warning;

[0046] If 0.5 ≤ δ´ < 0.8, it is analyzed that the risk warning level of the selected patch object is a second-level warning;

[0047] If 0.3 ≤ δ´ < 0.5, it is analyzed that the risk warning level of the selected patch object is a third-level warning;

[0048] If 0 ≤ δ´ < 0.3, it is analyzed that the risk warning level of the selected patch object is a fourth-level warning.

[0049] A method for intelligent recognition and risk warning of patches based on deep learning, the method includes:

[0050] S10: Periodically photograph the target area through satellite remote sensing technology, process the multi-source data of the target area obtained in different periods and perform 3D modeling to obtain several 3D remote sensing images. Use the watershed algorithm to perform raster segmentation on the 3D remote sensing images to obtain several 3D remote sensing images to be analyzed, and intelligently identify the patch objects and the boundaries of each patch object in each 3D remote sensing image to be analyzed;

[0051] S20: Analyze the boundary change data of each patch object according to the boundary information of each patch object, and evaluate the change risk index of each patch object based on the abnormal boundary change data;

[0052] S30: Determine the real-time change type of each patch object according to the abnormal boundary change data of each patch object, and collect the target patch object. Based on the coincidence situation of the abnormal trend between the target patch object and the adjacent patch objects of the target patch object, find the associated patch objects of the target patch object;

[0053] S40: Analyze the risk warning levels of each patch object, and determine the risk warning time of each patch object according to the analysis results.

[0054] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0055] 1. The present invention constructs three-dimensional remote sensing images of the target area at different times through remote sensing image technology, and realizes the intelligent recognition of patches based on the three-dimensional remote sensing images, without the need for manual on-site investigation of the target area, improving the acquisition efficiency of patch data.

[0056] 2. The present invention analyzes the abnormal boundary change data of patch objects from the change direction angle, change volume and change scale of patch objects, simplifies the processing complexity of patch data to a certain extent, and combines the abnormal degree of change of the abnormal boundary change data in time to analyze the change risk situation of each patch object in real time, improving the analysis accuracy.

[0057] 3. The present invention searches for associated patch objects of a patch object according to the abnormal coincidence situation of adjacent patch objects, and based on the search results and the shortest distance values of the patch object from each adjacent patch object, realizes the global risk warning of the patch object. Description of the Drawings

[0058] Figure 1 It is a schematic working flow diagram of the method for intelligent patch recognition and risk warning based on deep learning of the present invention. Detailed Embodiments

[0059] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0060] As Figure 1 shown, the present invention provides a technical solution for a method and system for intelligent patch recognition and risk warning based on deep learning. The system for intelligent patch recognition and risk warning based on deep learning includes a patch intelligent recognition module, a change risk index evaluation module, a patch association analysis module and a risk warning module;

[0061] The patch intelligent recognition module periodically photographs the target area through satellite remote sensing technology, processes the multi-source data of the target area obtained at different times, and performs 3D modeling to obtain several 3D remote sensing images. The watershed algorithm is used to rasterize the 3D remote sensing images to obtain several 3D remote sensing images to be analyzed. The patch objects and the boundaries of each patch object in each 3D remote sensing image to be analyzed are intelligently recognized. The watershed algorithm considers the segmentation of the image according to the composition of the watershed.

[0062] The patch intelligent recognition module includes a 3D remote sensing image construction unit, a raster segmentation unit, and an intelligent recognition unit.

[0063] The 3D remote sensing image construction unit obtains the multi-source data of the target area at different times through satellite remote sensing technology. The multi-source data includes satellite remote sensing images, lidar point clouds, ground control points, and digital elevation models. The digital elevation model is obtained based on satellite remote sensing images. The specific obtaining method belongs to the prior art. Based on the multi-source data, 3D remote sensing images of the target area at different times are obtained. The specific method for obtaining 3D remote sensing images of the target area based on multi-source data belongs to the prior art. The specific method steps are: multi-source data integration → lidar data optimization → ground object classification and contour extraction → sparse and dense reconstruction → terrain and building modeling → texture mapping and fusion → data fusion and accuracy improvement → dynamic scene integration → visualization and decision support.

[0064] The raster segmentation unit rasterizes the 3D remote sensing images using the watershed algorithm to obtain several 3D remote sensing images to be analyzed.

[0065] The intelligent recognition unit takes the catchment basins in the 3D remote sensing images as patch objects and the watersheds in the 3D remote sensing images as the boundaries of the patch objects.

[0066] The change risk index evaluation module analyzes the boundary change data of each patch object based on the boundary information of each patch object, and evaluates the change risk index of each patch object based on the abnormal boundary change data.

[0067] The change risk index evaluation module includes a boundary change data acquisition unit, a boundary change vector data analysis unit, and a change risk index evaluation unit.

[0068] The boundary change data acquisition unit is used to collect the acquisition time of the three-dimensional remote sensing image to be analyzed, and collect the acquisition time of the initial remote sensing image of the target area. The collected acquisition times are sorted in chronological order. Based on the sorting result, the three-dimensional remote sensing images obtained at two adjacent acquisition times are compared. The three-dimensional remote sensing images include the three-dimensional remote sensing image to be analyzed and the initial remote sensing image of the target area. Based on the comparison result, the abnormal boundary change data of each patch object is analyzed, and a boundary change data set of each patch object is generated. The specific method is as follows:

[0069] The three-dimensional remote sensing images to be analyzed are numbered in the order of their acquisition times. The numbering result is: j = 1, 2, …, n; n represents the total number of three-dimensional remote sensing image maps to be analyzed;

[0070] The three-dimensional remote sensing image 1 is compared with the initial remote sensing image, and the three-dimensional remote sensing image j is compared with the three-dimensional remote sensing image j - 1. The change direction angle, change volume, and change scale corresponding to the patch object numbered i in the three-dimensional remote sensing images 1 and j are obtained respectively, where i = 1, 2, …, m, representing the numbers corresponding to each patch object in the initial remote sensing image, and m represents the total number of patch objects;

[0071] The three-dimensional remote sensing images to be analyzed are respectively subjected to image registration processing with the initial remote sensing image. Image registration is a technology for aligning different images taken of the same scene, that is, finding the point-to-point mapping relationship between the images, or establishing an association for a certain feature of interest. Based on the registration processing result, an alignment point A is randomly selected j , and the alignment point refers to a random point in the overlapping area after the image registration processing of the three-dimensional remote sensing image to be analyzed and the initial remote sensing image. Taking the alignment point A j as the coordinate origin, a three-dimensional space coordinate system j and a three-dimensional space coordinate system p are respectively constructed;

[0072] In the three-dimensional space coordinate system p, the central coordinate U ip = (X ip , Y ip , Z ip ) of the patch object i is obtained. According to the distance formula between two points in space, the distance value d ip between each boundary point of the patch object i in the three-dimensional space coordinate system j and the coordinate U ipj is calculated. The boundary point coordinate B ipj corresponding to the maximum value of d ipj = (X ipj , Y ipj , Z ipj ) is obtained. Through the angle formula of three-dimensional space vectors for the vector and the vector Calculate the included angle to obtain the change direction angle β of the patch object i in the remotely sensed image j to be analyzed ipj , where b represents a constant greater than 0;

[0073] Calculate the difference between the regional volume corresponding to the patch object i in the remotely sensed image j to be analyzed and the regional volume corresponding to the patch object i in the initial remotely sensed image, and obtain the change volume S of the patch object i in the remotely sensed image j to be analyzed ipj ;

[0074] Denote d ipj Take the minimum value as the change scale C of the patch object i in the remotely sensed image j to be analyzed ipj ;

[0075] The boundary change dataset M of the patch object i i ={(β ip1 , S ip1 , C ip1 ),…,(β ipj , S ipj , C ipj )};

[0076] The boundary change vector data analysis unit determines whether each abnormal boundary change data stored in the boundary change dataset of the selected patch object meets the screening conditions. The screening conditions are: the values of β ipj , S ipj , and C ipj are all 0. If satisfied, the corresponding abnormal boundary change data is removed from the boundary change dataset of the selected patch object. If not satisfied, the corresponding abnormal boundary change data is retained in the boundary change dataset of the selected patch object. Traverse all patch objects to obtain the abnormal boundary change datasets of each patch object. The selected patch object refers to a randomly selected patch object;

[0077] The change risk index evaluation unit evaluates the real-time change risk index of each patch object according to the continuity in time and the degree of abnormal movement in time of each abnormal boundary change data stored in the abnormal boundary change dataset of each patch object. The specific method is as follows:

[0078] Let the abnormal boundary change dataset of the patch object i be M´ i , calculate the difference S i between the change volume stored in the (v + 1)-th position in M´ i and the change volume stored in the v-th position in M´ i(v+1→v) , calculate the ratio f i(v+1→v) between S i(v+1→v) and t, and obtain the set N i of the degree of abnormal movement of the patch object i with respect to the change volume, Ni ={f i(1+1→1) , …, f i(v+1→v)}, according to W i =P´ i / P i calculate the first degree of change of the patch object i, where P i represents the overall average of N i , P´ i represents the overall standard deviation of N i , t represents the shooting interval time of the satellite remote sensing technology for the target area, v = 1, 2, …, V, representing the numbering of each element stored in M´ i in the order from left to right, and V represents the total number of elements stored in M´ i ;

[0079] Calculate the difference C i between the change scale stored in the (v + 1)-th position in M´ i and the change scale stored in the v-th position in M´ i(v+1→v) , calculate the ratio f´ i(v+1→v) between C i(v+1→v) and t, and obtain the change degree set N´ i of the patch object i with respect to the change scale, N´ i ={f´ i(1+1→1) , …, f´ i(v+1→v)}, according to W´ i =Q´ i / Q i calculate the second degree of change of the patch object i, where Q i represents the overall average of N´ i , and Q´ i represents the overall standard deviation of N´ i ;

[0080] Calculate the difference β i between the change direction angle stored in the (v + 1)-th position in M´ i and the change direction angle stored in the v-th position in M´ i(v+1→v) . If β i(v+1→v) = 0, then according to the discrete degree of β i(v+1→v) in time, obtain the first influence coefficient h i(v+1→v) of the third degree of change of the patch object i. If β i(v+1→v) ≠ 0, then calculate the difference between β i(v+1→v) and t to obtain the second influence coefficient h´ i(v+1→v) of the third degree of change of the patch object i, and calculate the product f´´ i(v+1→v) between h´ i(v+1→v) and 1 + h i(v+1→v)Perform calculations to obtain the degree of change set N´´ of the patch object i with respect to the change direction angle i , N´´ i ={f´´ i(1+1→1) ,…, f´´ i(v+1→v)}, and calculate the second degree of change of the patch object i according to W´´ i =R´ i / R i . Here, R i represents the overall average value of N´´ i , and R´ i represents the overall standard deviation of N´´ i ;

[0081] Predict the change risk index of the patch object i at the current moment according to E i =g1×|W i |+g2×|W´ i |+g3×|W´´ i |. Among them, g1, g2, and g3 respectively represent the weight coefficients corresponding to the change volume, change scale, and change direction angle;

[0082] The patch association analysis module determines the real-time change type of each patch object according to the abnormal boundary change data of each patch object, and collects the target patch object. Based on the coincidence of the change trend between the target patch object and its adjacent patch objects, find the associated patch objects of the target patch object;

[0083] The patch association analysis module includes a change type judgment unit, a change coincidence analysis unit, and an association analysis unit;

[0084] The change type judgment unit judges whether the change direction angle of the patch object i is within the range of [η, π - η]. If not, it means that the change type of the patch object i is soil erosion. If so, when the change rate of the change scale and change volume of the patch object i is the same, judge that the change type of the patch object i is vegetation coverage. When the change rate of the change scale and change volume of the patch object i is different, judge that the change type of the patch object i is land use type change. For example: cultivated land is converted into construction land and there are construction traces on the cultivated land. Among them, η represents the angle degree and 0° ≤ η ≤ 30°, and π = 180°;

[0085] The change coincidence analysis unit collects patch objects with the change type of soil erosion. A patch object is randomly selected from the collected patch objects, and adjacent patch objects of the selected patch object are searched for. The adjacent patch object refers to the shortest distance between the patch object to be analyzed and the selected patch object within a set range, and the line connecting any point between the two patch objects does not pass through other patch objects. Other patch objects refer to patch objects other than the selected patch object and the patch object to be analyzed. The patch object to be analyzed refers to a randomly selected patch object from the patch objects other than the selected patch object among the collected patch objects. According to the coincidence of the selected patch object and the adjacent patch object with respect to the change direction angle and the change scale, the change coincidence coefficient between the selected patch object and the adjacent patch object is predicted. The specific method is as follows:

[0086] Calculate the difference S between the change scale of the selected patch object at time L and the change scale of the selected patch object at time L - t L-t→L Perform the calculation, and for exp(S L-t→L ) and exp(S L-2×t→L-t ) calculate the ratio s L-2×t→L to obtain the change trend of the selected patch object in the time period [L - 2×t, L]. If s L-2×t→L ≥1, it means that the change trend of the selected patch object in the time period [L - 2×t, L] is a positive change. If s L-2×t→L ≥1, it means that the change trend of the selected patch object in the time period [L - 2×t, L] is a positive change. If 0 < s L-2×t→L <1, it means that the change trend of the selected patch object in the time period [L - 2×t, L] is a negative change. Here, exp() represents the exponential function with base e and e = 2.73, and L represents the real-time time;

[0087] Judge whether the real-time change trends of the selected patch object and the adjacent patch object in the time period [L0, L] are the same. Calculate the first change image coefficient of the selected patch object and the adjacent patch object according to D1 = [H / (n - 1)]×exp(-F), where H represents the number of times the real-time change trends of the selected patch object and the adjacent patch object are the same in the time period [L0, L], F represents the degree of dispersion of the same change trend in time, and L0 represents the initial shooting time of the target area;

[0088] Judge whether the terminal sides of the change direction angles of the selected patch object and the adjacent patch object intersect. The terminal side is a ray. For example, the direction vector The ray where it is located is the terminal side of the patch object i in the three-dimensional space coordinate system j. If they do not intersect, the second movement image coefficient D2 of the patch object and the adjacent patch object is selected as 0. If they intersect, the second movement image coefficient D2 of the patch object and the adjacent patch object is selected as 1 - exp(-γ), where γ represents the included angle value corresponding to the intersection of the terminal side of the change direction angle of the selected patch object and the terminal side of the change direction angle of the adjacent patch object;

[0089] Predict the movement coincidence coefficient between the selected patch object and the adjacent patch object according to K = λ1×D1 + λ2×D2, where λ1 and λ2 both represent proportionality coefficients and λ1 + λ2 = 1;

[0090] When the movement coincidence coefficient between the selected patch object and the adjacent patch object is greater than the selected threshold, the adjacent patch object is taken as the associated patch object of the selected patch object. When the movement coincidence coefficient between the selected patch object and the adjacent patch object is less than or equal to the selected threshold, the adjacent patch object is not taken as the associated patch object of the selected patch object. Traverse all adjacent patch objects of the selected patch object to find the associated patch objects of the selected patch object;

[0091] The risk warning module is used to analyze the risk warning level of each patch object and determine the risk warning time of each patch object according to the analysis result;

[0092] The risk warning module includes a risk warning level analysis unit and a risk warning time determination unit;

[0093] The risk warning level analysis unit collects the total number w of the associated patch objects of the selected patch object, and combines the shortest distance values between the selected patch object and each associated patch object to determine the risk warning level of each patch object. The specific method is as follows:

[0094] Take the shortest distance value between the selected patch object and each associated patch object as the base, and take the included angle between the terminal side of the change direction angle of the selected patch object and the terminal side of the change direction angle of each associated patch object as the exponent to construct an exponential function. The exponential function model is ζ = τ ψ , calculate the correlation coefficient between the selected patch object and each associated patch object. If the terminal side of the change direction angle of the selected patch object and the terminal side of the change direction angle of each associated patch object do not intersect, it is considered that the included angle is 0. Map the mean value δ of the correlation coefficient to the range from 0 to 1 to obtain the mapped value δ´. Take δ´ as the risk warning coefficient of the selected patch object and analyze the risk warning level of the selected patch object;

[0095] The risk warning coefficient δ´ of each patch object other than the selected patch object is 0;

[0096] If 0.8 ≤ δ´ ≤ 1, it is analyzed that the risk warning level of the selected patch object is a first-level warning;

[0097] If 0.5 ≤ δ´ < 0.8, it is analyzed that the risk warning level of the selected patch object is a second-level warning;

[0098] If 0.3 ≤ δ´ < 0.5, it is analyzed that the risk warning level of the selected patch object is a third-level warning;

[0099] If 0 ≤ δ´ < 0.3, it is analyzed that the risk warning level of the selected patch object is a fourth-level warning;

[0100] The risk warning time determination unit determines the risk warning time of each patch object according to the risk warning level of each patch object. The risk warning time = L + (ε - δ´) × t, where ε represents the number corresponding to the risk warning level of the patch object. For example, if the risk warning level of the patch object is a first-level warning, then ε = 1.

[0101] A method for intelligent recognition and risk warning of patches based on deep learning, the method includes:

[0102] S10: Periodically photograph the target area through satellite remote sensing technology, process the multi-source data of the target area obtained at different times and perform 3D modeling to obtain several 3D remote sensing images. Use the watershed algorithm to perform raster segmentation on the 3D remote sensing images to obtain several 3D remote sensing images to be analyzed, and intelligently identify the patch objects and the boundaries of each patch object in each 3D remote sensing image to be analyzed;

[0103] S20: Analyze the boundary change data of each patch object according to the boundary information of each patch object, and evaluate the change risk index of each patch object based on the abnormal boundary change data;

[0104] S30: Determine the real-time change type of each patch object according to the abnormal boundary change data of each patch object, and collect the target patch object. Based on the coincidence of the abnormal trend between the target patch object and its adjacent patch objects, find the associated patch objects of the target patch object;

[0105] S40: Analyze the risk warning level of each patch object, and determine the risk warning time of each patch object according to the analysis result.

[0106] Example 1: Assume that there are associated patch objects J1 and J2 for the selected patch object. Let the shortest distance value between the selected patch object and the associated patch object J1 be τ1, and the shortest distance value between the selected patch object and the associated patch object J2 be τ2. Let the angle between the terminal side of the change direction angle of the selected patch object and the terminal side of the change direction angle of the associated patch object J1 be ψ1, and the angle between the terminal side of the change direction angle of the selected patch object and the terminal side of the change direction angle of the associated patch object J2 be ψ2. According to the exponential function Calculate the correlation coefficient between the selected patch object and the associated patch object J1 according to the exponential function Calculate the correlation coefficient between the selected patch object and the associated patch object J2 according to the exponential function, and map (ζ1 + ζ2) / 2 to the range of 0 to 1 to obtain the risk warning coefficient ζ´ of the selected patch object;

[0107] Assume ζ´ = 0.9. Since 0.8 ≤ ζ´ = 0.9 ≤ 1, it is analyzed that the risk warning level of the selected patch object is a first-level warning, that is, ε = 1;

[0108] Assume t = 7, unit: day, L = 15:30:48 on April 29, 2024. Then the risk warning time of the selected patch object = L + (ε - δ´) × t = 08:18:48 on April 30, 2024;

[0109] Therefore, the risk warning time of the selected patch object is 08:18:48 on April 30, 2024.

[0110] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-restrictive. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present invention. Any reference signs in the claims should not be regarded as limiting the claims involved.

Claims

1. An intelligent recognition and risk warning system for map patches based on deep learning, characterized in that: The system includes a patch intelligent recognition module, a change risk index assessment module, a patch correlation analysis module, and a risk warning module; The patch intelligent recognition module periodically photographs the target area through satellite remote sensing technology, processes the multi-source data of the target area obtained at different times and performs 3D modeling to obtain several 3D remote sensing images, uses the watershed algorithm to perform raster segmentation on the 3D remote sensing images to obtain several 3D remote sensing images to be analyzed, and intelligently identifies the patch objects and the boundaries of each patch object in each 3D remote sensing image to be analyzed; The change risk index assessment module analyzes the boundary change data of each patch object according to the boundary information of each patch object, and evaluates the change risk index of each patch object based on the abnormal boundary change data; The patch correlation analysis module determines the real-time change type of each patch object according to the abnormal boundary change data of each patch object, collects the target patch object, and combines the coincidence situation of the abnormal trend between the target patch object and the adjacent patch objects of the target patch object to find the associated patch objects of the target patch object; The patch correlation analysis module includes a change type judgment unit, an abnormal coincidence analysis unit, and a correlation analysis unit; The change type judgment unit judges whether the change direction angle of the patch object i is within the range of [η, π - η]. If not, it means that the change type of the patch object i is soil erosion. If so, when the change scale of the patch object i is the same as the change rate of the change volume, it is judged that the change type of the patch object i is vegetation coverage. When the change scale of the patch object i is different from the change rate of the change volume, it is judged that the change type of the patch object i is land use type change, where η represents the angle number and 0° ≤ η ≤ 30°, π = 180°, i = 1, 2, …, m, representing the numbers corresponding to each patch object in the initial remote sensing image, and m represents the total number of patch objects; The abnormal coincidence analysis unit collects the patch objects with the change type of soil erosion, randomly selects a patch object from the collected patch objects, finds the adjacent patch objects of the selected patch object. The adjacent patch object means that the shortest distance between the patch object to be analyzed and the selected patch object is within the set range and the line connecting any point on the two patch objects does not pass through other patch objects. According to the coincidence situation of the selected patch object and the adjacent patch object regarding the change direction angle and the change scale, the abnormal coincidence coefficient between the selected patch object and the adjacent patch object is predicted. The specific method is as follows: Calculate the difference S between the change scale of the selected patch object at time L and the change scale of the selected patch object at time L-t L-t→L Calculate the ratio s L-t→L between exp(S L-2×t→L-t ) and exp(S L-2×t→L ), and obtain the abnormal trend of the selected patch object in the time period [L-2×t, L]. If s L-2×t→L ≥1, it means that the abnormal trend of the selected patch object in the time period [L-2×t, L] is a positive abnormal movement. If s L-2×t→L ≥1, it means that the abnormal trend of the selected patch object in the time period [L-2×t, L] is a positive abnormal movement. If 0 < s L-2×t→L < 1, it means that the abnormal trend of the selected patch object in the time period [L-2×t, L] is a reverse abnormal movement. Here, exp() represents the exponential function with base e and e = 2.73, and L represents the real-time time; Judge whether the real-time abnormal trends of the selected patch object and the adjacent patch object are the same within the time period of [L0, L], and calculate the first abnormal image coefficient of the selected patch object and the adjacent patch object according to D1 = [H / (n - 1)] × exp(-F), where H represents the number of times that the real-time abnormal trends of the selected patch object and the adjacent patch object are the same within the time period of [L0, L], F represents the dispersion degree of the same abnormal trend in time, L0 represents the initial shooting time of the target area, and n represents the total number of 3D remote sensing images to be analyzed; Determine whether the terminal sides of the change direction angles of the selected patch object and the adjacent patch object intersect. If they do not intersect, the second change image coefficient D2 between the selected patch object and the adjacent patch object is 0. If they intersect, the second change image coefficient D2 between the selected patch object and the adjacent patch object is 1 - exp(-γ), where γ represents the included angle value corresponding to the intersection of the terminal sides of the change direction angles of the selected patch object and the adjacent patch object; Predict the change coincidence coefficient between the selected patch object and the adjacent patch object according to K = l1×D1 + l2×D2, where l1 and l2 both represent proportionality coefficients and l1 + l2 = 1; When the change coincidence coefficient between the selected patch object and the adjacent patch object is greater than the selected threshold, the associated analysis unit regards the adjacent patch object as the associated patch object of the selected patch object. When the change coincidence coefficient between the selected patch object and the adjacent patch object is less than or equal to the selected threshold, the adjacent patch object is not regarded as the associated patch object of the selected patch object. Traverse all adjacent patch objects of the selected patch object to find the associated patch objects of the selected patch object; The risk warning module is used to analyze the risk warning levels of each patch object and determine the risk warning times of each patch object according to the analysis results.

2. The intelligent patch recognition and risk warning system based on deep learning according to claim 1, characterized in that: The patch intelligent recognition module includes a three-dimensional remote sensing image construction unit, a raster segmentation unit, and an intelligent recognition unit; The three-dimensional remote sensing image construction unit acquires multi-source data of the target area at different times through satellite remote sensing technology. The multi-source data includes satellite remote sensing images, lidar point clouds, ground control points, and digital elevation models, and obtains three-dimensional remote sensing images of the target area at different times based on the multi-source data; The raster segmentation unit uses the watershed algorithm to segment the three-dimensional remote sensing image into several three-dimensional remote sensing images to be analyzed; The intelligent recognition unit regards the catchment basins in the three-dimensional remote sensing image as patch objects and the watersheds in the three-dimensional remote sensing image as the boundaries of the patch objects.

3. The intelligent recognition and risk warning system for patches based on deep learning according to claim 2, wherein: The change risk index evaluation module includes a boundary change data acquisition unit, a boundary change vector data analysis unit, and a change risk index evaluation unit; The boundary change data acquisition unit is used to collect the acquisition times of the three-dimensional remote sensing images to be analyzed, and collect the acquisition times of the initial remote sensing images of the target area, sort the collected acquisition times in chronological order, and based on the sorting results, compare the three-dimensional remote sensing images obtained at two adjacent acquisition times. Based on the comparison results, analyze the abnormal boundary change data of each patch object and generate the boundary change data set of each patch object; The boundary change vector data analysis unit determines whether each abnormal boundary change data stored in the boundary change dataset of the selected patch object meets the screening conditions. The screening conditions are that the values of the change direction angle, change volume, and change scale of the patch object are all 0. If it meets the conditions, the corresponding abnormal boundary change data is removed from the boundary change dataset of the selected patch object. If it does not meet the conditions, the corresponding abnormal boundary change data is retained in the boundary change dataset of the selected patch object. All patch objects are traversed to obtain the abnormal boundary change datasets of each patch object; The change risk index evaluation unit evaluates the real-time change risk index of each patch object according to the continuous situation in time and the degree of abnormal movement in time of each abnormal boundary change data stored in the abnormal boundary change datasets of each patch object.

4. The intelligent patch recognition and risk warning system based on deep learning according to claim 3, characterized in that: The specific method for the abnormal boundary change data acquisition unit to generate the abnormal boundary change datasets of each patch object is as follows: According to the chronological order of the acquisition time of the three-dimensional remote sensing images to be analyzed, number each three-dimensional remote sensing image to be analyzed. The numbering result is: j = 1, 2,..., n; Compare the three-dimensional remote sensing image to be analyzed 1 with the initial remote sensing image, and compare the three-dimensional remote sensing image j to be analyzed with the three-dimensional remote sensing image j - 1 to be analyzed, respectively obtaining the change direction angle, change volume, and change scale corresponding to the patch object numbered i in the three-dimensional remote sensing images 1 and j to be analyzed; Perform image registration processing on each 3D remote sensing image to be analyzed with the initial remote sensing image, and randomly select an alignment point A based on the registration result j , taking the alignment point A j as the coordinate origin, respectively construct a three-dimensional space coordinate system j and a three-dimensional space coordinate system p; Obtain the central coordinate U of the patch object i in the three-dimensional space coordinate system p ip = (X ip , Y ip , Z ip ), calculate the distance value d ip between each boundary point of the patch object i in the three-dimensional space coordinate system j and the coordinate U according to the distance formula between two points in space, obtain the boundary point coordinate B ipj corresponding to the maximum value of d ipj = (X ipj , Y ipj , Z ipj , Z ipj ), calculate the included angle between the vector and the vector through the included angle formula of three-dimensional space vectors, and obtain the change direction angle β ipj of the patch object i in the remote sensing image j to be analyzed, where b represents a constant greater than 0; Calculate the difference between the regional volume corresponding to the patch object i in the remotely sensed image to be analyzed and the regional volume corresponding to the patch object i in the initial remotely sensed image, and obtain the change volume S of the patch object i in the remotely sensed image j to be analyzed ipj ; Denote d ipj The minimum value of is used as the change scale C of the patch object i in the remotely sensed image j to be analyzed ipj ; Boundary change dataset M of patch object i i ={(β ip1 ,S ip1 ,C ip1 ),…,(β ipj ,S ipj ,C ipj )}。 5. The intelligent recognition and risk warning system for map patches based on deep learning according to claim 4, characterized in that: The specific method for the change risk index evaluation unit to evaluate the real-time change risk index of each patch object is as follows: Let the abnormal boundary change dataset of the patch object i be M´ i , for the change volume stored at the (v + 1)-th position in M´ i and the change volume stored at the v-th position in M´ i , calculate the difference S i(v+1→v) between them. Calculate the ratio f i(v+1→v) of S i(v+1→v) to t, and obtain the abnormal degree set N i of the patch object i with respect to the change volume. N i = {f i(1+1→1) , …, f i(v+1→v)}. According to W i = P´ i / P i , calculate the first abnormal degree of the patch object i. P i represents the overall average value of N i , P´ i represents the overall standard deviation of N i . t represents the shooting interval time of the satellite remote sensing technology for the target area. v = 1, 2, …, V represents the numbering of each element stored in M´ i in the order from left to right. V represents the total number of elements stored in M´ i ; Calculate the second change degree \(W'\) of the patch object \(i\) according to the calculation method of the first change degree i ; For M' i Calculate the difference β i between the transition direction angle stored at the (v + 1)-th position in and the transition direction angle stored at the v-th position in M' i(v+1→v) If β i(v+1→v) = 0, then obtain the first influence coefficient h i(v+1→v) of the third abnormal degree of the patch object i according to the discrete degree of β i(v+1→v) in time. If β i(v+1→v) ≠ 0, then calculate the difference between β i(v+1→v) and t to obtain the second influence coefficient h' i(v+1→v) of the third abnormal degree of the patch object i. Calculate the product f'' i(v+1→v) between h' i(v+1→v) and 1 + h i(v+1→v) to obtain the abnormal degree set N'' i of the patch object i with respect to the transition direction angle. N'' i = {f'' i(1+1→1) , …, f'' i(v+1→v)}. Calculate the second abnormal degree of the patch object i according to W'' i = R' i / R i . R i represents the overall average value of N'' i , and R' i represents the overall standard deviation of N'' i ; According to E i = g1 × |W i | + g2 × |W´ i | + g3 × |W´´ i | predict the transition risk index of the patch object i at the current moment, where g1, g2, and g3 respectively represent the weight coefficients corresponding to the transition volume, transition scale, and transition direction angle.

6. The intelligent patch recognition and risk warning system based on deep learning according to claim 5, characterized in that: The risk warning module includes a risk warning level analysis unit and a risk warning time determination unit; The risk warning level analysis unit collects the total number w of associated patch objects of the selected patch object, and combines the shortest distance values from the selected patch object to each associated patch object to determine the risk warning levels of each patch object; The risk warning time determination unit determines the risk warning time of each patch object according to the risk warning levels of each patch object. The risk warning time = L + (ε - δ´) × t, where ε represents the number corresponding to the risk warning level of the patch object.

7. The intelligent recognition and risk warning system for map patches based on deep learning according to claim 6, characterized in that: The specific method for the risk warning level analysis unit to determine the risk warning levels of each patch object is as follows: Take the shortest distance value from the selected patch object to each associated patch object as the base, and take the angle between the terminal side of the change direction angle of the selected patch object and the terminal side of the change direction angle of each associated patch object as the exponent to construct an exponential function, and calculate the correlation coefficient between the selected patch object and each associated patch object. If the terminal side of the change direction angle of the selected patch object does not intersect with the terminal side of the change direction angle of each associated patch object, it is considered that the angle is 0. Map the mean value δ of the correlation coefficients to the range from 0 to 1 to obtain the mapped value δ´. Take δ´ as the risk warning coefficient of the selected patch object, and analyze the risk warning level of the selected patch object; The risk warning coefficient δ´ of each patch object other than the selected patch object is 0; If 0.8 ≤ δ´ ≤ 1, it is analyzed that the risk warning level of the selected patch object is a first-level warning; If 0.5 ≤ δ´ < 0.8, it is analyzed that the risk warning level of the selected patch object is a secondary warning; If 0.3 ≤ δ´ < 0.5, it is analyzed that the risk warning level of the selected patch object is a tertiary warning; If 0 ≤ δ´ < 0.3, it is analyzed that the risk warning level of the selected patch object is a quaternary warning.

8. A deep learning-based intelligent identification and risk warning method for map patches applied to the deep learning-based intelligent identification and risk warning system for map patches according to any one of claims 1-7, characterized in that: The method includes: S10: Periodically photograph the target area through satellite remote sensing technology, process the multi-source data of the target area obtained at different times and perform 3D modeling to obtain several 3D remote sensing images. Use the watershed algorithm to perform raster segmentation on the 3D remote sensing images to obtain several 3D remote sensing images to be analyzed, and intelligently identify the patch objects and the boundaries of each patch object in each 3D remote sensing image to be analyzed; S20: Analyze the boundary change data of each patch object according to the boundary information of each patch object, and evaluate the change risk index of each patch object based on the abnormal boundary change data; S30: Determine the real-time change type of each patch object according to the abnormal boundary change data of each patch object, collect the target patch object, and find the associated patch object of the target patch object based on the coincidence of the abnormal trend between the target patch object and its adjacent patch objects; S40: Analyze the risk warning level of each patch object, and determine the risk warning time of each patch object according to the analysis result.

Images