Estimation method, estimation system, and program

Abstract

This invention provides an estimation method that can accurately estimate landslide risk areas at any given location, including locations where no buildings are present nearby. [Solution] The estimation method is performed by an estimation system 1 that estimates landslide risk areas C on a map image P having topographic information and building information. The estimation system 1 includes a training step in which a deep learning model 10 is trained to input a map image P using a training image set TR which is a pair of a training map image T and a designated image R showing landslide risk areas C on the map image T, and determines whether or not each point on the input map image P is a landslide risk area C; an image generation step in which a composite image CP is generated in which a virtual building image BI is added to a part of the map image P to be estimated; and a determination step in which the composite image CP is input to the deep learning model 10 and determines whether or not each point on the map image P to be estimated is a landslide risk area C.

Description

[Technical Field]

[0001] The present invention relates to an estimation method, an estimation system, and a program. [Background technology]

[0002] Prefectures designate hazard zones (landslide risk zones) where there is a risk of collapse of steep slopes, debris flows, and landslides occurring, based on the "Act on the Promotion of Measures against Landslides in Landslide Warning Zones, etc." (Landslide Prevention Act), in order to protect human lives from landslides. Hazard zones are broadly divided into landslide warning zones (yellow zones), which are areas where there is a risk of harm to the lives or bodies of residents, and landslide special warning zones (red zones), which are areas where there is a risk of significant harm to the lives or bodies of residents due to damage to buildings, etc.

[0003] The boundaries of the warning zone are determined by personnel identifying areas at risk of landslides using topographic maps and aerial photographs, and then conducting on-site surveys of the identified areas. Warning zones are reviewed approximately every five years. The surveys required to designate warning zones require a tremendous amount of time and effort, and there is a risk of overlooking dangerous areas that should be designated as warning zones due to topographical changes such as land development. Furthermore, it is difficult to foresee the potential risk of landslides when developing undesignated areas.

[0004] Based on the above, attempts have been made to estimate hazardous areas from map images using deep learning models (Non-Patent Literature 1). Non-Patent Literature 1 discloses an attempt to estimate areas at risk of landslides from red relief maps, standard maps, and aerial photographs using Mask R-CNN (Mask Region Based-Convolutional Neural Networks), one of the deep learning models. [Prior art documents] [Non-patent literature]

[0005]

Non-Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0006] However, in such a method, since the relationship between buildings and sediment disaster risk areas is considered to determine the sediment disaster risk areas, there is a problem that the sediment disaster risk areas with a risk of sediment disasters cannot be accurately estimated in the vicinity of locations where there are no buildings.

[0007] The present invention has been made under the above circumstances, and an object thereof is to provide an estimation method, an estimation system, and a program capable of accurately estimating sediment disaster risk areas at any location including locations where there are no buildings in the vicinity.

Means for Solving the Problems

[0008] In order to achieve the above object, an estimation method according to a first aspect of the present invention is an estimation method executed by an estimation system that estimates sediment disaster risk areas on a map image having terrain information and building information, a training step of training a deep learning model that inputs the map image using a training image set that is a set of the map image for training and an image indicating sediment disaster risk areas on the map image, and determines whether each location on the input map image is a sediment disaster risk area; An image generation step of generating a composite image in which a virtual building image is added to a part of the map image to be estimated; A determination step of inputting the composite image into the deep learning model and determining whether each point on the map image to be estimated is a landslide disaster risk area; including.

[0009] The building image may be an image of a group of buildings configured by two-dimensionally arranging figures representing the outer shape of the buildings seen from above. It may also be like this.

[0010] Before executing the image generation step, when the map image indicating the points where disasters actually occurred is input into the deep learning model, the intervals, number of rows, and number of columns of the buildings in the building image are adjusted so that the detection accuracy of determining that the points where disasters actually occurred are landslide disaster risk areas is the best. It may also be like this.

[0011] The deep learning model is a model that executes tasks of object detection and segmentation, and determines whether each pixel is a landslide disaster risk area. It may also be like this.

[0012] In the training step and the image generation step, a first image representing the inclination amount of each point, a second image representing the valley degree of each point, and a third image representing the presence or absence of buildings at each point are generated from the map image, the first image, the second image, and the third image are converted into a three-channel image, and the three-channel image is generated as an image to be input into the deep learning model. It may also be like this.

[0013] The map image includes a red stereoscopic map image having the terrain information. The aforementioned red relief map image is subjected to HSV conversion to generate the first image, which represents the slope of each point in terms of saturation, and the second image, which represents the ridge-valley gradient of each point in terms of brightness. It would be acceptable to do so.

[0014] An estimation system according to a second aspect of the present invention is: An estimation system for estimating landslide risk areas on a map image containing topographic information and building information, A deep learning model that determines whether each point on the map image, which is the target of estimation, is in a landslide hazard area, A control unit that trains the deep learning model using a training image set which is a set of training map images and images showing landslide risk areas on the map images, The system includes an image generation unit that generates a composite image in which a virtual building image is added to a portion of the map image that is the target of estimation, The control unit, The synthesized image is input to the deep learning model, which determines whether each point on the map image, which is the target of estimation, is located in a landslide hazard area.

[0015] A program relating to the third aspect of the present invention is: A computer that estimates landslide risk areas on map images containing topographic and building information, A training step involves training a deep learning model that determines whether each point on an input map image is in a landslide hazard zone, using a training image set which is a pair of a training map image and an image showing a landslide hazard zone on the map image. Image generation step: Generates a composite image in which a virtual building image is added to a part of the map image that is the target of estimation, A determination step in which the synthesized image is input to the deep learning model and it is determined whether or not each point on the map image that is the target of estimation is in a landslide hazard area, Make it run. [Effects of the Invention]

[0016] According to the estimation method, estimation system, and program of the present invention, it is possible to accurately estimate landslide risk areas at any point, including locations where there are no buildings nearby. [Brief explanation of the drawing]

[0017] [Figure 1] This is a block diagram showing the functional configuration of an estimation system according to an embodiment of the present invention. [Figure 2] (A) is a diagram showing an example of a conventional map image input to a deep learning model. (B) is a diagram showing an example of a map image generated by the image generation unit in Figure 1 and input to a deep learning model. [Figure 3] This figure shows an example of a building image and its parameters. [Figure 4] This figure shows an example of input and output images for a deep learning model. [Figure 5] Figure 1 is a block diagram showing the hardware configuration of the estimated system. [Figure 6] Figure 1 is a flowchart of the estimation process, which is the main flow of the estimation system. [Figure 7] This is a flowchart of the subroutine for the pre-training process. [Figure 8] This is a flowchart of the subroutine for the building image optimization process. [Figure 9] This is a flowchart of the subroutine for the judgment process. [Figure 10] (A) is a graph showing the change in the accuracy of detecting disaster locations when the spacing between virtual buildings is changed. (B) is a graph showing the change in the accuracy of detecting disaster locations when the number of columns and rows of virtual buildings is changed. [Figure 11] This table compares the recall, precision, and F1 score of the estimation system shown in Figure 1, using a map image P with the location of the landslide as the center of the image, with the recall, precision, and F1 score of the conventional system. [Figure 12]Figure 1 is a table showing the number of successful detections, demonstrating that the estimation system can detect a wide range of landslide risk areas. [Figure 13] This is a block diagram showing a modified example of the functional configuration of the estimation system according to an embodiment of the present invention. [Figure 14] This figure shows the image conversion process in the image generation unit of Figure 13. [Figure 15] Figure 13 is a flowchart of the pre-training subroutine in the estimation system. [Figure 16] Figure 13 is a flowchart of the subroutine for building image optimization processing in the estimation system. [Figure 17] Figure 13 is a flowchart of the subroutine for the decision-making process in the estimation system. [Figure 18] Figure 13 is a flowchart of the image generation process in the estimation system. [Figure 19] Figure 13 is a graph showing how the accuracy of detecting disaster locations changes when the number of rows and columns of virtual buildings and the spacing between buildings are changed in the estimation system. [Figure 20] Figures 1 and 13 show a table comparing the recall, precision, and F1 score of the estimation system using a map image P with the location of the landslide at its center, with those of the conventional system. [Figure 21] Figure 13 is a table showing the number of successful detections, demonstrating that the estimation system can detect a wide range of landslide risk areas. [Modes for carrying out the invention]

[0018] Embodiments of the present invention will be described in detail below with reference to the drawings. In each drawing, the same or equivalent parts are denoted by the same reference numerals. In the embodiments described below, expressions such as “having,” “including,” or “containing” also include the meaning of “consisting of” or “composed of.”

[0019] [Estimation System] As shown in Figure 1, the estimation system 1 comprises a storage unit 2 that stores the map image P to be estimated, and a deep learning model 10 that determines the landslide hazard risk areas C included in the map image P. The estimation system 1 estimates the landslide hazard risk areas C using the deep learning model 10.

[0020] The estimation system 1 also includes a control unit 11 and an image generation unit 12. The control unit 11 provides overall control over the processing of the estimation system 1. The control unit 11 can acquire various data, including map image data, from an external server device (not shown) via a communication network. The map image P stored in the storage unit 2 is generated based on the acquired map image data.

[0021] [Deep Learning Models] The deep learning model 10 is a model that performs object detection and segmentation tasks. For example, a Mask R-CNN (Mask Region based Convolutional Neural Networks) can be used as such a model. When Mask R-CNN is used, the deep learning model 10, upon input of a map image P, determines whether each point on the input map image P is a landslide hazard risk area C, pixel by pixel. If the map image P contains a landslide hazard risk area C, the deep learning model 10 performs segmentation between the landslide hazard risk area C and other regions, and outputs an estimated image PR that includes the shape of the landslide hazard risk area C and the bounding box B that encloses the landslide hazard risk area C.

[0022] [Map image] Map image P is a two-dimensional map image representing the topography, structures, etc. of the land as seen from above. Sediment disaster risk area C is a warning area based on the Sediment Disaster Prevention Act, which refers to an area where there is a risk of collapse of steep slopes, debris flows, and landslides occurring.

[0023] Map image P is a bitmap image of a two-dimensional map containing topographic information and building information. In map image P, the pixel value of each pixel represents the topographic information and building information of the corresponding point. Topographic information refers to information about the terrain at each point, such as height and gradient, obtained from a digital elevation model or the like. Building information refers to information about buildings, including the size and outline of the building structures constructed at each point. In this embodiment, the map image P to be estimated is composed of a set of two images: a 3D map image P1 and a standard map image P2. The 3D map image P1 and the standard map image P2 are images that represent the same area.

[0024] The 3D map image P1 contains topographic information. For example, a red 3D map can be used as the 3D map image P1. A red 3D map is an image that uses the saturation and brightness of red to represent 3D terrain such as slopes, ridges, and valleys in 2D. Red 3D maps are particularly good at extracting the characteristics of steep slopes. Note that the 3D map image P1 is not limited to a red 3D map; for example, a CS 3D map may also be used. A CS 3D map is a map that enables visual and intuitive terrain interpretation by combining curvature and slope. It is a map that represents 3D by coloring three terrain quantities, "elevation," "slope," and "undulation (curvature)," with different tones, and then layering multiple images and applying transparency.

[0025] Standard map image P2 includes building information. For example, a standard map image provided by the Geospatial Information Authority of Japan can be used as standard map image P2. In standard map image P2, natural objects such as rivers and mountains, and artificial structures such as buildings, are represented by geometric shapes, with buildings represented as shapes with a light orange interior and an orange exterior. Typical standard map images are provided as vector tile images that describe map information based on coordinates and attributes. Standard map image P2 is a bitmap image extracted from a standard map image in vector tile format. By extracting standard map image P2 from a standard map image in vector tile format, map objects can be extracted and drawn at any scale without degradation of image quality.

[0026] Map image P may also include aerial photographs (image file format, e.g., JPEG or PNG) and land use maps. The land use map is a map created from land use subdivision mesh data from the National Land Numerical Information. The land use subdivision mesh data is vector data that shows the land use situation represented by 11 types of items determined from map symbols and the color tones of satellite images in 100m (1 / 10 subdivision) mesh units. The land use map is a bitmap image that shows six types of land—agricultural land, building use, roads, railways, and other land—in different colors.

[0027] The 3D map image P1 and the standard map image P2 are, for example, 512 x 512 pixels. The standard map image P2 is created by stitching together four 256 x 256 raster images made from vector tile images at zoom level 15. As a result, one pixel corresponds to 3.8m, and the entire image represents an area of ​​approximately 2km x 2km. Both the 3D map image P1 and the standard map image P2 are 3-channel color images (red (R), green (G), and blue (B)).

[0028] [Control Unit] The control unit 11 acquires map image data provided from a server device such as the Geospatial Information Authority of Japan or a data distribution support platform, and generates a 3D map image P1 and a standard map image P2 of the same size representing the inspection area as map images P based on the acquired map image data. If there are multiple inspection areas, a map image P is generated for each inspection area. The control unit 11 stores the generated map images P in the storage unit 2. The 3D map image P1 and the standard map image P2 are superimposed so that the same locations on the map are in the same location, and are input to the deep learning model 10.

[0029] The control unit 11 acquires map image data provided from a server device such as the Geospatial Information Authority of Japan or a data distribution support platform. Furthermore, the control unit 11 acquires information from a server device of the Ministry of Land, Infrastructure, Transport and Tourism, for example, information created by prefectural surveys, indicating areas that were previously designated as landslide hazard zones C. Based on the acquired map image data and the information indicating areas that were previously designated as landslide hazard zones C, the control unit 11 generates multiple training image sets TR, which are sets of training map images T comprising a 3D map image P1 of the same size representing the same area, a standard map image P2, and a designated image R on the map image T indicating landslide hazard zones C (areas where collapse of steep slopes is mainly a concern). The control unit 11 stores the generated multiple training image sets TR in the storage unit 2.

[0030] The control unit 11 trains the deep learning model 20 using a training image set TR, which is a set of a training map image T and a designated image R that shows a landslide hazard area C on the map image T.

[0031] The control unit 11 acquires information indicating the actual locations where disasters occurred from the server equipment of the prefectural government or the information administrator. The server equipment of the information administrator includes, for example, a data distribution support platform. Based on the acquired location information, the control unit 11 acquires map image data of those locations provided by the server equipment, such as the Geospatial Information Authority of Japan or the data distribution support platform. Based on the acquired map image data, the control unit 11 generates an adjustment map image (adjustment image) O comprising a 3D map image P1 of the same size representing the same area centered on the location where the disaster occurred, and a standard map image P2. The control unit 11 stores the generated multiple adjustment images O in the storage unit 2. The adjustment image O is the image in which the disaster actually occurred at the center.

[0032] [Image generation unit] The image generation unit 12 generates a composite image CP by adding a building image BI of a virtual building E (see Figure 2(B)) to a portion of the map image P that is the target of estimation. The control unit 11 inputs the composite image CP generated by the image generation unit 12 into the deep learning model 10, which has been trained as described above, and has it determine the landslide hazard area C in the composite image CP.

[0033] The image generation unit 12 receives the map image P to be estimated as input. The image generation unit 12 adds building image BI to the map image P. In conventional systems, as shown in Figure 2(A), the standard map image P2 in the map image P may be extracted from the vector tile image into a map image P2' with buildings removed, and a map image P2'' with only buildings extracted from the standard map image P2 and everything else transparent, and these are input to the deep learning model 10. In contrast, as shown in Figure 2(B), the image generation unit 12 may add building image BI to a part of the map image P2'' from which only buildings have been extracted from the standard map. The building image BI is added to the center of the map image P2''. The map image P consists of a total of 9 layers (9 channels) of input images: the 3D map image P1 (a 3-channel color image with R, G, and B channels), the map image P2' with buildings removed (a 3-channel color image with R, G, and B channels), and the map image P2'' with only buildings extracted and the building image BI added (a 3-channel color image with R, G, and B channels).

[0034] [Building image] As shown in Figure 3, the building image BI is an image of a group of buildings composed of a two-dimensional arrangement of shapes (e.g., rectangles) representing the outline of the buildings E as seen from above. These shapes may be circles or other shapes, not just rectangles. When the control unit 11 inputs an adjustment image O indicating the location where a disaster actually occurred to the deep learning model 10, it adjusts the spacing D, number of columns N, and number of rows M of the buildings E in the building image BI so that the detection accuracy is best when the location where a disaster actually occurred is determined to be a landslide hazard area C.

[0035] In this building image BI, the area of ​​building E consists of the building itself and the surrounding margin. Building E is set to a size of 6x6 pixels, equivalent to, for example, 15m x 15m (the size of a typical house), the building itself is set to a size of 4x4 pixels, and the margin is set to a size of 2 pixels. The building itself is set to the same color as the building in the standard map.

[0036] As shown in Figure 1, the control unit 11 instructs the image generation unit 12 to read an adjustment image O from the storage unit 2, which indicates that the location is where a disaster actually occurred in the landslide hazard area C. The image generation unit 12 then combines multiple building images BI, each with different spacing D, number of columns N, and number of rows M of buildings E, with the map image of the read adjustment image O to generate multiple composite images OP.

[0037] The control unit 11 instructs the deep learning model 10 to determine whether each point on each composite image OP generated by the image generation unit 12 is in a landslide hazard risk area C. The control unit 11 then determines whether the location where the disaster actually occurred is included in the locations determined to be in a landslide hazard risk area C. The control unit 11 calculates the detection accuracy required to detect the location where the disaster actually occurred as part of a landslide hazard risk area C. The control unit 11 then searches for the optimal spacing D, number of columns N, and number of rows M of buildings E in the building image BI that result in the best detection accuracy.

[0038] [Summary of structure] As described above, in the estimation system 1, the control unit 11 trains the deep learning model 10 using the training image set TR, and then further adjusts the spacing D, number of columns N, and number of rows M of buildings E in the building image BI as described above. Subsequently, as shown in Figure 4, the control unit 11 causes the image generation unit 12 to synthesize the building image BI, with the adjusted spacing D, number of columns N, and number of rows M of buildings E, onto the map image P to be estimated, thereby generating a composite image CP. The control unit 11 then causes the deep learning model 10 to determine whether each point on the composite image CP is a landslide hazard risk area C. The set of points determined to be landslide hazard risk areas C becomes the estimated landslide hazard risk area. This clarifies the location and size of the estimated landslide hazard risk area C.

[0039] [Hardware configuration] The estimation system 1 shown in Figure 1 is realized, for example, by a computer having the hardware configuration shown in Figure 5 implementing a software program. Specifically, the estimation system 1 includes a CPU (Central Processing Unit) 21 that controls the entire device, an AI accelerator 22 such as a GPU (Graphics Processing Unit) or NPU (Neural Processing Unit) that performs training and judgment processing of deep learning models, a main memory 23 that operates as a workspace for the CPU 21, an external memory 24 that stores programs 30 etc. executed by the CPU 21, an operation unit 25, a display unit 26, a communication interface 27, and an internal bus 29 that connects these.

[0040] The CPU 21 and AI accelerator 22 implement the functions of the estimation system 1 by executing the program 30.

[0041] The main memory 23 consists of RAM (Random Access Memory), etc. Programs 30 executed by the CPU 21 and AI accelerator 22 are loaded into the main memory 23 from external memory 24. The main memory 23 is also used as a workspace (temporary data storage area) for the CPU 21 and AI accelerator 22.

[0042] The external memory 24 consists of non-volatile memory such as flash memory and hard disk. The external memory 24 pre-stores a program 30 for execution by the CPU 21 and the AI ​​accelerator 22.

[0043] The control unit 25 consists of devices such as a keyboard and a mouse, and an interface device that connects these devices to the internal bus 29.

[0044] The display unit 26 is composed of a display device such as a liquid crystal monitor.

[0045] The communication interface 27 is an interface that transmits and receives data with external devices via a communication network. Data input and output are performed through this communication interface 27.

[0046] The functions of Estimation System 1 can be implemented in a computer system consisting of one or more computers, each including one or more processors and one or more memory units, including non-temporary storage media. Multiple computers communicate with each other via a connected communication network to implement the functions of Estimation System 1. For example, some of the functions of Estimation System 1 may be implemented in one computer, while other functions may be implemented in other computers.

[0047] [Estimation method] Next, the estimation process, i.e., the estimation method, performed by the estimation system 1 according to this embodiment will be explained with reference to the flowcharts in Figures 6, 7, 8, and 9.

[0048] As shown in Figure 6, in the estimation process, the estimation system 1 executes a pre-training subroutine (step S1; training step) that trains the deep learning model 10 using a training image set TR (see Figure 1), which is a set of a training map image T and a designated image R showing a landslide hazard area C on the map image T.

[0049] In the pre-training subroutine, as shown in Figure 7, first, the control unit 11 acquires map image data from a server device such as the Geospatial Information Authority of Japan or a data distribution support platform, and acquires information indicating areas that have been previously designated as landslide hazard risk areas C from a server device of the Ministry of Land, Infrastructure, Transport and Tourism (step S11). Subsequently, based on the acquired map image data and the information indicating areas that have been previously designated as landslide hazard risk areas C, the control unit 11 processes the map image data by cropping or resizing it, and generates multiple training image sets TR, which are pairs of training map image T comprising a 3D map image P1 of the same size representing the same area and a standard map image P2, and a designated image R on the map image T indicating landslide hazard risk areas C (areas where collapse of steep slopes is mainly a concern) (step S12). The generated training image sets TR are stored in the storage unit 2.

[0050] Next, the control unit 11 inputs the adjustment map image T from the generated training image set TR to the deep learning model 10, and trains the deep learning model 10 so that the discrepancy between the landslide hazard risk area C determined by the image output from the deep learning model 10 in response to that input and the landslide hazard risk area C in the specified image R is reduced (step S13). In this training, the deep learning model 10 is trained using the training image set TR for learning, and the hyperparameters of the training of the deep learning model 10 are verified using the training image set TR for verification. Furthermore, the deep learning model 10 is evaluated using the training image set TR for testing.

[0051] The control unit 11 stores information such as the internal parameters of the trained deep learning model 10 (step S14). After step S14 is completed, the control unit 11 terminates the pre-training subroutine.

[0052] Returning to Figure 6, the control unit 11 then executes a subroutine for building image optimization processing (step S2; adjustment step). In the building image optimization processing subroutine, as shown in Figure 8, the control unit 11 first extracts multiple evaluation points based on map image data acquired from an external server device and information indicating the locations where disasters actually occurred, acquired from a server device of a prefecture or information administrator (step S21). Subsequently, the control unit 11 generates an adjustment image O for each evaluation point (step S22). As mentioned above, this adjustment image O is an image of the same size and scale as the training map image T and the map image P to be estimated, and it is an image that includes the locations where landslides occurred in that area. The generated adjustment image O is stored in the storage unit 2.

[0053] Next, the control unit 11 loads the internal parameters obtained through training the deep learning model 10 into the deep learning model 10 (step S23). Then, the control unit 11 increases the spacing D of buildings E in the building image BI by +1 from the minimum value to the maximum value (step S24), and increases the number of columns N and rows M of buildings E by +1 from the minimum value to the maximum value (step S25), until the number of columns N and rows M of buildings E reach their maximum values ​​(step S29), and the spacing D of buildings E reaches its maximum value (step S30), performing a loop process. In this loop process, with each iteration, the control unit 11 causes the image generation unit 12 to generate multiple composite image OPs (step S26), inputs the multiple composite image OPs into the deep learning model 10 (step S27), and calculates the detection accuracy, which is the percentage of locations where a disaster actually occurred that are determined to be landslide hazard risk areas (step S28).

[0054] After the loop processing is complete, the control unit 11 determines the optimal values ​​for the number of columns N, the number of rows M, and the spacing D of building E that provide the best detection accuracy (step S31). Figure 10(A) shows the detection accuracy (recall rate) of disaster locations when the number of columns N and the number of rows M of the virtual building E are fixed at 5 and the spacing D of building E is changed. In this case, 8 pixels, which has the highest recall rate, is determined as the optimal value for the spacing D of building E. Figure 10(B) also shows the detection accuracy (recall rate) of disaster locations when the spacing D of the virtual building E is fixed at 8 pixels and the number of columns N and the number of rows M are changed. In this case, 5, which has the highest recall rate, is determined as the optimal values ​​for the number of columns N and the number of rows M of building E. Returning to Figure 8, after step S31 is executed, the control unit 11 terminates the building image optimization processing subroutine.

[0055] Returning to Figure 6, the control unit 11 waits until it is commanded to start the determination (step S3; No). Once it is commanded to start the determination (step S3; Yes), the control unit 11 executes the determination processing subroutine (step S4).

[0056] In the judgment processing subroutine, as shown in Figure 9, the control unit 11 generates map images P, which are the estimation targets for each of the multiple inspection areas, based on map image data provided from an external server device (step S41). The generated map images P are stored in the storage unit 2. Subsequently, the control unit 11 causes the image generation unit 12 to generate a composite image CP, which is a part of the estimation target map image P read from the storage unit 2 with a virtual building image BI added to it (step S42; image generation step). This building image BI is an image in which the spacing D of buildings E, the number of columns N, and the number of rows M have been adjusted in the building image optimization process (see Figure 8).

[0057] Next, the control unit 11 loads the internal parameters obtained from the pre-training process into the deep learning model 10 (step S43). Subsequently, the control unit 11 inputs a composite image CP of the undetermined inspection area into the deep learning model 10 (step S44; determination step), and causes the deep learning model 10 to determine whether each point on the map image P is in a landslide hazard risk area C, and outputs the determination result (step S45; determination step).

[0058] Next, the control unit 11 determines whether or not there is a map image P of an undetermined inspection area (step S46). As long as there is a map image P of an undetermined inspection area (step S46; Yes), the control unit 11 repeats the input of a composite image CP of the undetermined inspection area (step S44) and the output of the determination result of the landslide hazard risk area C (step S45).

[0059] If there is no map image P of an undetermined inspection area (step S46; No), the control unit 11 terminates the determination processing subroutine. After terminating the determination processing subroutine, the system returns to Figure 6, and the control unit 11 terminates the estimation process.

[0060] [Estimated accuracy] In the estimation system 1 described above, an evaluation is performed to determine whether or not a total of 167 locations—84 locations where landslides actually occurred in areas not designated as landslide risk zone C, and 83 locations that are not steep slopes and do not pose a risk of landslides—should be estimated as landslide risk zone C. Figure 11 shows the recall, precision, and F1 score of the estimation system 1 according to this embodiment when using a map image P with the location where a landslide occurred as the center of the image, and also shows the recall, precision, and F1 score of a conventional deep learning model. The recall is the probability that a risk location was correctly identified, the precision is the probability that the location identified as a risk location was correct, and the F1 score is the harmonic mean of the recall and precision.

[0061] As shown in Figure 11, the estimation system 1 according to this embodiment shows significant improvements in recall, precision, and F1 score compared to conventional systems. The estimation system 1 can accurately detect the location where a disaster actually occurred and reduce false positives.

[0062] The estimation accuracy shown in Figure 11 represents the estimation accuracy when estimating the landslide risk area C using an image centered on the location where the disaster occurred. On the other hand, Figure 12 shows the number of successful detections in the estimation system 1 according to this embodiment, for determining 5 of the 84 locations where the aforementioned landslide occurred, when the location where the landslide occurred was displaced in the longitude and latitude directions from the center of the map image P. In Figure 12, the horizontal axis represents the displacement in the longitude direction from the disaster location at the center of the map image P (unit: ×10 -4 The graph shows the latitude (in degrees), with the vertical axis representing the latitudinal displacement from the disaster site at the center of map image P (unit: ×10). -4 The graph shows the degree of displacement, and the number in each cell indicates the number of successful detections for the corresponding longitude and latitude displacements. As shown in Figure 12, the number of successful detections is 5 not only at the center of map image P, but also in the surrounding area including the center. This indicates that estimation system 1 can accurately estimate not only the center of map image P, but also risk points shifted from the center, as landslide risk areas C.

[0063] The three-dimensional map image P1, map image P2', and map image P2'' (see Figures 2(A) and 2(B)) that make up map image P are each 3-channel color images. In other words, in estimation system 1, the deep learning model 10 receives a total of 9 layers (3 x 3 channels) of images as input. However, most deep learning models 10 that perform image processing are designed to receive 3-channel images as input. Therefore, by creating and inputting a 3-channel image instead of the 9-channel image input to the deep learning model 10, it is possible to use the deep learning model 10 for estimating landslide hazard areas C without modifying the input interface of a typical image processing deep learning model 10.

[0064] To achieve conversion to a 3-channel color image, the estimation system 1B shown in Figure 13 includes an image generation unit 12B instead of an image generation unit 12. The image generation unit 12B converts a map image P', which consists of 6-channel map images P1 and P2'' obtained by removing map image P2' from a total of 9-channel map image P, into a 3-channel image that can be input to a general image processing deep learning model 10. Map image P2' is a map image obtained by removing buildings from a standard map and contains information such as contour lines and roads, but this information is thought to be correlated with the information contained in the 3D map images P1 and P2''.

[0065] As shown in Figure 14, map image P' consists of a set of two images: a 3D map image P1 and a map image P2'' from which only buildings have been extracted from a standard map image. Here, 3D map image P1 is assumed to be a red 3D map image.

[0066] As shown in Figure 14, the image generation unit 12B generates a first image P representing the slope amount of each point from the map image P'. S And the second image P, which shows the degree of ridge / valley at each point. V And the third image P, which shows the presence or absence of buildings at each location. B The system generates the following: The slope amount represents the degree of the terrain's incline, and the ridge-valley degree is a numerical representation of whether the terrain is ridge-shaped or valley-shaped. The ridge-valley degree is an index that indicates whether the terrain is ridge-shaped or valley-shaped; if the terrain is ridge-shaped, the ridge-valley degree shows a positive value. If the terrain is valley-shaped, the ridge-valley degree shows a negative value. If the terrain is steep, the absolute value of the ridge-valley degree increases, and if the terrain is flat, the ridge-valley degree becomes 0. The ridge-valley degree is described, for example, in Japanese Patent Publication No. 7508722. The image generation unit 12B performs HSV conversion on the red 3D map image P1 to generate the first image P S As a result, an image is generated that represents the degree of slope at each point using saturation, and the second image P VAs an example, an image representing the valley degree of each location in terms of brightness is generated. The HSV color space is defined by three elements: hue H, saturation S, and brightness V. In HSV conversion, the pixel value (R, G, B) of the three-dimensional map image P1 in RGB format is converted into the pixel values (H, S, V) of hue H, saturation S, and brightness V according to the following formula.

[0067] Here, for each pixel, C max = max(R, G, B), and C min = min(R, G, B). The conversion formula from RGB to HSV is expressed as follows. S = (C max - C min ) / C max V = C max C max = R: H = 60 × (G - B) / (C max - C min ) C max = G: H = 60 × (B - R) / (C max - C min ) + 120 C max = B: H = 60 × (R - G) / (C max - C min ) + 240 When R, G, B = 0: H = 0 The image generation unit 12B generates an image P H based on the H of each pixel converted by the above formula, generates a first image P S based on the S of each pixel, and generates a second image P V based on the V of each pixel.

[0068] Furthermore, the image generation unit 12B performs grayscale conversion on the bitmap format map image P2” obtained by extracting only buildings from the standard map image which is a vector tile image, to generate a third image P B .

[0069] Furthermore, the image generation unit 12B generates the first image P S , the second image P V and the third image P BThis is converted into a 3-channel image. In Figure 14, the first image P S This is converted to the red image R, and the second image P V This is converted to a green image G, and the third image P B This is converted to the blue image B. However, it is not limited to this, and the first image P S This is converted to image G or image B, and the second image P V This is converted to image B or image R, and the third image P B The image may be converted to image R or image G. Here, a 3-channel image is converted to red image R, green image G, and blue image B, but it is not limited to this. The image does not have to be a separate channel for each color.

[0070] Furthermore, the image generation unit 12B generates the third image P B In contrast, the image is binarized so that the building portion and the non-building portion can be distinguished. Binarization may be performed so that the building becomes white and the background becomes black, as shown in Figure 14, or the building becomes black and the background becomes white. Note that if the 3D map image P1 is a red 3D map image, then the image P of hue H H This results in an image with a brightness value of almost 0, and therefore lacks topographic information. However, the image generation unit 12B generates the third image P B and image P H The combined image may be input to the deep learning model 10 as, for example, image B shown in Figure 14.

[0071] As shown in Figure 13, the image generation unit 12B converts the map image P' to be estimated into a 3-channel composite image CP' and then inputs it to the deep learning model 10. Furthermore, the image generation unit 12B also converts the training map image T' and the adjustment image O' into a 3-channel training map image T'' and composite image OP', and then inputs them to the deep learning model 10. Note that when generating the composite images OP' and CP', the image generation unit 12B converts the third image P BAfter synthesizing the building image BI, a 3-channel color image (R, G, B) is generated. The image generation unit 12B differs from the image generation unit 12 in that it generates a training map image T'' based on the training map image T'.

[0072] The estimation process of estimation system 1B in Figure 13 is basically the same as the estimation process shown in Figure 6. However, it differs from estimation system 1 in Figure 1 in that the pre-training process in step S1 (Figure 15), the building image optimization process in step S2 (Figure 16), and the determination process in step S4 (Figure 17) are performed by the image generation unit 12B.

[0073] As shown in Figure 15, in the pre-training process, an image generation process is performed between steps S12 and S13 (step S50). That is, in the pre-training process, after the training image set TR (map image T' + specified image R) is generated, the image generation process is performed before training the deep learning model 10 (step S50).

[0074] As shown in Figure 16, in the building image optimization process, the image generation process is executed instead of step S26 (generation of composite image) (step S50). In the image generation process, the building image BI is combined with the adjustment image O' (see Figure 13) to generate the composite image OP'.

[0075] As shown in Figure 17, in the determination process, the image generation process is executed instead of step S42 (generation of composite image) (step S50). In the determination process, the building image BI is combined with the map image P' to be estimated (see Figure 13) to generate the composite image CP'.

[0076] As shown in Figure 18, in the image generation process, the image generation unit 12B performs HSV conversion on the 3D map image (red 3D map) P1 to generate the first image P S , second image P V (Step S51) generates the third image P, which is a grayscale image of only the building. Next, the image generation unit 12B generates the third image P BThe image generation unit 12B then generates the image (step S52). Next, the image generation unit 12B determines whether or not it has been called from the pre-training process (step S53). If it has not been called from the pre-training process (step S53; No), the image generation unit 12B synthesizes the building image BI with the adjustment map image O' or the map image P' to be estimated to generate the synthesized image OP' and the map image CP' to be estimated (step S54).

[0077] If it is determined that the process is being called from the pre-training process (step S53; Yes), after step S54 is executed, the image generation unit 12B generates the first image P S , second image P V , third image P B Convert to a 3-channel image (step S55). At this time, the third image P B The image is then binarized. After step S55 is executed, the image generation unit 12B terminates the image generation process.

[0078] In estimation system 1, Mask R-CNN is used as the deep learning model 10, whereas in estimation system 1B, YOLO (You Only Look Once) v8 is used as the deep learning model 10.

[0079] A comparative evaluation of Estimated System 1 and Estimated System 1B was conducted using simulations. The evaluation results are described below.

[0080] The simulation used a red 3D map image, cropped into patch images equivalent to zoom level 15 (approximately 1.3km x 1.3km). The test data consisted of 167 locations in Nagasaki Prefecture: 84 locations where landslides occurred in non-landslide hazard zones, and 83 locations that were not steep slopes and did not pose a risk of landslides. For pre-training, the following number of image sets, with landslide hazard zones on steep slopes designated as risk zones, were used as training data. In the case of Mask R-CNN: 24,200 pairs (of which 1,306 pairs were used for validation) For YOLOv8: 24,418 pairs (of which 1,288 pairs were used for verification)

[0081] Figure 19 shows the detection accuracy (recall rate) of disaster locations in estimation system 1B when the spacing D of building E and the number of columns N and rows M of the virtual building E are changed. In estimation system 1, as shown in Figures 10(A) and 10(B), the optimal combination of number of columns N and rows M and spacing D was (number of columns N and rows M = 5, spacing D = 8 pixels), whereas in estimation system 1B, one of the optimal combinations of number of columns N and rows M and spacing D was (number of columns N and rows M = 5, spacing D = 7 pixels). This is thought to be because the deep learning models 10 are different in estimation system 1 and estimation system 1B.

[0082] [Estimated accuracy] Figure 20 shows the recall, precision, and F1 score for estimation systems 1 and 1B when using a map image P with the location of the landslide as the center of the image, and also shows the recall, precision, and F1 score for conventional deep learning models. Estimation system 1B is represented by a 3-channel YOLOv8 (number of columns N, number of rows M=5, interval D=7 pixels), and estimation system 1 is represented by a 9-channel Mask R-CNN (number of columns N, number of rows M=5, interval D=8 pixels). As shown in Figure 20, it was found that estimation systems 1 and 1B according to this embodiment show significant improvements in recall, precision, and F1 score compared to conventional systems. Furthermore, comparing estimation system 1 and estimation system 1B, it was found that estimation system 1B is superior in recall and F1 score.

[0083] Figure 21 shows the number of successful detections of the landslide risk area C in the estimation system 1B according to this embodiment, when the location of the landslide was displaced in the longitude and latitude directions from the center of the map image P. In Figure 21, as in Figure 12, the horizontal axis represents the displacement in the longitude direction from the disaster location at the center of the map image P (unit: ×10 -4 The graph shows the latitude (in degrees), with the vertical axis representing the latitudinal displacement from the disaster site at the center of map image P (unit: ×10). -4The graph shows the degree(s), and the value in each cell indicates the number of successful detections for the corresponding displacement in the longitude and latitude directions. As shown in Figure 21, the number of successful detections is 5 not only at the center of map image P, but also in the surrounding area including the center. The area with 5 successful detections is approximately 1.5 times wider in estimation system 1B shown in Figure 21 compared to estimation system 1 shown in Figure 12.

[0084] As described in detail above, according to the estimation method, estimation system 1, and program 30 of this embodiment, a landslide hazard area C is estimated based on a composite image CP in which a virtual building image BI is added to a part of the map image P that is the target of estimation, assuming that a building E exists in the target of estimation. Therefore, it is possible to accurately estimate landslide hazard areas C where there is a risk of collapse of steep slopes and debris flows occurring at any point, including points where there are no buildings nearby.

[0085] According to this embodiment, the building image BI is an image of a group of buildings composed of a two-dimensional arrangement of rectangles representing the outline of buildings E as seen from above. Using such an image for the building image BI results in higher accuracy in estimating the landslide hazard area C than adding an image of a group of buildings where buildings E are randomly placed. This is because an image in which buildings E are arranged regularly allows for a more uniform arrangement of buildings E without overlap on the image. Furthermore, it has been confirmed that adding the building image BI to a portion of the map image P that is the target of estimation, rather than adding it to the entire map image P, results in higher accuracy in estimating the landslide hazard area C.

[0086] The detection accuracy of the actual disaster location varies depending on the spacing D, number of columns N, and number of rows M of buildings E in the building image BI. Therefore, the spacing D, number of columns N, and number of rows M of buildings E are adjusted to achieve the best detection accuracy of the actual disaster location. This allows for accurate estimation of sediment disaster risk areas C, which are at risk of landslides, debris flows, and mudslides on steep slopes. In this embodiment, the number of columns N and number of rows M are set to the same number, but they may be adjusted separately. The size of buildings E may also be adjusted in the same way as the spacing D, number of columns N, and number of rows M of buildings E.

[0087] As mentioned above, the deep learning model 10 is not limited to Mask R-CNN. Other deep learning models 10 can be used as well, as long as they perform object detection and segmentation tasks, such as Faster R-CNN or YOLO (You Only Look Once). When YOLOv8 was adopted as the deep learning model 10, it was found that high recall and F1 score could be obtained while maintaining high precision.

[0088] Furthermore, by converting a multi-channel color image to a 3-channel color image, as in estimation system 1B, a general deep learning model 10 can be used to determine landslide hazard areas C without modifying the input interface to be multi-channel. Note that in estimation system 1B, the 3D map image P1 was a red 3D map, but this is not the only option. Even if an image other than a red 3D map is used as the 3D map image P1, it is sufficient if information about the relief of each point can be extracted from that image as two-variable parameters. For example, the amount of slope and the direction of slope can be considered. In this case, image transformations other than HSV conversion may be applied depending on the properties of the image.

[0089] By utilizing estimation systems 1 and 1B, it is expected that, for example, prefectures can streamline and expedite the process of designating landslide hazard zones C. Furthermore, it is expected that estimate systems 1 and 1B can be used by civil engineering consultants and engineers involved in land development, disaster prevention, and infrastructure development to assess the risk of landslides in undeveloped areas not designated as landslide hazard zones C.

[0090] The hardware and software configurations of estimated systems 1 and 1B are examples only and can be changed and modified as needed.

[0091] The core processing part of the estimation systems 1 and 1B, which consist of a CPU 21, an AI accelerator 22, main memory 23, external memory 24, an operation unit 25, a display unit 26, a communication interface 27, and an internal bus 29, can be implemented using a normal computer system, not a dedicated system. For example, the estimation systems 1 and 1B that perform the above processing may be configured by distributing a computer program for executing the above operations on a computer-readable recording medium (flexible disk, CD-ROM, DVD-ROM, etc.) and installing the computer program on a computer. Alternatively, the computer program may be stored in the memory of a server device on a communication network such as the Internet, and the estimation systems 1 and 1B may be configured by a normal computer system downloading it.

[0092] If the functions of estimated systems 1 and 1B are realized through a division of labor between the OS (operating system) and application programs, or through collaboration between the OS and application programs, then only the application program portion may be stored on a recording medium or storage unit.

[0093] It is also possible to superimpose a computer program onto a carrier wave and distribute it via a communication network. For example, a computer program could be posted on a bulletin board system (BBS) on a communication network and distributed via the network. This computer program could then be launched and executed under the control of the OS, similar to other application programs, thereby enabling the aforementioned processing.

[0094] This invention allows for various embodiments and modifications without departing from the broad spirit and scope of the invention. Furthermore, the embodiments described above are for illustrative purposes only and do not limit the scope of the invention. In other words, the scope of this invention is indicated not by the embodiments, but by the claims. Various modifications made within the scope of the claims and the equivalent scope of the meaning of the invention are considered to be within the scope of this invention. [Industrial applicability]

[0095] This invention is intended for use in estimating areas at risk of sediment-related disasters. [Explanation of Symbols]

[0096] 1, 1B Estimation system, 2 Memory unit, 10 Deep learning model, 11 Control unit, 12, 12B Image generation unit, 21 CPU, 22 AI accelerator, 23 Main memory, 24 External memory, 25 Operation unit, 26 Display unit, 27 Communication interface (I / F), 29 Internal bus, 30 Program, 50 Image conversion unit, B Bounding box, C Sediment disaster risk area, CP, CP' Composite image, D Interval, E Building, M Number of rows, N Number of columns, O, O' Adjustment image, OP, OP' Composite image, P, P' Map image, P1 3D map image, P2 Standard map image, P2', P2'' Map image, PR Estimation image, R Specified image, T, T', T'' Map image, TR Training image set

Claims

1. An estimation method performed by an estimation system for estimating landslide risk areas on a map image containing topographic information and building information, A training step involves training a deep learning model that uses a training image set, which is a set of training map images and images indicating landslide hazard risk areas on the map images, to input the map images and determine whether each point on the input map image is in a landslide hazard risk area. Image generation step: Generates a composite image in which a virtual building image is added to a part of the map image that is the target of estimation, A determination step in which the synthesized image is input to the deep learning model and it is determined whether or not each point on the map image that is the target of estimation is in a landslide hazard area, An estimation method that includes [this].

2. The aforementioned building image is an image of a group of buildings composed of a two-dimensional arrangement of figures representing the external outlines of the buildings as seen from above. The estimation method according to claim 1.

3. Before executing the image generation step, the deep learning model is input with the map image showing the actual location where the disaster occurred. The adjustment step includes adjusting the spacing between buildings, the number of rows, and the number of columns in the building image so that the detection accuracy is best when the location where the disaster actually occurred is determined to be a landslide hazard area. The estimation method according to claim 2.

4. The deep learning model described above is This is a model that performs object detection and segmentation tasks. For each pixel, a determination is made as to whether or not it is in a landslide hazard area. The estimation method according to claim 1.

5. In the aforementioned training step and the aforementioned image generation step, From the aforementioned map image, a first image representing the slope at each point, a second image representing the ridge-valley degree at each point, and a third image representing the presence or absence of buildings at each point are generated. The first image, the second image, and the third image are converted into a 3-channel image. The three-channel image is generated as an image to be input to the deep learning model. The estimation method according to claim 1.

6. The aforementioned map image includes a red relief map image having the aforementioned topographic information, The aforementioned red relief map image is subjected to HSV conversion to generate a first image in which the degree of slope at each point is represented by saturation, and a second image in which the degree of ridges and valleys at each point is represented by brightness. The estimation method according to claim 5.

7. An estimation system for estimating landslide risk areas on a map image containing topographic information and building information, A deep learning model that determines whether each point on the map image, which is the target of estimation, is in a landslide hazard area, A control unit that trains the deep learning model using a training image set which is a set of training map images and images showing landslide risk areas on the map images, The system includes an image generation unit that generates a composite image in which a virtual building image is added to a portion of the map image that is the target of estimation, The control unit, The synthesized image is input to the deep learning model, which determines whether each point on the map image, which is the target of estimation, is located in a landslide hazard area. Estimation system.

8. A computer that estimates landslide risk areas on map images containing topographic and building information, A training step involves training a deep learning model that determines whether each point on an input map image is in a landslide hazard zone, using a training image set which is a pair of a training map image and an image showing a landslide hazard zone on the map image. Image generation step: Generates a composite image in which a virtual building image is added to a part of the map image that is the target of estimation, A determination step in which the synthesized image is input to the deep learning model and it is determined whether or not each point on the map image that is the target of estimation is in a landslide hazard area, A program that executes the command.

Images