Vegetation management system and vegetation management method

Abstract

To efficiently formulate a shooting plan and a trimming plan in vegetation management.SOLUTION: A vegetation management system includes: a data acquisition section that acquires a satellite image of a power transmission line arrangement region; a site work situation collection section that collects a situation of a trimming work executed at a site of the power transmission line arrangement region; and a planning section that divides the power transmission line arrangement region into a plurality of partial regions, manages a status relating to the trimming work and the acquisition of the satellite image in association with each of the partial regions, and formulates a plan for the trimming work and a plan for shooting the satellite image on the basis of the status. The status includes a non-shooting status, a shooting status, a clear waiting status, and a cleared status. The planning section acquires an order of executing the trimming work on the plurality of partial regions by performing reinforcement learning by an agent of which an action is along a direction of progress of handling risk by trimming on the basis of the status of the plurality of partial regions.SELECTED DRAWING: Figure 9

Description

Technical Field

[0001] The present invention relates to a vegetation management system and a vegetation management method using satellite images.

Background Art

[0002] In power transmission and distribution lines mainly deployed in forested mountainous areas, accidents such as line breaks and fires may occur due to contact with vegetation, resulting in power outages. As shown in FIG. 1, the contact of vegetation with power transmission and distribution lines includes patterns where trees fall due to strong bad weather or diseases, patterns where branches and trunks are blown and entangled, and patterns where trees grow and hang on the power transmission lines. The patterns of falling and being blown are mainly due to short-term weather changes, while the pattern of trees growing and hanging on the power transmission lines is due to long-term tree growth.

[0003] Not only in Japan where natural disasters such as typhoons and floods occur frequently, but also in regions with dry climates where forests are prone to wildfires, such as North America, vegetation management is actively carried out. Once a power outage occurs, not only the cost of restoration but also the loss of electricity sales opportunities during that period becomes a problem. Therefore, active countermeasures are being taken for vegetation management around power transmission and distribution lines. The methods of vegetation management are also progressing digitization using drones and satellite images, and effects such as reduction of patrol costs, prevention of overlooking, and quantification of work volume and risk are expected.

[0004] From the perspective of risk quantification, Patent Document 1 discloses a method for creating a database for each tree and estimating the time when a tree will grow and come into contact with equipment based on the height of the equipment on top of the tree, and further, a method for setting the planned date and route for felling based on the estimated felling time of the tree and the location information of the tree. Specifically, Patent Document 1 states: "The timing of tree felling can be determined without inspecting the trees in advance." "The tree felling instruction device 1 comprises an input unit 2, a display unit 3, a processing unit 4, and a storage unit 5. The equipment data update unit 42 of the processing unit 4 updates the equipment data 61 in the storage unit 5 when it acquires data related to equipment changes from the input unit 2. The tree data update unit 43 updates the tree data 71 in the storage unit 5 when it becomes necessary to change the data due to equipment changes or when it acquires data related to tree felling from the input unit 2. The tree growth rate estimation unit 44 estimates the future tree growth rate for each tree type from the tree growth rate data 81 and weather data 91 in the storage unit 5. The felling time estimation unit 45 estimates the time when the tree will grow and come into contact with the equipment as the felling time, based on the estimated tree growth rate and the height of the equipment on the tree. The felling scheduled date and route setting unit 46 sets the scheduled felling date and route based on the estimated tree felling time and tree location information." However, developing a tree database itself is a very difficult task, and when managing geographically wide areas, methods using aerial imagery such as satellite imagery are being considered. Remote sensing technologies that capture and analyze images from the air to grasp macroscopic vegetation trends have also been disclosed. [Prior art documents] [Patent Documents]

[0005] [Patent Document 1] Japanese Patent Publication No. 2008-003855 [Overview of the Initiative] [Problems that the invention aims to solve]

[0006] Combining a method of formulating logging plans after understanding vegetation growth risks with a technology that uses aerial imagery to assess forest volume makes it seem possible to assess vegetation growth risks and plan the logging sequence using aerial images, as shown in Figure 2. However, acquiring aerial images itself incurs costs and takes time, and there is a challenge in that it is not possible to create logging plans while always keeping all information on the forest area under management up to date.

[0007] While paid high-resolution satellite imagery allows for the diagnosis of vegetation contact from above, the area captured in a single image is limited, requiring multiple acquisitions. New high-resolution satellite image acquisition is performed by requesting tasking from the operating company, and the image is taken within approximately two weeks when the specified conditions, such as cloud cover and elevation angle, are met. Therefore, updating the entire controlled area to the latest data takes a considerable amount of time.

[0008] Summer, when vegetation risks are highest, is the ideal time for risk assessment using satellite imagery and also a good opportunity for logging. The rainy season precedes summer, during which both imagery and logging are halted. Since the optimal start times for both satellite imagery and logging coincide, it is necessary to proceed with satellite imagery and logging simultaneously. If logging plans are made after satellite imagery is completed, logging contractors can be cut off from unnecessary areas. However, if logging work is carried out in areas that are still being newly imagered, the pre-emptive risk assessment work through imagery becomes useless. Moreover, since the true current risks can be identified in the areas where imagery has been taken, there may be areas that require action even if it means changing the logging plan.

[0009] When planning a shooting schedule, one approach is to refer to past contact risk assessment results and use satellite imagery for confirmation. However, prediction errors are inevitable, so a method is needed to dynamically reflect the actual vegetation risk, which becomes clear sequentially through partial or complete shooting of the area, into the logging plan.

[0010] Thus, when managing vegetation, it is necessary to formulate a base plan without overlapping photography and logging, and to dynamically modify the plan in response to the risks that are identified as they arise. Therefore, the present invention aims to improve the efficiency of photography planning and logging planning in vegetation management. [Means for solving the problem]

[0011] To achieve the above objective, one representative vegetation management system of the present invention is a vegetation management system for managing vegetation around a power transmission facility, comprising: a data acquisition unit that acquires satellite images of the power transmission line area; a field work status collection unit that collects the status of logging work carried out on site in the power transmission line area; and a planning unit that divides the power transmission line area into a plurality of sub-regions, manages the status of the logging work and the acquisition of satellite images in association with each sub-region, and formulates a plan for the logging work and a plan for taking satellite images based on the status, wherein the status includes an uncaptured status where satellite images have not been taken, a capturing status where satellite images are to be acquired, a clearing waiting status where it has been evaluated that there is a risk that needs to be cleared by the logging work, and a cleared status indicating that the risk has been cleared regardless of whether the satellite images have been taken or not, and the planning unit is characterized in that it acquires an order for carrying out the logging work for the plurality of sub-regions by performing reinforcement learning with an agent whose action is the direction of progress in dealing with the risk by logging based on the status of the plurality of sub-regions. Furthermore, one representative vegetation management method of the present invention is a vegetation management method for managing vegetation around a power transmission facility, wherein the vegetation management system includes at least the step of obtaining satellite images of the power transmission line area; a planning step of dividing the power transmission line area into a plurality of sub-areas, managing the status of logging work and acquisition of the satellite images in association with each sub-area, and formulating a plan for logging work and a plan for taking satellite images based on the status; and an update step of updating the plan for logging work and the plan for taking satellite images to reflect the status of logging work carried out on site in the power transmission line area. The status includes an uncaptured status where satellite images have not been captured, a capturing status where satellite images are to be acquired, a clearing status where risks that need to be cleared by logging operations have been assessed, and a cleared status indicating that the risks have been cleared regardless of whether satellite images have been captured or not. The planning step and the update step are characterized in that, based on the status of the multiple sub-regions, reinforcement learning is performed by an agent whose action is the direction of progress in addressing risks through logging, thereby obtaining an order in which to perform the logging operations on the multiple sub-regions. [Effects of the Invention]

[0012] According to the present invention, photography plans and logging plans for vegetation management can be efficiently formulated. Other problems, configurations, and effects will be clarified by the following description of embodiments. [Brief explanation of the drawing]

[0013] [Figure 1] Diagram illustrating the problem of vegetation contacting power transmission and distribution lines. [Figure 2] A diagram illustrating the optimization of logging plans by combining aerial images. [Figure 3] A diagram illustrating the overall invention. [Figure 4] Diagram illustrating the method of cutting the grid according to the present invention. [Figure 5] A diagram illustrating how to plan the shooting schedule of the present invention. [Figure 6]Schematic diagram of the update of the shooting and logging plan of the present invention [Figure 7] Diagram for explaining the grid management of the present invention [Figure 8] Diagram for explaining the environment and reward setting of reinforcement learning of the present invention [Figure 9] Functional block of the present invention [Figure 10] Hardware configuration [Figure 11] Flowchart of reinforcement learning

Mode for Carrying Out the Invention

[0014] The embodiments of the present invention will be described with examples. In each embodiment, blocks, hardware, and processes with the same numbers basically perform the same operations, so the descriptions are omitted.

Examples

[0015] The flow of optimizing the logging plan from the satellite image imaging of the present invention will be described using FIG. 3. In this embodiment, long-term maintenance aimed at clearing the risk of power line contact due to vegetation growth is the main target. The overhead height of the power line is high and the vegetation grows slowly, so there is a margin of time for contact compared to the distribution line, but once a fire occurs, the affected area is extensive. In North America where the land area is large, the area where the power lines are arranged is extensive, but in order to reduce the risk of power outages caused by vegetation, the entire area is periodically maintained. The power line is installed at a higher position than the distribution line, and since the vegetation grows slowly, a schedule for maintaining the area over several years is often arranged.

[0016] The vegetation management system collects aerial images, 3D data, weather data, soil data, GIS data such as the arrangement information of power transmission and distribution lines and past power outage locations (S1001). Based on these data, the vegetation management system generates a vegetation database around the power line by image processing (S1002). The information on the locations where power outages occurred in the past represents the susceptibility to power outages in that area macroscopically.

[0017] While it is difficult to keep aerial images constantly up-to-date, if images have been taken even once in the past, the vegetation management system uses those images, weather data, and soil data to predict future growth and forecast contact risks at the present and future timings of planned logging (S1003). Because this contact risk is based on images, it functions as a predicted risk map.

[0018] Flows S1001 to S1003 can be prepared before the shooting plan is created and can be done asynchronously with the execution of the plan. The method for generating the predicted risk map involves detecting vegetation using aerial images taken in the past and predicting the changes over time from that point in time. Vegetation detection can be done at the forest level or by managing vegetation individually. When managing vegetation individually, more detailed data is required than for forest areas. If only forest areas are used, accurate estimation can be achieved using semantic segmentation, etc. For individual vegetation management, techniques for separating individual vegetation by detecting tree canopies are required, using 3D data or aerial images taken at multiple times.

[0019] Once a tree canopy is detected, spectral values ​​can be extracted from the center of the canopy and compared with a plant spectral library to separate it into individual tree species. Managing data on individual trees has several advantages compared to managing forests on a domain basis. For example, information on tree species and age can be managed. Tracking tree species, age, soil, and weather information allows for accurate prediction of future growth and leads to a more accurate understanding of vegetation contact risk.

[0020] When evaluating contact risk, summer is the time of year when the risk is highest, so taking photographs in summer allows for accurate risk prediction. Winter images can also be useful for classifying tree species if leaf fall and leaf color are detected, but this is only for supplementary purposes, and the priority of photographing decreases after the tree species and location have been identified. Summer arrives after the rainy season, making logging easier, and it is conceivable that photography and logging will proceed simultaneously. To avoid wasting photographs, a means of avoiding overlap with logging is necessary. Based on this predicted risk map, a photography plan and a logging plan are formulated (S1004). In the system of the present invention, the method for formulating the photography plan is acquired through reinforcement learning. The method of executing reinforcement learning will be explained later.

[0021] When planning the imaging schedule, it is advisable to coordinate it with the tasking method for new satellite imaging. For example, as shown in Figure 4, new satellite image acquisition is performed along the north-south direction. This depends on the satellite's orbit, so it is advisable to create and manage a grid along the orbital direction of the satellite images to be used. The grid size should be the satellite's observation width (Swath) as the size of one side. Alternatively, this can be set as the maximum value, and the width of each of the multiple divisions can be used as the size of one side.

[0022] Satellite imaging modes include long-strip observation, multi-strip wide-area observation, and stereo observation. The long-strip mode, which continuously takes images directly below the target, has high resolution and is suitable for detailed assessment of contact risk. Therefore, it is best to use the long-strip mode as the base mode and, if necessary, combine it with two-strip imaging. For each grid, it is best to manage the status of: not yet photographed and contact risk not cleared (risk unknown), during photography, contact risk identified in multiple stages but not yet cleared (i.e., waiting for clearing), and cleared regardless of the photography status. The vegetation management system uses reinforcement learning to acquire the optimal actions of agents when given the managed status for each grid.

[0023] The vegetation management system of the present invention operates by using the action guidelines of the acquired agents, receiving progress reports on on-site work, and sequentially updating the photography and logging plans while synchronizing the risk clearance status and the photography status (S1005).

[0024] Since reinforcement learning generally acquires behaviors through simulations that repeatedly reproduce the situation, we prepare a logging simulator that generates observation spaces for a status map to manage the risk clearance status of power lines, a predicted risk map, and a location management map to manage the location of logging agents, as shown in Figure 6.

[0025] Figure 7 is an explanatory diagram of the logging simulator. The logging simulator is one of the functions of the image capture and logging planning unit 104, which will be described later. The satellite image capture planning unit within the logging simulator is expected to behave almost identically in the simulator as in the actual operation. Figure 11 shows the flow of performing reinforcement learning using the logging simulator. First, the logging simulator determines the number of episodes to be used for reinforcement learning, and then determines how many steps to take within each episode and executes a loop. The steps may be taken in units of one day, or the unit time may be adjusted according to the period required for satellite image capture and the actual progress speed of the work. At each step, the logging simulator determines whether new satellite image capture has been completed (S1101).

[0026] The logging simulator receives satellite images once new imaging is complete. Grid columns that have already been imaged should be removed from the imaging plan for that year. Within the newly imaged area, the vegetation around power lines is analyzed, and the process is retried based on the predicted risk locations as described in procedure S1003, updating the predicted risk map as needed (S1102).

[0027] Next, the logging simulator updates the shooting plan based on the risk prediction results (S1103). The shooting plan is selected based on the predicted risk map, selecting one column along the satellite's direction of travel (north-south), and changing the status of the unchecked squares (grid) in that column to "SHOOTING". The "SHOOTING" status is maintained for approximately two weeks. These two weeks represent the tasking time for new satellite imagery, and the number may vary depending on the specifications of the satellite being ordered. After the period ends, the logging simulator updates the status map based on the risk assessment values ​​obtained from the image analysis.

[0028] Figure 5 illustrates how to prioritize the imaging plan. Several methods can be considered for prioritizing imaging. Simply put, consider the satellite's direction of travel as the column direction, and based on the risk prediction map, count the squares with a high contact risk. Prioritize imaging in columns with many squares determined to be high risk. If the numbers are the same, count the number of squares with medium risk and low risk, and prioritize imaging in columns with a larger total number. If the logging start point is known, it is best to avoid overlap and start imaging from multiple columns away from the location.

[0029] There are several other ways to prioritize. You can prioritize the column with the most grids in the "awaiting completion" status. Alternatively, you can prioritize the column that divides consecutive blocks of grids in the "awaiting completion" status. This is because dividing the grids in the "awaiting completion" status into multiple blocks and performing the logging work on each block can make the logging work more efficient. You can also prioritize the column that contains power line squares that are split into two segments in a single shot. This is effective when the strategy is to complete the shooting of one power line square first when the power line branches in two directions.

[0030] In conjunction with the characteristics of such satellite imagery, the logging simulator acquires a logging plan using reinforcement learning (S1104). The logging simulator divides the region into a grid and defines the squares that need to be visited based on the actual location information of power lines, and the agent acquires a logging action plan using reinforcement learning. The imaging plan is based on vegetation growth risk information predicted from past data, but there is a need for a method to dynamically reflect uncertainty due to prediction errors and actual vegetation risks that become apparent sequentially from new imaging into the logging plan. Reinforcement learning is expected to be effective in addressing this.

[0031] As input, the agent is given the three channels (status map, predicted risk map, and agent position information) output from the logging simulator as an observation space. The agent is given the three-channel observation space over one or more time points, and learning progresses by rewarding each of the agent's possible action spaces. The agent's action space refers to, for example, which direction it moves on the grid, i.e., the direction of logging. It is best to define a discrete action space where it moves up, down, left, and right. Of course, it is also possible to allow diagonal movement and have eight directions. When multiple observation spaces are provided, it is best to provide action spaces such as how many times the logging action in each direction of movement is performed consecutively. When the agent steps on a square, the risk is cleared, and on the status map, it changes to cleared = CHECKED (value of the status square: 0).

[0032] Furthermore, the logging simulator handles time-series information and updates the risk clearance status map, as shown in Figure 8 (S1105), reflecting the satellite imaging plan and results, as well as the agent's execution results. The logging simulator is intended to be updated hourly and daily.

[0033] Initially, the locations that logging agents should visit are set to "Not Checked" and "Photon incomplete" (NOT_CHECK). This is determined by the intersection of power line location data and a grid divided according to the high-resolution satellite's orbit. Areas without power lines are given a status of -2, areas being photographed are given a status of -1, areas where the risk has been cleared are given a status of 0, areas that are not cleared and photon incomplete are given a status of 1, and locations where new satellite imagery has been taken and the immediate risk has been identified are given a status of 2-5, classified according to the risk level. When the risk assessment value is lower than low risk (RISK_L), meaning that there is clearly no risk, it can be considered that the risk has been cleared without the logging agent having to visit.

[0034] These are expected to be controlled by risk assessment thresholds. To visualize the simulator's operation, a graphical user interface may be provided that visualizes a status map by performing color rendering for each of the aforementioned states, and displays it together with a logging agent location management map.

[0035] In reinforcement learning, it is necessary to acquire actions that can respond to future events, but it is not possible to obtain future correct data. Therefore, the following approach can be taken. If a large amount of satellite imagery is already available, one possible method is to predict the risk value based on a past point in time, and treat images newer than that point in time as new images. This allows the information from images available at that time to be used to reflect the actual risk prediction value in the training.

[0036] Alternatively, during training, it is necessary to learn about future events under the assumption that there is an error between the actual risk value obtained by analyzing newly acquired images and the risk value predicted from past acquired images. If future risk predictions were perfect, a simple optimization plan of the acquisition order would suffice, but since the accuracy of prediction algorithms is almost never perfect, uncertainty exists, and the aforementioned error must be taken into account. To simulate future risks, during training, a predetermined variation is added to the risk prediction value as a substitute. In other words, it is best to learn means to deal with the discrepancy between the information on immediate risks revealed by new acquisitions and the values ​​in the risk prediction map. In the execution phase of the plan, it is not necessary to add the above error to the risk prediction map.

[0037] The actions an agent learns during the reinforcement learning phase are highly dependent on the reward design. The goal is to optimize the logging order through coordinated actions with satellite imagery, and for example, the rewards shown in Figure 8 are appropriate. A negative reward is given when the agent passes through an area without power lines (NO_TX_LINE, state cell value: -2) or a cell that is currently tasked with imagery (SHOOTING, state cell value: -1).

[0038] A reward is given when visiting a location that is not cleared and photography is incomplete (NOT_CHECK, status square value: 1), or when moving to a risk-identifying square (RISK_L, RISK_M, RISK_H, RISK_VH, status square value: 2-5). The method for dividing the risk-identifying squares should have a mechanism that allows adjustment of the number of divisions and threshold parameters.

[0039] When passing through a CHECKED square that has already been cleared of risk, no new reward is generated. However, the reward decreases over time, aiming to reduce the number of times the agent passes through the same location. Since satellite imagery is performed in the north-south direction (column direction), it is desirable for the risk-cleared squares to also fill in in the column direction. This is because if all the squares in a column before satellite imagery are risk-cleared, there is no need to photograph that column, making photography unnecessary and thus cost-effective. Therefore, a slightly larger reward is given when an entire vertical column changes to CHECKED, and when all NOT_CHECK squares change to CHECKED, one episode of reinforcement learning is completed and a large reward is given (S1106). Reinforcement learning is then performed so that the total reward obtained in each episode is high. However, no reward is given if the risk is diagnosed as below a predetermined value through imagery and the square automatically becomes CHECKED, rather than being handled by a logging agent.

[0040] Reinforcement learning agents learn either value-based, policy-based, or both. To manage the number of states, which is proportional to the combination of observation and action spaces, deep reinforcement learning approaches incorporating neural networks are used when dealing with large spaces. The DQN (Deep Q-Network), which has a proven track record in human-to-human Go, was the starting point, and advancements are progressing. Because computation becomes difficult to converge as the observation space expands, it is best to limit reinforcement learning to an area that can be physically managed within one year. The observation space should be limited to a size where, if functioning correctly, one episode can be completed in a few hundred steps. The process should be terminated after repeating episodes and confirming the acquisition of a predetermined reward.

[0041] Figure 9 shows the functional blocks for executing an actual photo-taking and logging plan using the reinforcement learning agent obtained in this flow. The vegetation management system includes an analysis data acquisition unit 101, a risk prediction unit 102, a photo-taking request unit 103, a photo-taking and logging plan unit 104, a display data generation unit 105, and a field work status collection unit 106.

[0042] The analysis data acquisition unit 101 works in conjunction with the image request unit 103 to acquire satellite images and download data such as weather, soil, and topography data used for growth prediction, as well as GIS data of power outage history and power line locations used for risk prediction.

[0043] The risk prediction unit 102 is responsible for calculating the risk assessment value at each status map location. By inputting this data into the logging planning unit 104, a logging plan is output using the results of reinforcement learning.

[0044] The logging plan presented to the user is processed by the output unit, the display data generation unit 105, and presented as data on the patrol order for each grid or as data on the patrol order for each line converted for each section of power lines within the grid.

[0045] To advance the logging plan and the optimization plan simultaneously, on-site information needs to be updated in a timely manner. For example, it would be good to have an app that allows users to upload on-site work status, or to create standardized electronic reporting forms, and have on-site logging workers input information daily. The function that receives this input is the on-site work status collection unit 106. The on-site work status collection unit identifies the work content reported by the workers and provides the information necessary for updating the status of the logging photography plan unit 104.

[0046] In the execution phase of the learning results, the photography and logging planning unit 104 outputs the next plan using the status of satellite photography during tasking, the growth risk predicted from past data, the work location of the logging agent, and the risk clearance status provided by the field work status collection unit 106, and provides it to the display data generation unit 105.

[0047] As explained in Figure 7, the imagery and logging planning unit 104 includes a logging simulator and an imagery and logging planning result output unit. The logging simulator includes a risk update unit, a satellite imagery planning unit, and a status update unit. The risk prediction unit 102 provides the risk prediction results to the risk update unit and the satellite imagery planning unit. In addition, the results of actions by the agent (movement and risk clearing actions) are input to the risk update unit and the status management unit.

[0048] The Risk Update Unit updates the risks based on the predicted risk results and the results of the actions taken, and outputs the results to the Status Management Unit. The Satellite Imagery Planning Unit formulates a satellite imagery plan based on the predicted risk results, and outputs the results to the Status Management Unit.

[0049] The status management unit updates the observation space (status map, predicted risk map, agent location information) and sets rewards using updated risks, satellite imagery plans, and agent action results. Based on the updated observation space and rewards, agents take the following actions. Furthermore, the status management unit outputs the shooting plan and the logging plan derived from the updated observation space to the shooting and logging plan result output unit. The shooting and logging plan result output unit outputs the shooting plan and logging plan to the display data generation unit 105.

[0050] The display data generation unit 105 generates and outputs display data for the shooting plan and the logging plan as appropriate. Therefore, the user can confirm both the logging plan, which is based on the shooting plan, and the shooting plan, which is based on the logging plan.

[0051] Figure 10 shows the hardware configuration for realizing the present invention. The CPU (Central Processing Unit) 413 starts the program of the present invention and executes the processing of the functional block shown in Figure 9. The data used for calculations is temporarily stored in memory 415. If high-speed calculations are required, the data may be transferred to the memory of the GPU (Graphics Processing Unit) 414, and the GPU's calculation support may be used. In particular, when using deep reinforcement learning, it is advisable to receive support for parallel calculations using the GPU.

[0052] The output data of the logging plan is displayed on the display device 411. Based on the displayed results, the user can decide on the logging plan and then place actual image orders, etc., using the user interface 416. The acquired data and predicted risk map can be managed on the storage device 412. The communication interface may be used when configuring the system with multiple computers or to execute actual calculation processing on an instance on the cloud. This server may be provided as an on-premises server or implemented on a public cloud. When implemented on a public cloud, the CPU 413, GPU 414, and memory 415 may be dynamically allocated and scaled according to the calculation scale. When using external computing resources, communication is done via the communication interface 417. The user also has terminal hardware that configures the computer in a similar manner. Progress of on-site work is uploaded from the terminal via the communication interface 417. The above process provides a method for optimizing logging plans.

[0053] As described above, the disclosed vegetation management system is a vegetation management system for managing vegetation around power transmission facilities, and comprises: an analysis data acquisition unit 101 as a data acquisition unit that acquires satellite images of the power transmission line layout area; a field work status collection unit 106 that collects the status of logging work carried out on site in the power transmission line layout area; and a photography logging plan unit 104 as a planning unit that divides the power transmission line layout area into a plurality of sub-regions, manages the status of the logging work and the acquisition of satellite images in association with each sub-region, and formulates a plan for the logging work and a plan for taking satellite images based on the status. The status includes an "uncaptured" status where satellite images have not been captured, a "captured" status where satellite images are to be acquired, a "waiting for clearing" status where risks that need to be cleared by logging operations have been assessed, and a "cleared" status indicating that the risks have been cleared regardless of whether satellite images have been captured or not. Based on the status of the multiple sub-regions, the planning unit acquires an order for performing the logging operations on the multiple sub-regions by performing reinforcement learning with an agent whose action is the direction of progress in addressing risks through logging. This configuration and operation allows for the efficient development of photography and logging plans for vegetation management.

[0054] Furthermore, the aforementioned sub-region is a grid demarcated by the orbital direction of the satellite used to capture the satellite image and the width direction perpendicular to the orbital direction, and the size of the grid in the width direction corresponds to the observation width of the satellite. Then, in formulating the plan for the logging operation, the planning unit sets a reward for all of the grids forming the circular rows to reach the cleared status before they reach the shooting status, and performs reinforcement learning. In formulating the plan for the shooting operation, the planning unit excludes rows where all grids have reached the cleared status due to the logging operation. Therefore, high-resolution satellite imagery can be effectively used to formulate imaging plans and logging plans.

[0055] Furthermore, the planning unit performs reinforcement learning by setting rewards corresponding to the risks determined by the analysis of newly captured satellite images. Furthermore, if the region of the "shooting in progress" status and the region of the "logging work" overlap, a penalty is set and the reinforcement learning is performed. In this way, by appropriately setting compensation, it is possible to formulate a logging plan that takes into account the relationship with the satellite image acquisition plan.

[0056] The system also includes a risk prediction unit 102 that diagnoses the contact risk, which is the risk of vegetation coming into contact with the power transmission facility, using at least the number of past power outages or satellite images taken in the past, and generates a predicted risk map that reflects growth over time in the risk assessment results, and the planning unit may formulate a plan for the logging work using the predicted risk map. This configuration allows for the effective use of past data to formulate both photography and logging plans.

[0057] Furthermore, when satellite images of the shooting results are obtained for a portion of the area in the shooting status, the planning unit evaluates the risk in multiple stages based on the satellite images and updates the status of the portion of the area that is evaluated as having a risk that requires clearing by the logging work to "awaiting clearing". This configuration allows for the development of long-term, detailed logging plans that address multiple levels of risk.

[0058] Furthermore, the planning unit may update the status of the sub-regions with sufficiently low risk to "cleared." This configuration reduces the number of grids where actual logging work is performed, thereby making logging operations more efficient.

[0059] When formulating a shooting plan, the planning department can select the row with the most high-risk grids as a candidate for the next new shooting session. Alternatively, the planning unit may select the column with the most grids in the "awaiting completion" status as the next candidate for new shooting. Alternatively, the planning unit may formulate a plan to designate a column that divides a block containing consecutive blocks with the "awaiting completion" status as a candidate for the next new shoot. In this way, by using appropriate indicators, it is possible to formulate an imaging plan that takes into account the relationship with the satellite image acquisition plan.

[0060] Furthermore, when the planning unit simulates the actual risks revealed by satellite images during reinforcement learning, it may add the error component of the vegetation growth prediction model as a risk prediction error. This configuration allows for risk forecasting by effectively utilizing the errors in the growth forecast model.

[0061] Furthermore, the system may also include a display data generation unit 105 as an output unit that outputs the plan for the logging work formulated based on the plan for the logging work, and the plan for the logging work formulated based on the plan for the logging work. This configuration allows users to review the logging plan and the photography plan, which mutually influence each other.

[0062] It should be noted that the present invention is not limited to the embodiments described above, and various modifications are included. For example, the embodiments described above are explained in detail to make the present invention easier to understand, and are not necessarily limited to those having all the configurations described. Furthermore, it is possible to replace or add configurations, not just delete them. [Explanation of Symbols]

[0063] 101: Analysis data acquisition unit, 102: Risk prediction unit, 103: Filming request unit, 104: Filming and logging plan unit, 105: Display data generation unit, 106: Field work status collection unit, 411: Display device, 412: Storage device, 413: CPU, 414: GPU, 415: Memory, 416: User interface, 417: Communication interface

Claims

1. A vegetation management system for managing vegetation around power transmission facilities, A data acquisition unit that obtains at least satellite images of the power line layout area, A field work status collection unit collects information on the status of logging work carried out at the site in the aforementioned power transmission line installation area, A planning unit divides the power line layout area into multiple sub-regions, manages the status of the logging work and the status of satellite image acquisition in association with each sub-region, and formulates a plan for the logging work and a plan for satellite image acquisition based on the status. Equipped with, The aforementioned status includes an uncaptured status where satellite images have not been captured, a capturing status where satellite images are being awaited to be acquired, a clearing status where it has been assessed that there is a risk that needs to be cleared by the logging work, and a cleared status indicating that the risk has been cleared regardless of whether satellite images have been captured or not. The vegetation management system is characterized in that the planning unit acquires an order for performing the logging work on the multiple sub-regions by performing reinforcement learning with an agent that has the direction of progress in addressing risks through logging as an action, based on the status of the multiple sub-regions.

2. A vegetation management system according to claim 1, The vegetation management system is characterized in that the aforementioned sub-region is a grid demarcated by the orbital direction of the satellite used to capture the satellite image and the width direction perpendicular to the orbital direction, and the size of the grid in the width direction corresponds to the observation width of the satellite.

3. A vegetation management system according to claim 2, The aforementioned planning department, In formulating the plan for the logging operation, reinforcement learning is performed by setting a reward for all of the grids forming the circular row to reach the "cleared" status before they reach the "shooting" status. The aforementioned planning department, A vegetation management system characterized in that, when formulating the plan for the aforementioned photography, rows in which all grids have been cleared by the aforementioned logging work are excluded.

4. A vegetation management system according to claim 1, The vegetation management system is characterized in that the planning unit performs reinforcement learning by setting a reward corresponding to the risk determined by the analysis of newly captured satellite images.

5. A vegetation management system according to claim 1, characterized in that a penalty is set when the partial area of ​​the "taking in progress" status and the partial area of ​​the "cutting work" overlap, and the reinforcement learning is performed.

6. A vegetation management system according to claim 1, The system further comprises a risk prediction unit that diagnoses the contact risk, which is the risk of vegetation coming into contact with the power transmission facility, using at least one of the number of past power outages or satellite images taken in the past, and generates a predicted risk map that reflects growth over time in the risk assessment results. The vegetation management system is characterized in that the planning unit formulates a plan for the logging operation using the predicted risk map.

7. A vegetation management system according to claim 1, The vegetation management system is characterized in that, when satellite images of the shooting results are obtained for a portion of the area in the shooting status, the planning unit evaluates the risk in multiple stages based on the satellite images and updates the portion of the area that is evaluated as having a risk requiring clearing by logging work to a clearing waiting status.

8. A vegetation management system according to claim 7, The vegetation management system is characterized in that the planning unit updates the status of the sub-areas with sufficiently low risk to a cleared status.

9. A vegetation management system according to claim 3, The aforementioned planning unit formulates a plan to designate the row with the most high-risk grids as the next candidate for new photography.

10. A vegetation management system according to claim 3, The vegetation management system is characterized in that the planning unit formulates a plan to designate the column with the most grids in the "awaiting clear" status as the next candidate for new photography.

11. A vegetation management system according to claim 3, The vegetation management system is characterized in that the planning unit formulates a plan to designate a column that divides a block with consecutive "awaiting completion" statuses as a candidate for the next new photo shoot.

12. A vegetation management system according to claim 1, The vegetation management system is characterized in that, when the planning unit simulates the actual risks revealed by satellite images during reinforcement learning, it adds the error component of the vegetation growth prediction model as a risk prediction error.

13. A vegetation management system according to claim 1, further comprising an output unit that outputs a plan for logging work formulated based on the plan for logging work and a plan for logging work formulated based on the plan for logging work, respectively.

14. A vegetation management method for managing vegetation around power transmission facilities, The vegetation management system The steps include at least obtaining satellite images of the power line layout area, A planning step involves dividing the power line installation area into multiple sub-regions, managing the status of logging operations carried out at the site of the power line installation area and the status of satellite image acquisition in association with each sub-region, and formulating a plan for logging operations and a plan for satellite image acquisition based on the status. The update step includes updating the plan for the logging work and the plan for taking satellite images in accordance with the status of logging work carried out at the site in the area where the power lines are installed, The aforementioned status includes an uncaptured status where satellite images have not been captured, a capturing status where satellite images are being awaited to be acquired, a clearing status where it has been assessed that there is a risk that needs to be cleared by the logging work, and a cleared status indicating that the risk has been cleared regardless of whether satellite images have been captured or not. A vegetation management method characterized in that the planning step and the updating step acquire an order for performing the logging work on the multiple sub-regions by performing reinforcement learning with an agent whose action is the direction of progress in addressing risks through logging, based on the status of the multiple sub-regions.

Images