Remote Sensing Task Generation System Based on News Text
By using a remote sensing mission generation system based on news text, combining deterministic computation and semantic understanding, the system solves the problems of accuracy and consistency in target coordinate acquisition in satellite remote sensing missions, realizes automated and reproducible remote sensing mission generation, and reduces the risk of observation failure and resource waste.
Patent Information
- Application Number
- CN202611336505.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-31
- Publication Date
- 2026-11-17
AI Technical Summary
In existing satellite remote sensing missions, target coordinate acquisition suffers from problems such as distortion of regional boundaries, directional offset, lag in the position of moving targets, lack of judgment on mission timeliness, and coordinate illusion in the output of large language models, resulting in waste of observation resources and high failure rate.
A remote sensing task generation system based on news text is adopted. Through the collaborative work of retrieval tools, news text extraction tools, large language models, location transformation tools, and task assembly tools, the system achieves the automatic conversion from unstructured news text to structured remote sensing tasks. Combined with deterministic computation and semantic understanding, it avoids the illusion risk of direct output from large language models.
It has achieved full-process automation, result consistency and reproducibility of remote sensing tasks, improved the accuracy of regional shape representation, reduced the risk of observation failure and resource waste, and improved task generation efficiency and success rate.
Smart Images

Figure CN123115541A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a remote sensing mission generation system based on news text, belonging to the field of satellite remote sensing mission planning and geographic information processing technology. Background Technology
[0002] Currently, target coordinate acquisition in satellite remote sensing missions typically employs the following two basic schemes:
[0003] Basic Solution 1: Manual Interpretation and Generation of Task Coordinates. Task planners manually read and determine the target location from news or publicly available information, then manually find the coordinates using map tools or manually draw the area boundaries, and finally manually enter the coordinates into the task planning system.
[0004] Basic Solution Two: Direct Coordinate Generation from Large Model, a simple combination of large model and map API. The large language model reads the news text and directly outputs the target's latitude and longitude, or it directly calls the geocoding API (a service interface that converts place name text into latitude and longitude coordinates; common commercial implementations include Amap Web API and Google Maps Geocoding API) provided by the map vendor to obtain the coordinates, which are then directly distributed as remote sensing task coordinates without further processing.
[0005] The drawbacks of the above method include:
[0006] 1) Distortion of Regional Boundary: Existing methods typically use the center point returned by geocoding as a reference when generating regional target coordinates, constructing a fixed-size bounding box as an approximate representation of the region. However, the actual shape of regional boundaries is often highly irregular. When using the bounding box approximation, parts of the rectangle will fall outside the target area, wasting observation resources and potentially causing missed measurements because the rectangle does not completely cover the true boundary. If the accuracy is improved by directly querying official administrative boundary data from map services, the number of original coordinate points in this boundary data usually reaches thousands to tens of thousands, far exceeding the processing limit of the satellite mission planning system for regional boundary coordinate points, making it unusable directly. This method cannot automatically achieve a balance between the mutually constraining indicators of boundary shape fidelity and the number of boundary coordinate points.
[0007] 2) Defects in Generating Directional / Offset Regions: News texts often describe regions with directional or positional offset semantics, such as "northeast." This means the target is neither a complete administrative division nor a single coordinate point, but rather a sub-region within a known area or a newly formed region offset along a specified direction from a known reference point. Existing technologies either divide the known complete area into sub-regions along directions like east, west, south, and north; or generate new regions by offsetting a known reference point along a specified direction and distance. Current solutions rely on manual drawing of region boundaries in map software based on experience. This results in inconsistent drawing outcomes among different operators and cannot be batch-processed or automated.
[0008] 3) Lag Target Positioning Defect: For targets with autonomous movement capabilities, existing solutions typically use the last observed position of the target as reported in news articles as the satellite's observation coordinates for imaging missions. However, there is usually a time interval of several hours or even longer between the news release and the actual execution of the observation mission. Since such targets possess a certain speed, their actual position may have deviated by tens or even hundreds of kilometers from the reported position within this time interval. Existing technologies do not incorporate information such as the target's estimated speed, the time difference between the observation time and the mission generation time, and whether the direction of movement is known to dynamically estimate the range of possible current positions of the moving target. This results in a high imaging failure rate because by the time the satellite executes observations based on the lagging position reported in the news, the target is no longer in that location.
[0009] 4) Lack of judgment on the timeliness value of tasks: Existing news aggregation and summarization automated systems usually only summarize or classify news texts, without taking into account the time difference between the news release time and the current time, whether the target has the ability to move, and whether the event is sudden, natural disaster or general news. This makes it easy to generate invalid observation tasks for targets that have lost their timeliness value, resulting in a waste of satellite resources.
[0010] 5) Hallucination Defect of Large Language Models: If the processing steps requiring precise numerical calculations, such as coordinate transformation, region clipping, and flight path estimation, are completely delegated to the large language model to output the results (i.e., the approach in Basic Scheme 2), the following problems will arise: The large language model essentially generates text or numerical sequences that conform to linguistic rules based on statistical probability, rather than performing deterministic mathematical operations. Therefore, although its output coordinate values may be reasonable in form, they may deviate substantially from the actual geographical location. This phenomenon is known in the industry as hallucination (referring to the model outputting results with high confidence that are reasonable in form but incorrect in content). In addition, repeatedly calling the large language model to generate coordinates for the same input text may result in inconsistent output results, failing to guarantee the determinism and reproducibility of the calculation results, and not meeting the requirements of satellite mission planning systems for coordinate accuracy and result traceability. Summary of the Invention
[0011] To address the issues of substantial deviations in obtaining target coordinates based on text using large language models, and the lack of determinism and reproducibility of the coordinates, this invention provides a remote sensing task generation system based on news text.
[0012] The present invention provides a remote sensing task generation system based on news text, comprising:
[0013] Search tools are used to search for target news on news websites by keywords and obtain the URL of the target news;
[0014] News text extraction tool, used to extract the complete text content of a target news website;
[0015] A large language model is used to perform structured parsing of complete news text content using pre-designed keyword rules, determine target attributes, target location names, and observational value; and output target attributes and target location names with observational value; the target attributes include point targets, moving targets, and regional targets;
[0016] The location transformation tool is used to obtain geospatial data of point targets based on the target location name, and to generate the original regional outline coordinates of moving targets or regional targets based on the target location name; the original regional outline coordinates are processed to obtain the corresponding regional outline representation.
[0017] A task assembly tool is used to assemble target attributes and geospatial data or regional contour representations with observational value to obtain structured task records; a confidence level is configured for each structured task record, and a corresponding task name is set; the task name and the corresponding structured task record are output as remote sensing tasks; the confidence level is used as reference data for remote sensing task planning.
[0018] According to the remote sensing task generation system based on news text of the present invention, the large language model communicates and collaborates with the retrieval tool, news text extraction tool, location conversion tool, and task assembly tool through the MCP protocol; each tool declares its name, function, and parameter format according to the MCP protocol; after reading the declaration, the large language model autonomously judges and initiates a call request to the corresponding tool based on the current task stage; after the corresponding tool executes the calculation logic, it returns the calculation result to the large language model in an agreed format; the large language model further judges and calls the tool to execute the next task based on the returned result, until the remote sensing task output is achieved.
[0019] According to the remote sensing task generation system based on news text of the present invention, the news text extraction tool first judges the target news URL. If it is an intermediate URL, it uses a decoding method to restore the real URL of the target news, and then extracts the complete news text content under the real URL of the target news.
[0020] The method for extracting complete news text content is as follows: First, remove elements from the webpage that are not related to the main content. Then, extract the title, publication time field, and main text paragraphs from the multiple sets of meta-information on the webpage in sequence, and then concatenate them to obtain the complete news text content.
[0021] According to the remote sensing task generation system based on news text of the present invention, the target location place name adopts a composite structure of country name plus place name;
[0022] The observation value is determined based on the target attributes. If it is a moving target, it is set to have observation value within 72 hours.
[0023] For point or regional targets, the observation value is set within 14 days for natural disaster events and within 72 hours for general events.
[0024] According to the remote sensing task generation system based on news text of the present invention, the location conversion tool obtains geospatial data of point targets and original regional outline coordinates of moving targets or regional targets based on a map, and represents the geospatial data or original regional outline coordinates with a code given in the map.
[0025] According to the remote sensing task generation system based on news text of the present invention, for regional targets pointing to complete administrative regions / sea areas, the location transformation tool uses a polyline simplification algorithm to compress the number of original regional contour coordinates to obtain a regional contour representation, including:
[0026] The original region outline coordinates are converted into multiple corresponding polylines, and the multiple polylines are connected in sequence to form a closed region;
[0027] For each polyline, determine all coordinate points and set a tolerance threshold; take the line segment formed by connecting the starting point and ending point of the current polyline as the initial approximation of the current polyline, and calculate the vertical distance from the middle coordinate point of the current polyline to the line segment.
[0028] The intermediate coordinate point corresponding to the maximum vertical distance is selected as the candidate dividing point. If the maximum vertical distance does not exceed the tolerance threshold, all intermediate coordinate points are removed. If the maximum vertical distance exceeds the tolerance threshold, the candidate dividing point is retained as the new endpoint to divide the current polyline into two sub-polylines. Then, the same processing as the current polyline is recursively performed on each sub-polyline to finally obtain the region contour representation after the number of coordinates is compressed.
[0029] According to the remote sensing task generation system based on news text of the present invention, the method for calculating the vertical distance from the middle coordinate point on the current polyline to the line segment is as follows:
[0030] The original region outline coordinates are represented as latitude and longitude coordinates;
[0031] Calculate the average latitude based on the latitude coordinates of all points on the current broken line. Then by average latitude Calculate the scaling factor :
[0032] ;
[0033] With scaling factor Perform a proportional scaling correction on the longitude coordinates of all points on the current polyline, and calculate the perpendicular distance from the intermediate coordinate point to the line segment:
[0034] ,
[0035] ,
[0036] In the formula For position coefficients, This refers to the corrected intermediate coordinate point determined by the corresponding latitude coordinates and the corrected longitude coordinates. The middle coordinate point, This is the corrected starting point coordinate on the current broken line. This refers to the corrected endpoint coordinates on the current broken line. The starting point coordinates and the endpoint coordinates The line segments formed Vertical distance;
[0037] In the compressed region outline representation, the upper limit of the number of coordinate points is set to... Then the minimum tolerance threshold for:
[0038] ,
[0039] In the formula To set the tolerance threshold, This represents the simplified algorithm for polyline.
[0040] According to the remote sensing task generation system based on news text of the present invention, for a regional target pointing to a directional local area, the position conversion tool first obtains the original regional outline coordinates of the complete region corresponding to the directional local area;
[0041] Calculate the geometric centroid coordinates of the closed region formed by the original region contour coordinates; use the longitude and latitude coordinates of the geometric centroid as the north-south and east-west boundaries, respectively; select the north-south or east-west boundary according to the directional description of the directional local region, and use the half-plane clipping algorithm to cut the original complete region into two parts, retaining the side facing the directional description as the region contour representation.
[0042] If the directional description contains two directional components, then on the basis of the first side region, the half-plane clipping algorithm corresponding to the second directional component is superimposed and executed, and the side facing the second directional component is retained as the region contour representation.
[0043] The method for calculating the geometric centroid coordinates is as follows:
[0044] ,
[0045] In the formula Let be the area of the closed region defined by the original region's outline coordinates. For the closed area vertex coordinates These are the coordinates of the geometric centroid.
[0046] According to the remote sensing task generation system based on news text of the present invention, for a regional target pointing to an area offset outside a reference point, the position conversion tool first obtains the reference point coordinates as the original regional outline coordinates; then, the directional description pointing to the outside of the reference point is converted into an azimuth angle; then, based on the reference point coordinates, azimuth angle and offset distance description, the coordinates of the offset center point after offset are solved using the great circle navigation forward solution formula; and a rectangular area is generated as the regional outline representation with the offset center point coordinates as the center and the description width and height.
[0047] Let the coordinates of the reference point be... azimuth angle is The angular distance obtained based on the offset distance is ,in Longitude as the reference point Calculate the coordinates of the offset center point using the latitude of the reference point. :
[0048] ;
[0049] ,
[0050] In the formula This is the offset distance. This is the average radius of the Earth.
[0051] According to the remote sensing task generation system based on news text of the present invention, for a moving target, the moving distance is calculated based on the target's length along the moving direction, the moving speed, and the time difference between the target's news observation time and the task generation time. :
[0052] ,
[0053] In the formula For movement speed, For the time difference, The length in kilometers of the target along the direction of movement;
[0054] If the target news does not provide the direction of movement, take the radius of the uncertain region as the center of the circle centered on the target's location at the time of news observation. The determined coordinates of the circular region are the coordinates of the original region outline:
[0055] ,
[0056] In the formula This indicates a minimum radius of 10 kilometers.
[0057] If the target news provides the direction of movement is According to the direction of movement Calculate the distance traveled The corresponding coordinate point is used as the center of the circle, and the radius of the uncertain region is taken. The determined coordinates of the circular region are the coordinates of the original region outline:
[0058] ;
[0059] The coordinates of the target are selected at equal angular intervals from the original region contour coordinates, and then connected sequentially to form an approximately circular polygon, which serves as the corresponding region contour representation.
[0060] The beneficial effects of this invention are as follows: This invention combines online news retrieval, natural language semantic understanding, and map geocoding services to automatically convert unstructured news text into structured satellite remote sensing observation tasks. It separates the processing steps requiring precise calculations (such as region cropping) from the judgment steps requiring semantic understanding (such as semantic parsing of news content), entrusting them to deterministic computing tools and large language models respectively. This effectively avoids the illusionary risks caused by large language models directly outputting coordinates.
[0061] This invention achieves fully automated generation of structured tasks with a unified output format, enabling batch processing of large-scale news data and significantly improving task generation efficiency and result consistency. By employing a compression method that prioritizes boundary shape fidelity and automatically constrains the number of coordinate points to an upper limit, it significantly improves the accuracy of region shape representation compared to rectangular approximation schemes; and significantly reduces the amount of data required for data transmission and storage compared to directly using unprocessed raw boundary data.
[0062] This invention provides a unified, reusable directional cutting algorithm and offset area construction algorithm, replacing manual point plotting and drawing. The processing results are repeatable and can be batch-processed and automated for large-scale similar news reports. The method for estimating the position of moving targets allows satellite observation coordinates to dynamically reflect the target's most likely current location range. Compared to directly using the lagging position recorded in news reports, this significantly reduces the risk of observation failure due to target displacement. When the flight direction is known, the estimated area range can be further narrowed, which helps improve the observation success rate and save satellite resources. This invention provides quantitative assessment results of mission timeliness value, avoiding the repeated scheduling of satellite observation missions for targets that have lost their observation value.
[0063] This invention separates the deterministic calculation process from the semantic understanding and judgment process, avoiding the illusion that the coordinate values directly output by the large language model are reasonable in form but incorrect in content; no matter how many times the same input is calculated, the calculation result can be completely consistent, which meets the requirements of satellite mission planning system for coordinate accuracy and traceability and reproducibility of results. Attached Figure Description
[0064] Figure 1 This is a schematic diagram illustrating the calling relationship between the large language model and other tools in the remote sensing task generation system based on news text described in this invention.
[0065] Figure 2 This is a flowchart of the task generation process of the remote sensing task generation system based on news text described in this invention. Detailed Implementation
[0066] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0067] Specific Implementation Method 1: Combination Figure 1 and Figure 2 As shown, this invention provides a remote sensing task generation system based on news text, including:
[0068] Search tools are used to search for target news on news websites by keywords and obtain the URL of the target news;
[0069] News text extraction tool, used to extract the complete text content of a target news website;
[0070] A large language model is used to perform structured parsing of complete news text content using pre-designed keyword rules, determine target attributes, target location names, and observational value; and output target attributes and target location names with observational value; the target attributes include point targets, moving targets, and regional targets;
[0071] The location transformation tool is used to obtain geospatial data of point targets based on the target location name, and to generate the original regional outline coordinates of moving targets or regional targets based on the target location name; the original regional outline coordinates are processed to obtain the corresponding regional outline representation.
[0072] A task assembly tool is used to assemble target attributes and geospatial data or regional contour representations with observational value to obtain structured task records; a confidence level is configured for each structured task record, and a corresponding task name is set; the task name and the corresponding structured task record are output as remote sensing tasks; the confidence level is used as reference data for remote sensing task planning.
[0073] This implementation adopts a collaborative architecture with a deterministic computing module and a semantic understanding module, which can communicate and collaborate through the MCP (Model Context Protocol).
[0074] The MCP protocol is a standardized communication protocol that specifies the declaration format of callable capabilities, the parameter format of call requests, and the return format of call results. The deterministic computation module pre-implements specific operations such as news retrieval and coordinate calculation as several independent functional units with fixed input / output formats, deterministic calculation processes, and the same input will inevitably produce the same output. These are called tool units, and the module declares the names, functional descriptions, and parameter formats of these tools according to the format specified in the MCP protocol. After reading the declaration, the large language model autonomously determines which tool should be called and what parameters should be passed in based on the content of the news text to be processed, and initiates a call request. After the deterministic computation module executes the calculation logic of the corresponding tool, it returns the calculation result to the large language model in the agreed format. The large language model determines the tool to be called next based on the returned result, and this process is repeated until the transformation from a news text to a complete structured remote sensing task is completed.
[0075] The tools included in the deterministic calculation module of this implementation, such as the retrieval tool, news text extraction tool, location conversion tool, and task assembly tool, are internally independent of and do not call any large language model. They only execute fixed retrieval or numerical calculation logic, thus possessing the characteristics of deterministic results, reproducibility, and traceability. For processing steps that require understanding natural language semantics and involve a certain degree of ambiguity, such as determining the location corresponding to the event described in the news, choosing which of multiple candidate locations with the same name returned by geocoding, and whether the target has mobility attributes, these are handled by the large language model, which has semantic understanding capabilities. By separating these two different types of processing tasks—deterministic numerical calculation and natural language semantic judgment—and assigning them to modules best suited for these tasks, the architectural risk of illusions caused by the large language model directly outputting coordinate values is avoided, while retaining the large language model's ability to handle the inherent uncertainty and ambiguity in natural language.
[0076] The search tool (search_hot_news) can retrieve relevant news items based on keywords within a pre-defined range of authoritative media domains; the news article extraction tool (extract_news_article) captures the title, publication time, and article content (including decoding of encrypted redirect links) of the webpage corresponding to a specified news link; the location conversion tool (geocode_place) can convert place name text into latitude and longitude coordinates or administrative boundary data (multiple candidates may be returned for the same place name, which will be selected by the semantic understanding module).
[0077] The location transformation tools include a directional cutting tool (cut_region_by_direction) and an external offset region building tool (build_external_offset_area) for regional targets; and a mobile target estimation tool (estimate_mobile_target_area) for moving targets. The directional cutting tool cuts out corresponding sub-regions from a known complete region according to the east, west, south, north and their combinations. The external offset region building tool uses a known reference point as a baseline and offsets it by a specified direction and distance to generate a new rectangular region. The mobile target estimation tool estimates the possible area range of the target based on the observation time, moving speed and moving direction.
[0078] Finally, the task assembly tool (generate_remote_sensing_tasks) based on a large language model is used to execute the above results in a predetermined order, realizing end-to-end generation from keywords to remote sensing task drafts, which are then reviewed and confirmed by the semantic understanding module.
[0079] In this embodiment, the above modules and their calling relationships are as follows: Figure 1 As shown, a host program can be used to carry and drive the semantic understanding module to complete the decision loop of tool selection, parameter passing, and result acquisition; the MCP tool service layer consists of the aforementioned tool modules.
[0080] Combination Figure 2 As shown, for search tools, a whitelist of authoritative media can be pre-selected, which is a list of domain names of reputable news websites with high credibility, such as official media domains like Xinhua News Agency and Reuters. The search process is conducted within the domains listed to exclude content with lower credibility, such as self-media and non-authoritative information sources. The search can preferentially use the Bing News search interface, combining domain range restrictions with search keywords into a query statement. If a Bing interface access key is not configured, the Google News RSS subscription interface can be used as an alternative, initiating keyword and domain-specific queries for each domain in the whitelist, and merging and deduplicating the search results from multiple domains before returning the title, source, publication time, and summary information.
[0081] Furthermore, the news text extraction tool first determines the target news URL. If it is an intermediate URL, it uses a decoding method to restore the real URL of the target news, and then extracts the complete news text content under the real URL of the target news.
[0082] The method for extracting complete news text content is as follows: First, remove elements from the webpage that are not related to the main content. Then, extract the title, publication time field, and main text paragraphs from the multiple sets of meta-information on the webpage in sequence, and then concatenate them to obtain the complete news text content.
[0083] The news text extraction tool is configured to decode encrypted redirect links. For example, links returned by Google News search results are not the actual URLs of the original news articles, but rather encrypted intermediate addresses that require a server-side redirect to reach the original text. If no processing is performed and a crawling request is directly sent to these intermediate addresses, the resulting content will only be a redirect prompt page without any text. Therefore, before crawling the text, it first determines whether the link to be crawled is an encrypted redirect link. If so, a dedicated decoding method is used to restore the actual URL of the original news article before a crawling request is sent to that actual URL. During the text crawling process, elements irrelevant to the text content, such as scripts, styles, and navigation bars, are first removed. Then, the title and publication time fields are extracted from multiple sets of meta information (meta tags) on the webpage. Finally, the text paragraphs are concatenated into complete text.
[0084] The target location name adopts a composite structure of country name plus place name;
[0085] The observation value is determined based on the target attributes. If it is a moving target, it is set to have observation value within 72 hours.
[0086] For point or regional targets, the observation value is set within 14 days for natural disaster events and within 72 hours for general events.
[0087] This implementation provides a quantitative assessment of timeliness value, which solves the drawback of existing methods that lack judgment on the timeliness value of tasks.
[0088] The large language model performs structured parsing by mapping free-formatted text content to judgment results of several fixed fields. If words related to moving targets such as ships or regions such as regions and cities appear in the text, the target attribute is determined to be a regional target, and its range should be represented by polygons; otherwise, it is determined to be a point target, represented by a single coordinate point. The model also determines whether the target itself has a moving attribute based on lexical features. Candidate place names are extracted from the text according to pre-designed place name matching rules, which include, but are not limited to, text fragments ending with common place name suffixes such as province / city / county / island / port / sea area. Composite structures of country names plus place names are expanded, retaining multiple candidate place names, and then disambiguation is performed on these candidate place names to improve the success rate of subsequent geocoding queries. Combining whether the target has a moving attribute, event-related lexical features, and the time difference between the news release time and the current time, the model determines whether the target corresponding to the news still has observational value.
[0089] The system can generate corresponding explanations to determine whether the time of news release exceeds the observation value period.
[0090] The location conversion tool obtains geospatial data of point targets and original regional outline coordinates of moving targets or regional targets based on the map, and represents the geospatial data or original regional outline coordinates with a code given in the map.
[0091] In this embodiment, the original region contour coordinate system can be WGS-84.
[0092] The location conversion tool can pre-build a thesaurus containing all administrative region names and sea area names. If the text of the place name to be queried or the information indicating the location matches the thesaurus, such as the text containing the words "Hebei," the query can be directly performed using the Gaode Maps API, which can provide boundary data accurate to the actual administrative division. If the thesaurus is not matched, such as Texas, USA, the Google Maps API is used first. If the Google Maps API does not return a valid result, or if the returned coordinates are determined to fall outside the target location, a fallback query is performed using a valid map corresponding to the actual coordinates to ensure that the true administrative region boundary data is obtained, rather than a simplified rectangular approximation.
[0093] The coordinates obtained through the Gaode Map interface are GCJ-02 encrypted coordinates used by map services, which have a systematic deviation from the WGS-84 coordinates used by satellite positioning. The GCJ-02 encrypted coordinates can be converted into WGS-84 coordinates using a general coordinate conversion algorithm before use.
[0094] As an example, for regional targets pointing to complete administrative regions / sea areas, the location transformation tool uses the Douglas-Peucker polyline simplification algorithm to compress the number of coordinates in the original regional outline, obtaining a regional outline representation, including:
[0095] The original region outline coordinates are converted into multiple corresponding polylines, and the multiple polylines are connected in sequence to form a closed region;
[0096] For each polyline, determine all coordinate points and set a tolerance threshold; take the line segment formed by connecting the starting point and ending point of the current polyline as the initial approximation of the current polyline, and calculate the vertical distance from the middle coordinate point of the current polyline to the line segment.
[0097] The intermediate coordinate point corresponding to the maximum vertical distance is selected as the candidate dividing point. If the maximum vertical distance does not exceed the tolerance threshold, the current polyline is considered to be approximated by connecting its two endpoints, and all intermediate coordinate points are eliminated. If the maximum vertical distance exceeds the tolerance threshold, the candidate dividing point is retained as the new endpoint to divide the current polyline into two sub-polylines. The same processing is then recursively performed on each sub-polyline to obtain the region contour representation with reduced coordinates. After processing, the number of coordinate points retained by the polyline is significantly reduced, while the deviation of the overall shape of the polyline from the original boundary is constrained by the tolerance threshold and does not exceed it.
[0098] When the target attribute is determined to be a regional target, the official boundary data of the administrative division can be queried based on the candidate location's administrative division code (Adcode, a unique numerical code assigned to each administrative division by Amap, for example, 110000 corresponds to Beijing; using the code for retrieval can accurately pinpoint the specific administrative division and avoid ambiguity caused by place names with the same name, such as multiple provinces having a location named Qinghe). The boundary data must also be converted to WGS-84 coordinates before use. The number of boundary coordinate points obtained from the query usually reaches several thousand to tens of thousands, and the number of coordinate points needs to be compressed to the range that the system can process, while maintaining the fidelity of the boundary shape as much as possible.
[0099] The method for calculating the perpendicular distance from the middle coordinate point on the current polyline to the line segment is as follows:
[0100] The original region outline coordinates are represented as latitude and longitude coordinates;
[0101] To eliminate the effect that the ground distance corresponding to a unit angle in the longitude direction is less than the ground distance corresponding to a unit angle in the latitude direction in high-latitude regions, and to avoid overestimating the east-west deviation, the average latitude is calculated based on the latitude coordinates of all coordinate points on the current broken line. Then by average latitude Calculate the scaling factor :
[0102] ;
[0103] With scaling factor Perform a proportional scaling correction on the longitude coordinates of all points on the current polyline, and calculate the perpendicular distance from the intermediate coordinate point to the line segment:
[0104] ,
[0105] ,
[0106] In the formula For position coefficients, This refers to the corrected intermediate coordinate point determined by the corresponding latitude coordinates and the corrected longitude coordinates. The middle coordinate point, This is the corrected starting point coordinate on the current broken line. This refers to the corrected endpoint coordinates on the current broken line. The starting point coordinates and the endpoint coordinates The line segments formed Vertical distance; This means restricting the calculated value to a closed interval. Inside, to ensure that the perpendicular foot lands on the line segment. Within its range, rather than along its extension line.
[0107] In the compressed region outline representation, the upper limit of the number of coordinate points is set to... Then the minimum tolerance threshold for:
[0108] ,
[0109] In the formula To set the tolerance threshold, This represents the simplified algorithm for polyline.
[0110] This implementation method, while acquiring real-world boundary data, automatically constrains the number of boundary coordinate points within the system's processing range and maintains the fidelity of the boundary shape as much as possible.
[0111] Regarding tolerance threshold The value of the tolerance threshold is such that the larger the tolerance threshold is set, the fewer coordinate points are retained after simplification, but the greater the deviation of the boundary shape from the original boundary; the smaller the tolerance threshold is set, the higher the fidelity of the boundary shape, but the number of retained points may still not meet the system's processing limit.
[0112] As an example, set a uniform upper limit for target points. The limit is 200, meaning that for any level of complexity, the number of coordinate points after compression will not exceed 200. Then, a binary search method is used to automatically determine the tolerance threshold that satisfies the upper limit constraint on the number of points. If the number of points still exceeds the specified range, the middle value within the range is used for simplification. This indicates that the tolerance threshold needs to be increased, and the lower bound of the search interval should be adjusted to the median value; if the simplified number of points does not exceed If the tolerance threshold is not reduced, it indicates that there is still room to reduce the fidelity. Adjust the upper bound of the search interval to the middle value. After repeated iterations, it can converge to the minimum tolerance threshold that can be obtained under the condition of satisfying the number of points constraint.
[0113] Perform tolerance threshold Once selected, regardless of the size of the target area or the complexity of the boundary shape, the number of boundary coordinate points can be automatically constrained within the system's processing range without manual parameter adjustment. Compared to processing methods that eliminate coordinate points at a fixed ratio or use a uniform fixed tolerance threshold, this implementation method can simultaneously take into account both the boundary shape fidelity and the constraint of the number of coordinate points, which are mutually restrictive technical indicators.
[0114] As an example, for a regional target pointing to a directional local area, the position transformation tool first obtains the original regional outline coordinates of the complete region corresponding to the directional local area;
[0115] Calculate the geometric centroid coordinates of the closed region formed by the original region contour coordinates; use the longitude and latitude coordinates of the geometric centroid as the north-south and east-west boundaries, respectively; select the north-south or east-west boundary according to the directional description of the directional local region, and use the half-plane clipping algorithm to cut the original complete region into two parts, retaining the side facing the directional description as the region contour representation.
[0116] If the directional description contains two directional components, then on the basis of the first side region, the half-plane clipping algorithm corresponding to the second directional component is superimposed and executed, and the side facing the second directional component is retained as the region contour representation.
[0117] The method for calculating the geometric centroid coordinates is as follows:
[0118] ,
[0119] In the formula Let be the area of the closed region defined by the original region's outline coordinates. For the closed area vertex coordinates These are the coordinates of the geometric centroid.
[0120] For example, if the target area described in a news text is the northeastern part of a province, the geocoding interface can only return the complete administrative boundaries of the province. Therefore, it's necessary to cut out sub-regions that conform to the directional description from the complete region. The coordinates of the geometric centroid (i.e., the area-weighted geometric center of the polygon) of the complete region are calculated. Then, using the longitude and latitude of the centroid as the north-south and east-west boundaries respectively, a half-plane clipping algorithm (Sutherland–Hodgman algorithm) is used to cut the original region into two parts, retaining the side facing the target direction. If the directional description contains two directional components, such as northeast, then two cuts are performed sequentially: first, the eastern part is retained according to the east-west boundary, and then the northern part is retained according to the north-south boundary. The region obtained after superimposing these two cuts is the final northeastern sub-region.
[0121] The specific processing method for performing half-plane clipping is as follows: along the dividing line or The original region boundary is traversed segment by segment. For two adjacent vertices on the boundary, if both are located on the side to be preserved, the corresponding segment boundary is preserved as is. If they are located on opposite sides of the dividing line (i.e., the corresponding segment boundary crosses the dividing line), the coordinates of the intersection point between the corresponding segment boundary and the dividing line are calculated at the crossing point using linear interpolation. The intersection point is then inserted into the output result to ensure that the new region boundary obtained after cutting is still a closed polyline with the beginning and end connected.
[0122] As an example, for a target area pointing to an offset region outside the reference point, the position conversion tool first obtains the reference point coordinates as the original region outline coordinates; then it converts the direction description pointing to the offset outside the reference point into an azimuth angle; then, based on the reference point coordinates, azimuth angle and offset distance description, it uses the great circle navigation forward solution formula to solve for the offset center point coordinates; and with the offset center point coordinates as the center, it generates a rectangular region as the region outline representation according to the description width and height.
[0123] Let the coordinates of the reference point be... azimuth angle is The angular distance obtained based on the offset distance is ,in Longitude as the reference point Calculate the coordinates of the offset center point using the latitude of the reference point. :
[0124] ;
[0125] ,
[0126] In the formula This is the offset distance. The average radius of the Earth is 6371.0088 kilometers.
[0127] When a news text describes a target located at a certain direction and distance from a known reference point (landmark), such as a sea area 50 kilometers northeast of a certain location, and the current target is neither within the territory of any known administrative division nor can its boundary data be obtained directly through geocoding, a new area must be generated based on the known reference point and offset according to direction and distance.
[0128] First, the direction description is converted into azimuth (0 degrees for true north and increasing clockwise, e.g., 45 degrees for northeast). Based on this, the Great-circle Navigation formula (which determines the coordinates of the point reached after traveling a specified distance from a known starting point along a specified azimuth) is used to solve for the coordinates of the new center point after offset. Finally, a rectangular area is generated with the new center point as the center and the specified width and height.
[0129] The reason for using the great circle navigation formula instead of the planar coordinate addition and subtraction method to calculate the offset coordinates is that the Earth's surface is approximately spherical, and the higher the latitude, the smaller the actual ground distance corresponding to a unit longitude difference. If the planar coordinate addition and subtraction method is used to calculate the offset, it will produce significant errors in mid-to-high latitude regions.
[0130] If the news text describes the location as having orthogonal components offset by several kilometers to the east and north (rather than a general directional name), the calculation is done in two steps: First, offset by the corresponding distance at an azimuth of 0 degrees (due north; a negative north offset is equivalent to an azimuth of 180 degrees) to obtain the midpoint. Then, using the midpoint as a reference, offset by the corresponding distance at an azimuth of 90 degrees (due east; a negative east offset is equivalent to an azimuth of 270 degrees) to obtain the final center point coordinates. Using the final center point as the center, offset by half the width or height in each of the four directions according to the given area width and height to calculate the coordinates of the four vertices of the rectangular area.
[0131] The two examples above demonstrate how a program can automatically and repeatedly cut out sub-regions from a known complete region according to direction, and generate new regions by offsetting them by direction and distance from a reference point.
[0132] As an example, for a moving target, the distance traveled is calculated based on the target's length along the direction of movement, its speed, and the time difference between the target's news observation time and the task generation time. :
[0133] ,
[0134] In the formula For movement speed, For the time difference, The length in kilometers of the target along the direction of movement;
[0135] If the target news does not provide the direction of movement, take the radius of the uncertain region as the center of the circle centered on the target's location at the time of news observation. The determined coordinates of the circular region are the coordinates of the original region outline:
[0136] ,
[0137] In the formula This indicates a minimum radius of 10 kilometers.
[0138] If the target news provides the direction of movement is According to the direction of movement Calculate the distance traveled The corresponding coordinate point is used as the center of the circle, and the radius of the uncertain region is taken. The determined coordinates of the circular region are the coordinates of the original region outline:
[0139] ;
[0140] The coordinates of the target are selected at equal angular intervals from the original region contour coordinates, and then connected sequentially to form an approximately circular polygon, which serves as the corresponding region contour representation.
[0141] This implementation method enables dynamic estimation of the possible current location range of a moving target.
[0142] Setting a minimum radius can prevent inaccurate range of uncertain areas due to excessively small calculated sailing distance values.
[0143] Providing direction of movement Under these conditions, the radius of the uncertain region is narrowed to 25% of the maximum possible movement distance to reflect the reduced uncertainty brought about by the known direction information.
[0144] Regardless of whether the navigation direction is known, the great circle navigation formula is used. Coordinate points are calculated around a determined center at equal angular intervals, and these points are connected sequentially to form an approximately circular polygon. This polygon serves as the output of the region's observation coordinates for the moving target. For example, the angle is divided into 24 equal parts, meaning a point is taken every 15 degrees.
[0145] In this embodiment, the confidence level in the task assembly tool is the smaller value between the confidence level of news content classification judgment and the confidence level of place name coordinate matching.
[0146] The task assembly tool takes the geographical location information (point coordinates, or compressed area polygon coordinates, labeled WGS-84) obtained from each stage, combines it with the determined target attributes and timeliness value assessment results, and assembles it into a structured task record with a uniform format. Each task record includes a confidence index, with a value range of [missing information]. For closed intervals, the closer the value is to 1, the higher the credibility of the task judgment result; the closer it is to 0, the lower the credibility, and manual review is recommended.
[0147] The confidence level is chosen to be the smaller value between the confidence level of news content classification judgment and the confidence level of place name coordinate matching because if the confidence level of any judgment in any processing link is low, it should be reflected in the overall confidence level of the entire task record. That is, the confidence level of the entire task record is determined by the link with the weakest judgment basis.
[0148] This invention also provides a structured reasoning prompt, corresponding to the remote_sensing_task_reasoning_prompt prompt template, to constrain the accessed large language model during task generation. It must first complete the target attribute judgment and then perform place name disambiguation; for moving targets, it must first complete the position estimation; and for directional regions, it must first obtain the complete region and then perform the segmentation. The corresponding tools are called in a predetermined order to prevent the large language model from skipping the deterministic calculation stage and directly outputting coordinate values that have not been calculated and verified.
[0149] In summary, the system of the present invention can produce technical effects such as controllable boundary data volume, automated region generation, and improved timeliness of moving target positioning.
[0150] In this invention, the geocoding service can be replaced by interfaces provided by other map vendors, such as OpenStreetMap's Nominatim geocoding service, Bing Maps, Baidu Maps, etc., as long as the technical concept of dynamically selecting the data source based on whether the place name points to the target area and obtaining the real boundary data of the area remains unchanged.
[0151] In addition to the Douglas-Peucker polyline simplification algorithm, the administrative region boundary coordinate point compression algorithm can be replaced by other similar simplification algorithms, such as the Visvalingam-Whyatt algorithm. Its processing principle is to remove the vertices that have the least impact on the overall area of the polygon after each removal. Combined with the technical idea of this invention to automatically determine the tolerance threshold that meets the upper limit constraint of the number of points using the binary search method, it can also achieve an automatic trade-off between the two indicators of boundary shape fidelity and coordinate point number constraint.
[0152] The news search channel can be replaced by other news search engines or commercial news data interfaces, and searches can be performed within the scope of pre-defined authoritative media domains.
[0153] The communication protocol between the semantic understanding module and the deterministic computation module, apart from the MCP protocol, can be replaced by other tool calling protocols or frameworks that have the three basic capabilities of tool capability declaration, call request initiation, and call result return. The architectural approach of entrusting deterministic computation to fixed tools and semantic understanding to large language models must be maintained.
[0154] In the system of the present invention, the uncertain region generated when the heading information is unknown can be changed from a circle to a fan shape, covering only the angle range of the possible movement direction of the target, so as to further narrow the range of the uncertain region.
[0155] This invention provides an architecture where deterministic computation and semantic understanding modules work collaboratively. Processing steps requiring precise and repeatable computation, such as news retrieval, boundary compression, region segmentation, and moving target estimation, are encapsulated into several independent tools with fixed input / output formats and deterministic calculation results. These tools are then provided to a large language model via a standardized protocol. The large language model only handles fuzzy judgments requiring semantic understanding, such as tool selection, candidate place name disambiguation, and determining whether a target possesses mobility attributes. This fundamentally avoids the illusionary risks associated with the large language model directly outputting coordinate values, while retaining its ability to handle the inherent uncertainties of natural language.
[0156] Automatic selection of map data source: Based on whether the place name text or the information of the location matches the preset keyword table, the system dynamically determines the corresponding map service to use for querying, ensuring that real administrative boundary data can be obtained while taking into account the query coverage of place names.
[0157] The number of administrative region boundary coordinate points is automatically constrained within a given upper limit while maintaining the fidelity of the boundary shape as much as possible: the polyline simplification algorithm is combined with the binary search method to automatically determine the tolerance threshold that makes the number of coordinate points converge to the target upper limit, without the need for manual parameter tuning for the shape complexity of each region.
[0158] The process involves directional cutting around the centroid of a region to generate sub-regions, and offsetting around a reference point by direction and distance to generate new regions: converting natural language directional descriptions such as sub-regions in a certain direction or a location at a certain distance in a certain direction in news text into a process of generating region coordinates that can be automatically and repeatedly executed by the program.
[0159] The existing method of directly using outdated news reports replaces the current method of dynamically adjusting the geographical area based on the time and the completeness of directional information. Simultaneously, it enables a graded and quantitative assessment of the timeliness and observational value of remote sensing missions.
[0160] While the invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that different dependent claims and features described herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other described embodiments.
Claims
1. A remote sensing task generation system based on news text, characterized in that... include, Search tools are used to search for target news on news websites by keywords and obtain the URL of the target news; News text extraction tool, used to extract the complete text content of a target news website; A large language model is used to perform structured parsing of complete news text content using pre-designed keyword rules, determine target attributes, target location names, and observational value; and output target attributes and target location names with observational value; the target attributes include point targets, moving targets, and regional targets; The location transformation tool is used to obtain geospatial data of point targets based on the target location name, and to generate the original regional outline coordinates of moving targets or regional targets based on the target location name; the original regional outline coordinates are processed to obtain the corresponding regional outline representation. A task assembly tool is used to assemble target attributes and geospatial data or regional contour representations with observational value to obtain structured task records; a confidence level is configured for each structured task record, and a corresponding task name is set; the task name and the corresponding structured task record are output as remote sensing tasks; the confidence level is used as reference data for remote sensing task planning.
2. The remote sensing task generation system based on news text according to claim 1, characterized in that, The large language model communicates and collaborates with the retrieval tool, news text extraction tool, location conversion tool, and task assembly tool via the MCP protocol; each tool declares its name, function, and parameter format according to the MCP protocol; after reading the declaration, the large language model autonomously makes a judgment based on the current task stage and initiates a call request for the corresponding tool. After the corresponding tool executes the calculation logic, it returns the calculation results to the large language model in a predefined format. The large language model then makes further judgments based on the returned results and calls the tool for the next task until the remote sensing task output is achieved.
3. The remote sensing task generation system based on news text according to claim 1, characterized in that, The news text extraction tool first determines the target news URL. If it is an intermediate URL, it uses a decoding method to restore the real URL of the target news, and then extracts the complete news text content under the real URL of the target news. The method for extracting complete news text content is as follows: First, remove elements from the webpage that are not related to the main content. Then, extract the title, publication time field, and main text paragraphs from the multiple sets of meta-information on the webpage in sequence, and then concatenate them to obtain the complete news text content.
4. The remote sensing task generation system based on news text according to claim 1, characterized in that, The target location name adopts a composite structure of country name plus place name; The observation value is determined based on the target attributes. If it is a moving target, it is set to have observation value within 72 hours. For point or regional targets, the observation value is set within 14 days for natural disaster events and within 72 hours for general events.
5. The remote sensing task generation system based on news text according to claim 1, characterized in that, The location conversion tool obtains geospatial data of point targets and original regional outline coordinates of moving targets or regional targets based on the map, and represents the geospatial data or original regional outline coordinates with a code given in the map.
6. The remote sensing task generation system based on news text according to claim 1, characterized in that, For regional targets pointing to complete administrative regions / sea areas, the location transformation tool uses a polyline simplification algorithm to compress the number of coordinates in the original regional outline, obtaining a regional outline representation, including: The original region outline coordinates are converted into multiple corresponding polylines, and the multiple polylines are connected in sequence to form a closed region; For each polyline, determine all coordinate points and set a tolerance threshold; take the line segment formed by connecting the starting point and ending point of the current polyline as the initial approximation of the current polyline, and calculate the vertical distance from the middle coordinate point of the current polyline to the line segment. The intermediate coordinate point corresponding to the maximum vertical distance is selected as the candidate dividing point. If the maximum vertical distance does not exceed the tolerance threshold, all intermediate coordinate points are removed. If the maximum vertical distance exceeds the tolerance threshold, the candidate dividing point is retained as the new endpoint to divide the current polyline into two sub-polylines. Then, the same processing as the current polyline is recursively performed on each sub-polyline to finally obtain the region contour representation after the number of coordinates is compressed.
7. The remote sensing task generation system based on news text according to claim 6, characterized in that, The method for calculating the perpendicular distance from the middle coordinate point on the current polyline to the line segment is as follows: The original region outline coordinates are represented as latitude and longitude coordinates; Calculate the average latitude based on the latitude coordinates of all points on the current broken line. Then by average latitude Calculate the scaling factor : ; With scaling factor The longitude coordinates of all points on the current polyline are proportionally scaled and corrected, and the perpendicular distance from the intermediate coordinate point to the line segment is calculated: , , In the formula For position coefficients, This refers to the corrected intermediate coordinate point determined by the corresponding latitude coordinates and the corrected longitude coordinates. The middle coordinate point, This is the corrected starting point coordinate on the current broken line. This refers to the corrected endpoint coordinates on the current polygonal line. The starting point coordinates and the endpoint coordinates The line segments formed Vertical distance; In the compressed region outline representation, the maximum number of coordinate points is set to [value]. Then the minimum tolerance threshold for: , In the formula To set the tolerance threshold, This represents the simplified algorithm for polyline.
8. The remote sensing task generation system based on news text according to claim 1, characterized in that, For a regional target pointing to a directional local area, the position transformation tool first obtains the original regional outline coordinates of the complete region corresponding to the directional local area; Calculate the geometric centroid coordinates of the closed region formed by the original region contour coordinates; use the longitude and latitude coordinates of the geometric centroid as the north-south and east-west boundaries, respectively; select the north-south or east-west boundary according to the directional description of the directional local region, and use the half-plane clipping algorithm to cut the original complete region into two parts, retaining the side facing the directional description as the region contour representation. If the directional description contains two directional components, then on the basis of the first side region, the half-plane clipping algorithm corresponding to the second directional component is superimposed and executed, and the side facing the second directional component is retained as the region contour representation. The method for calculating the geometric centroid coordinates is as follows: , In the formula Let be the area of the closed region defined by the original region's outline coordinates. For the closed area vertex coordinates These are the coordinates of the geometric centroid.
9. The remote sensing task generation system based on news text according to claim 1, characterized in that, For a target area pointing to the outside of the reference point, the position conversion tool first obtains the reference point coordinates as the original area outline coordinates; then it converts the direction description pointing to the outside of the reference point into an azimuth angle; then, based on the reference point coordinates, azimuth angle and offset distance description, it uses the great circle navigation forward solution formula to solve for the offset center point coordinates; and then, with the offset center point coordinates as the center, it generates a rectangular area according to the description width and height as the area outline representation. Let the coordinates of the reference point be... azimuth angle is The angular distance obtained based on the offset distance is ,in Longitude as the reference point Calculate the coordinates of the offset center point using the latitude of the reference point. : ; , In the formula This is the offset distance. This is the average radius of the Earth.
10. The remote sensing task generation system based on news text according to claim 1, characterized in that, For a moving target, the distance traveled is calculated based on the target's length along the direction of movement, its speed, and the time difference between the target's news observation time and the task generation time. : , In the formula For movement speed, For the time difference, The length in kilometers of the target along the direction of movement; If the target news does not provide the direction of movement, take the radius of the uncertain region as the center of the circle centered on the target's location at the time of news observation. The determined coordinates of the circular region are the coordinates of the original region outline: , In the formula This indicates a minimum radius of 10 kilometers. If the target news provides the direction of movement is According to the direction of movement Calculate the distance traveled The corresponding coordinate point is used as the center of the circle, and the radius of the uncertain region is taken. The determined coordinates of the circular region are the coordinates of the original region outline: ; The coordinates of the target are selected at equal angular intervals from the original region contour coordinates, and then connected sequentially to form an approximately circular polygon, which serves as the corresponding region contour representation.
