A real-time grid tile inferencing map service method and system

Abstract

The application discloses a kind of real-time grid tile inference map service method and system, belong to geographic information system and remote sensing technical field.The method includes service system receives tile request, calculates corresponding geographic space coordinate range according to tile request, according to geographic space coordinate range, read the original data block covering this range from distributed storage, based on MICP protocol, original data block is distributed to CPU rendering pipeline and GPU inference pipeline and is handled in parallel;Wherein, the MICP protocol is specially used for the application layer protocol between computing scheduling unit and intelligent inference engine for task distribution and state synchronization, the visual base map generated by CPU rendering pipeline and the intelligent interpretation result generated by GPU inference pipeline are fused, and are responded as service return.This application realizes grid tile real-time rendering and intelligent inference parallel processing based on MICP protocol, provides map service of instant view.

Description

Technical Field

[0001] This invention relates to the fields of geographic information systems and remote sensing technology, specifically to a real-time raster tile inference map service method and system. Background Technology

[0002] In the fields of Geographic Information Systems (GIS) and remote sensing, with the continuous enrichment of Earth observation methods, the daily generation of satellite imagery and drone data is exploding. However, existing technologies exhibit a significant disconnect between the "visualization" and "in-depth analysis" of data. Traditional tile map services (such as TMS and WMTS standards) primarily address the problem of browsing large datasets online. Their core approach is to preprocess raw data to generate static image tiles in a pyramid structure. While this method ensures smooth browsing, it essentially only provides surface-level visual information and cannot perform semantic understanding of image content. Furthermore, the pre-tiling mode leads to significant storage waste and long data update cycles. Meanwhile, with the development of deep learning technology, although remote sensing intelligent interpretation (AI) has achieved breakthroughs in accuracy, current engineering applications mostly employ offline batch processing. This requires users to first download the complete image files and then run the algorithm locally or on a dedicated server, which is time-consuming and cannot be seamlessly integrated with map browsing. In addition, the aforementioned systems generally rely on common network protocols (such as HTTP / 1.1 or RESTful APIs), which can lead to problems such as high protocol overhead, inability to transmit task priorities, and difficulty in achieving sub-millisecond task distribution in high-concurrency scenarios.

[0003] Although existing platforms support online processing of remote sensing data, none of them have seamlessly integrated deep learning inference capabilities into the standard tile request-response chain, nor do they have a dedicated communication protocol for the collaborative scheduling of CPU rendering and GPU inference tasks, resulting in the inability to achieve a low-latency interactive experience of 'request as inference'.

[0004] To address the aforementioned issues, there is an urgent need for a real-time raster tile inference map service method and system to resolve the problems associated with traditional methods. Summary of the Invention

[0005] The purpose of this invention is to provide a real-time raster tile inference map service method and system. By introducing the lightweight MICP protocol, it achieves sub-millisecond collaborative scheduling of CPU rendering and GPU inference dual pipelines, completing parallel processing of real-time raster tile visualization and deep learning interpretation, and providing a "what you see is what you get" intelligent map service. This solution solves the problem of the separation between visualization and intelligent analysis in traditional map services, significantly reduces response latency and storage overhead, and pioneers a new generation of real-time inference map service mode.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for providing real-time raster tile inference map services includes: Step 1: The service system receives the tile request and calculates the corresponding geospatial coordinate range based on the tile request; Step 2: Based on the geospatial coordinate range, read the original data blocks covering that range from the distributed storage; Step 3: Based on the MICP protocol, distribute the raw data blocks to the CPU rendering pipeline and the GPU inference pipeline for parallel processing; wherein, the MICP protocol is an application layer protocol specifically used for task distribution and state synchronization between the computing scheduling unit and the intelligent inference engine. Step 4: Merge the visualization basemap generated by the CPU rendering pipeline with the intelligent interpretation results generated by the GPU inference pipeline, and return it as a service response.

[0007] Furthermore, in step 1, the tile request parameters are ( z , x , y ),in, x For tile column numbering, y For tile row numbers, z To scale the hierarchy.

[0008] Further, in step 1, the corresponding geospatial coordinate range is calculated based on the tile request, specifically as follows: The pixel coordinates of the top-left corner of the tile are calculated as follows: ; ; In the formula, For tile dimensions; Convert it to geographic coordinates as follows: ; ; In the formula, For target longitude, For the target latitude, This represents the total number of slices.

[0009] Further, in step 2, based on the geospatial coordinate range, the original data blocks covering that range are read from the distributed storage, specifically as follows: Based on the geospatial coordinate range, the original data block is located and read from the distributed storage using vector pyramid indexing technology; the vector pyramid index only stores the bounding box, physical storage pointer and metadata of the original data block.

[0010] Furthermore, in step 3, based on the MICP protocol, the raw data blocks are distributed to the CPU rendering pipeline and the GPU inference pipeline for parallel processing, specifically as follows: Step 301: The computation scheduling unit encapsulates the tile request into a MICP message. The header of the MICP message contains version, type, task ID, dynamic priority, data length and checksum fields, and the body of the MICP message includes encrypted geographic coordinates and tensor data. Step 302: Based on dynamic priority, distribute tasks to target GPU nodes through a dynamic routing mechanism; Step 303: The CPU rendering pipeline and the GPU inference pipeline process the same raw data block in parallel; Step 304: Maintain the state consistency of each node through the heartbeat mechanism of the MICP protocol; Step 305: When an error is detected, execute the hierarchical error handling strategy.

[0011] Furthermore, the dynamic priority The calculation formula is: ; In the formula, α and β are weighting coefficients, Distance represents the Euclidean distance between the tile and the center of the user's view, and Timeliness is the task timeliness factor.

[0012] Furthermore, in step 303, the CPU rendering pipeline and the GPU inference pipeline process the same raw data block in parallel, specifically as follows: The CPU rendering pipeline performs dynamic resampling, radiometric correction and stretching, and slicing on the original data blocks sequentially to generate an 8-bit RGB visualization base map; the GPU inference pipeline performs tensor transformation and normalization on the original data blocks, inputs them into a preloaded deep learning model for forward inference, and outputs a probability map or feature map.

[0013] Further, in step 305, the hierarchical error handling strategy specifically includes: For communication errors, the exponential backoff algorithm is used for retransmission; If a node fails, perform a failover and route the task to a backup node; For data errors where the checksum fails, discard the message and trigger a retransmission.

[0014] Furthermore, in step 4, the visualization basemap generated by the CPU rendering pipeline and the intelligent interpretation result generated by the GPU inference pipeline are fused and returned as a service response, specifically: Using an overlay rendering mode, the intelligent interpretation results are overlaid as a semi-transparent layer on the visualization base map to synthesize a new image as a service response. Using a data separation mode, a response package containing multiple data parts is returned. The response package includes a visual base map and intelligent interpretation results, which are overlaid and rendered by the user.

[0015] The present invention also provides a real-time raster tile inference map service system, applied to the above-mentioned real-time raster tile inference map service method, comprising: a data storage unit, a computing scheduling unit, a tile rendering engine, an intelligent inference engine, and a service fusion unit connected in sequence; The data storage unit is used for distributed storage of raw geospatial raster data and provides a data reading interface based on a vector pyramid index. The computation scheduling unit, as the system control center, is used to receive client tile requests, calculate the corresponding geospatial coordinate range based on the tile request, read the original data block covering the range from the data storage unit based on the geospatial coordinate range, and distribute the original data block to the tile rendering engine and the intelligent inference engine based on the MICP protocol; wherein, the MICP protocol is an application layer protocol specifically used for task distribution and state synchronization between the computation scheduling unit and the intelligent inference engine. The tile rendering engine, deployed on a CPU computing node, is used to dynamically visualize and render the received raw data blocks, generating a visual base map. The intelligent inference engine is deployed on a GPU computing node and embeds a deep learning pre-trained model to perform real-time feature extraction and interpretation of the received raw data blocks, generating intelligent interpretation results. The service fusion unit is used to fuse the visualization base map with the intelligent interpretation result and return it to the client as a service response.

[0016] In summary, the present invention has at least one of the following beneficial technical effects: 1. This invention enables online intelligent reasoning of remote sensing data with "instant WYSIWYG" results, significantly improving the timeliness of geospatial information acquisition. By constructing a collaborative computing architecture based on the MIP protocol on the server side, the system can simultaneously complete visual rendering of images and feature extraction from deep learning models within milliseconds of responding to front-end tile requests. This means that while browsing maps, users can obtain high-value intelligent analysis results such as land cover classification and change detection in real time without going through the cumbersome process of "data download, local computation, and result feedback." This has revolutionary significance for time-sensitive application scenarios such as disaster emergency response and real-time monitoring.

[0017] 2. The real-time tile rendering and inference technology employed in this invention fundamentally solves the problems of massive storage redundancy and data update lag inherent in traditional tile map services. In traditional technologies, to meet the browsing needs of different levels, hundreds of millions of static tile images must be pre-generated. This not only occupies enormous storage space, but also requires full tile re-slicing once the original images are updated or the model algorithm is upgraded, resulting in extremely high maintenance costs. In contrast, this invention utilizes vector pyramid indexing and dynamic resampling technology to perform on-demand calculations directly based on the original raster data, saving over 70% of storage resources and ensuring that users always access the latest data and model calculation results, significantly reducing system operation and maintenance costs.

[0018] 3. This invention significantly expands the depth of human-computer interaction and the flexibility of analysis in map services. Traditional static tile services can only provide fixed visual outputs, while this invention service-orientedizes the inference process, transforming map services from a simple "display tool" into an interactive "analysis tool." Users can adjust inference parameters (such as confidence thresholds and target category filtering) in real time on the client side, and the server can immediately provide the corresponding calculation results. This dynamic interactive capability allows the same data source to adapt to the personalized needs of different users, thereby greatly enhancing the application value and user experience of geographic information services in vertical fields such as precision agriculture and urban planning. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the method flow provided in an embodiment of the present invention; Figure 2 A system architecture block diagram provided for embodiments of the present invention; Figure 3 This is a schematic diagram of intelligent agricultural crop growth monitoring provided in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0021] like Figure 1 As shown, this invention provides a method for Real-Time Raster Tiling Inference Map Service (RTIMS), comprising: Step 1: The service system receives the tile request and calculates the corresponding geospatial coordinate range based on the tile request; Step 2: Based on the geospatial coordinate range, read the original data blocks covering that range from the distributed storage; Step 3: Based on the Map Inference Communication Protocol (MICP), the raw data blocks are distributed to the CPU rendering pipeline and the GPU inference pipeline for parallel processing; wherein, the MICP protocol is an application layer protocol specifically used for task distribution and state synchronization between the computing scheduling unit and the intelligent inference engine. Step 4: Merge the visualization basemap generated by the CPU rendering pipeline with the intelligent interpretation results generated by the GPU inference pipeline, and return it as a service response.

[0022] In step 1, the service system receives a tile request and calculates the corresponding geospatial coordinate range based on the tile request, specifically: The service receives tile requests from the front end, typically following standard Tile Map Service (TMS) or Open Geospatial Consortium (OGC) standards. The request parameters include ( z , x , y ),in, x For tile column numbering, y For tile row numbers, z To scale the hierarchy; The system first needs to calculate the bounding box of the tile in a geographic coordinate system (usually WGS84, EPSG: 4326) to index data from storage later. Assuming the map uses the industry-standard WebMercator projection (EPSG: 3857), the longitude λ and latitude of the tile's top-left corner (North-West Corner) are... The calculation formula is as follows: ; ; In the formula, For tile size (usually 256 or 512); Convert it to geographic coordinates as follows: ; ; In the formula, For target longitude, For the target latitude, This represents the total number of slices. The present invention provides Table 1, which records or explains the value range of the parameters involved in step 1; Table 1 Explanation of Formula Parameters and Symbols in Step 1 In step 2, based on the geospatial coordinate range, the original data blocks covering that range are read from the distributed storage, specifically: To achieve "real-time tiles," the system does not directly read pre-generated images but instead reads the original GeoTIFF or other raster data. This invention abandons the traditional raster pyramid pre-tiling scheme and innovatively proposes Vector Pyramid Index (VPI) technology to establish a quadtree index of the original image. Based on the geographical range calculated in step 1, the original data block covering the area is quickly located in distributed storage. VPI only stores the bounding box, physical storage pointer, and metadata of the data, without storing the actual pixel data. This structure makes the storage overhead of the multi-level index only about 1% of the traditional scheme (VPI / Raster≈0.01), and completely preserves the band information of the original data, supporting dynamic access of arbitrary resolution and band combinations. When accessing data dynamically, a smart caching strategy is used to check if the cached data for that region already exists in memory or video memory. The cache key is generated by the hierarchy, coordinates, and task type (KEY=MD5({ z , x , y If the cache is hit and the result is within its validity period, the system returns the result directly; otherwise, the real-time inference process is triggered.

[0023] In step 3, based on the MICP protocol, the raw data blocks are distributed to the CPU rendering pipeline and the GPU inference pipeline for parallel processing, specifically as follows: Step 301: The computation scheduling unit encapsulates the tile request into a MICP message. The header of the MICP message includes version, type, task ID, dynamic priority, data length, and checksum fields. The body of the MICP message includes encrypted geographic coordinates and tensor data, specifically: Because standard network protocols (such as HTTP) have limitations when handling high-concurrency, low-latency heterogeneous computing tasks (CPU rendering vs. GPU inference), such as large header overhead, inability to carry task priority metadata, and lack of heartbeat synchronization mechanism, this invention specifically designs a lightweight Map Inference Communication Protocol (MICP). Its message header occupies only 13 bytes and embeds task ID, priority, checksum, and encryption identifier. It is specifically designed to achieve sub-millisecond task distribution and state synchronization between the computing scheduling unit and the intelligent inference engine. The computational scheduling unit encapsulates tile requests into MICP messages. The message format is defined as a binary structure with a fixed 16-byte header and a variable-length body. Message types include: 0x01 (task request), 0x02 (result return), and 0x03 (heartbeat detection). The complete format is mathematically represented as follows: ; The header field is shown in Table 2 (the symbol "||" indicates concatenation). Table 2 Header Fields Table

[0024] The dynamic priority The calculation formula is: ; In the formula, α and β are weighting coefficients (default α=0.6, β=0.4), Distance represents the Euclidean distance (in pixels) between the tile and the center of the user's viewport, and Timeliness is the task timeliness factor (value 0-1).

[0025] The body contains encrypted geographic coordinates and tensor data, encrypted using AES-256.

[0026] The MICP protocol can effectively solve resource contention problems. This implementation process is used within collaborative computing units. Based on the urgency of the task, it determines whether to use a "synchronous wait" mode or an "asynchronous pipeline" mode. The synchronous mode ensures real-time performance and is suitable for high-priority tasks at the center of the user's current view, while the asynchronous mode improves computational throughput and is suitable for tasks in pre-loaded or edge regions.

[0027] Step 302: Based on dynamic priority, distribute tasks to target GPU nodes through a dynamic routing mechanism, specifically as follows: Based on the priority field of the MICP message, the system uses a routing formula to select the GPU node, the calculation formula of which is as follows: ; in, Indicates the first Task queue length per GPU node GPUUtil i For utilization, γ is an adaptive coefficient (default γ=0.5). If Priority>0.8, routing is done to high-performance nodes; otherwise, routing is done to edge nodes. This routing strategy comprehensively considers spatial locality (distance from the center of the user's viewport) and time sensitivity (such as disaster monitoring tasks requiring higher priority), which is something that general load balancing algorithms (such as round-robin and minimum connection count) cannot achieve.

[0028] Dynamic routing mechanisms can effectively resolve the contradiction between high concurrency in map services and high computational consumption in deep learning. This implementation is a dynamic closed loop involving both hardware and software collaboration. A monitoring agent collects real-time data on the software queuing depth and hardware utilization of each node in the cluster. Upon receiving a new task, these two heterogeneous metrics are weighted using a weighting coefficient β to synthesize a unified load cost score. The argmin algorithm then identifies the node with the lowest overall cost for task distribution and immediately performs local state pre-increment to prevent the "thundering herd effect" under high concurrency. This ensures that inference tasks are always routed to the most available computing resources, achieving globally optimal resource utilization and reducing collaborative computing latency.

[0029] Step 303: The CPU rendering pipeline and the GPU inference pipeline process the same raw data block in parallel, specifically as follows: After the GPU node decrypts the message, it performs inference, while the CPU rendering pipeline runs in parallel. Specifically, the system defines the raw data blocks read from distributed storage as... (Raw Data Matrix) and distribute it to two parallel pipelines for processing.

[0030] Assume the original geospatial raster data blocks read are a three-dimensional matrix. Its dimensions are represented as H×W×C, where H is the height of the data block (number of pixel rows), W is the width of the data block (number of pixel columns), and C is the number of bands (e.g., 3 for RGB images, 4 or more for multispectral images).

[0031] Pipeline A: CPU Visualization Pipeline. This pipeline aims to convert raw high dynamic range (HMR) or multi-band data into 8-bit RGB images suitable for web display. The specific process is as follows: (1) Resampling: Since the resolution of the original image is inconsistent with the resolution of the screen display, dynamic resampling is required (such as bilinear interpolation or cubic convolution interpolation). (2) Radiometric correction and stretching: Histogram stretching or lookup table (LUT) mapping is performed on the DN values ​​of the remote sensing image to convert it into an RGB image matrix suitable for human viewing. ; (3) Slicing and cutting: Cut the processed large image into standard tile size.

[0032] Pipeline B: GPU Deep Learning Inference Pipeline. This pipeline utilizes the parallel computing capabilities of the GPU to process deep learning inference data. The specific process for land cover identification or feature extraction is as follows: (1) Tensor transformation: transforming tensor into ... Convert to the tensor format required for deep learning models Then, normalize the values ​​(e.g., map the values ​​to [0, 1] or standardize them) to obtain a normalized tensor. ; (2) Model reasoning: Input is a deep learning model (such as U-Net, DeepLabV3+, etc.) preloaded into GPU memory. (3) Forward computation: The model outputs a probability map or feature map. For example, for binary classification problems (such as extracting buildings), the output is the probability P (pixel∈Building) that each pixel belongs to a building; (4) Threshold segmentation and vectorization (optional): Set a threshold θ to convert the probability map into a binary mask, or further use a contour tracking algorithm (such as Marching Squares) to generate vector polygons. ; Step 304: Maintain the state consistency of each node through the heartbeat mechanism of the MICP protocol, specifically as follows: Consistency is ensured through the MICP heartbeat mechanism (sending a type 0x03 message every 50ms). The total time consumption is calculated using the following formula: ; in, The total service response time, The processing time for pipeline A (rendering), The processing time for pipeline B (inference) For the time consumed by result fusion (e.g., overlaying an inference mask onto an RGB image), This is the overhead for the MICP protocol (measured <2ms).

[0033] Step 305: When an error is detected, execute the hierarchical error handling strategy, specifically as follows: Error handling in the MICP protocol is a core component for ensuring high system availability. It employs a multi-level error handling framework, including error classification, detection mechanisms, handling strategies, and system recovery. The error handling process is tightly coupled with MICP message types, specifically as follows: (1) Error classification and detection mechanism The MICP protocol categorizes errors into three types and detects them in real time through message verification and status heartbeats: ① Communication error: message loss or timeout (e.g., network jitter).

[0034] ② Node error: GPU node failure or overload.

[0035] ③ Data error: Checksum failure or decryption error.

[0036] Error detection relies on the following mathematical conditions (with a threshold parameter): ①Timeout detection: If the response time (default ), trigger error flag ; ② Checksum verification: Checksum of the received message Compared with calculated values The comparison is based on the following error conditions: ; Abnormal heartbeat: If continuous If the next heartbeat message is not responded to, the node is determined to be faulty. ).

[0037] (2) Hierarchical processing strategy A dynamic hierarchical processing strategy is adopted for different error types, as shown in Table 3. Table 3 Dynamic Hierarchical Processing Strategy Table (3) Downgrade and recovery mechanism Regarding the degradation strategy, when all GPU nodes become unavailable, the system automatically degrades to CPU inference mode (using a lightweight model). As for the recovery strategy, after a node recovers, it re-registers via heartbeat messages, and the system gradually migrates tasks back to the GPU to avoid instantaneous overload.

[0038] In step 4, the visualization basemap generated by the CPU rendering pipeline and the intelligent interpretation result generated by the GPU inference pipeline are fused and returned as a service response, specifically: The system will generate pipe A (Visual image) and generated by pipe B or (The reasoning results) are encapsulated; This invention provides two modes, namely: Mode 1 (Overlay Rendering Mode): The inference results are overlaid as a semi-transparent layer on the server side. The process involves combining the above data into a new image and returning it to the client. Mode 2 (Data Separation Mode): Returns a response package containing multiple parts of data (such as MultipartResponse), including map tile images and inference results in GeoJSON format, which are overlaid and rendered by the client browser.

[0039] like Figure 2As shown, the present invention also provides a real-time raster tile inference map service system, applied to the above-mentioned real-time raster tile inference map service method, comprising: a data storage unit, a computing scheduling unit, a tile rendering engine, an intelligent inference engine, and a service fusion unit connected in sequence; The data storage unit is used for distributed storage of raw geospatial raster data and provides a data reading interface based on a vector pyramid index. The computation scheduling unit, as the system control center, is used to receive client tile requests, calculate the corresponding geospatial coordinate range based on the tile request, read the original data block covering the range from the data storage unit based on the geospatial coordinate range, and distribute the original data block to the tile rendering engine and the intelligent inference engine based on the MICP protocol; wherein, the MICP protocol is an application layer protocol specifically used for task distribution and state synchronization between the computation scheduling unit and the intelligent inference engine. The tile rendering engine, deployed on a CPU computing node, is used to dynamically visualize and render the received raw data blocks, generating a visual base map. The intelligent inference engine is deployed on a GPU computing node and embeds a deep learning pre-trained model to perform real-time feature extraction and interpretation of the received raw data blocks, generating intelligent interpretation results. The service fusion unit is used to fuse the visualization base map with the intelligent interpretation result and return it to the client as a service response.

[0040] To further illustrate the technical solution of this invention and its effectiveness in practical applications, a detailed description is provided below using the specific application scenario of "intelligent agricultural crop growth monitoring." In this embodiment, the system backend is deployed on a high-performance computing cluster in the cloud, configured with a distributed object storage system for storing massive amounts of high-resolution multispectral satellite imagery (including red, green, blue, and near-infrared bands), and equipped with computing nodes featuring professional-grade deep learning graphics cards (NVIDIA T4) to support RTIMS services. The frontend is a standard Web GIS browser interface, specifically as follows... Figure 3 As shown; When agricultural users access the monitoring platform via a browser and zoom in on a specific farm area they manage (e.g., zoom level 16), the front end automatically sends a series of tile requests to the server, including the zoom level and row / column numbers. Upon receiving the request, the RTIMS service's computational scheduling unit first parses the corresponding geospatial range based on the request parameters and then uses vector pyramid indexing technology to quickly locate and read the latest multispectral raw data block covering that farm area from distributed storage. At this point, the system no longer searches for pre-cut images but directly obtains the raw GeoTIFF data stream.

[0041] Subsequently, the system immediately initiated a synchronous parallel processing flow. On one hand, the CPU rendering pipeline extracted the red, green, and blue visible light bands from the raw data, performed radiometric calibration, histogram stretching, and color correction, generating a true-color RGB base map reflecting the actual topography of the farmland. On the other hand, the GPU inference pipeline simultaneously received this multispectral data, normalized it, and input it into the "crop health assessment model" pre-loaded in the video memory. This model is built based on a deep convolutional neural network (such as the U-Net architecture), uses the near-infrared band to calculate the normalized vegetation index (NDVI) in real time, and identifies specific areas where crop growth is hindered, water is lacking, or diseases occur based on preset agronomic model parameters, generating corresponding classification probability masks.

[0042] Finally, the service fusion unit overlays the CPU-generated real-world base map with the GPU-generated semi-transparent vegetation health heatmap (e.g., using red to highlight diseased areas and green to display healthy areas), and returns the composite to the client instantly via an HTTP response. Users not only see clear images of the farmland on their browser's map but also intuitively view the crop growth analysis layer precisely overlaid on the image. The entire process is completed instantly as the user drags and zooms the map, eliminating the need for manual area selection or waiting for offline reports, achieving a highly efficient "browse-and-analyze" monitoring experience.

[0043] Tested on 10,000 z=16 level tile requests under the same hardware environment (NVIDIA T4×4, 10GbE network): The average response time of this invention in RTIMS is 87ms; in the traditional pre-slicing + offline inference mode (requiring download before analysis): >5s; in client-side TensorFlow.js inference (Chrome browser): >2.3s, with a memory overflow rate of 12%. Furthermore, in terms of storage, this invention eliminates the need to pre-generate inference result tiles, saving 73.6% of storage space compared to a full preprocessing solution.

[0044] Embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0045] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0046] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0047] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0048] Contents not described in detail in this specification are prior art known to those skilled in the art. It is hereby indicated that the above description is intended to help those skilled in the art understand this invention, but does not limit the scope of protection of this invention. Any equivalent substitutions, modifications, improvements, or simplifications of the above descriptions that do not depart from the essential content of this invention fall within the scope of protection of this invention.

Claims

1. A method for real-time raster tile inference map service, characterized in that, include: Step 1: The service system receives the tile request and calculates the corresponding geospatial coordinate range based on the tile request; Step 2: Based on the geospatial coordinate range, read the original data blocks covering that range from the distributed storage; Step 3: Based on the MICP protocol, distribute the raw data blocks to the CPU rendering pipeline and the GPU inference pipeline for parallel processing; wherein, the MICP protocol is an application layer protocol specifically used for task distribution and state synchronization between the computing scheduling unit and the intelligent inference engine. Step 4: Merge the visualization basemap generated by the CPU rendering pipeline with the intelligent interpretation results generated by the GPU inference pipeline, and return it as a service response.

2. The real-time raster tile inference map service method according to claim 1, characterized in that, In step 1, the tile request parameters are ( z , x , y ),in, x For tile column numbering, y For tile row numbers, z To scale the hierarchy.

3. The real-time raster tile inference map service method according to claim 2, characterized in that, In step 1, the corresponding geospatial coordinate range is calculated based on the tile request, specifically as follows: The pixel coordinates of the top-left corner of the tile are calculated as follows: ； ； In the formula, For tile dimensions; Convert it to geographic coordinates as follows: ； ； In the formula, For target longitude, For the target latitude, This represents the total number of slices.

4. The real-time raster tile inference map service method according to claim 3, characterized in that, In step 2, based on the geospatial coordinate range, the original data blocks covering that range are read from the distributed storage, specifically: Based on the geospatial coordinate range, the original data block is located and read from the distributed storage using vector pyramid indexing technology; the vector pyramid index only stores the bounding box, physical storage pointer and metadata of the original data block.

5. The real-time raster tile inference map service method according to claim 4, characterized in that, In step 3, based on the MICP protocol, the raw data blocks are distributed to the CPU rendering pipeline and the GPU inference pipeline for parallel processing, specifically as follows: Step 301: The computation scheduling unit encapsulates the tile request into a MICP message. The header of the MICP message contains version, type, task ID, dynamic priority, data length and checksum fields, and the body of the MICP message includes encrypted geographic coordinates and tensor data. Step 302: Based on dynamic priority, distribute tasks to target GPU nodes through a dynamic routing mechanism; Step 303: The CPU rendering pipeline and the GPU inference pipeline process the same raw data block in parallel; Step 304: Maintain the state consistency of each node through the heartbeat mechanism of the MICP protocol; Step 305: When an error is detected, execute the hierarchical error handling strategy.

6. The real-time raster tile inference map service method according to claim 5, characterized in that, The dynamic priority The calculation formula is: ； In the formula, α and β are weighting coefficients, Distance represents the Euclidean distance between the tile and the center of the user's view, and Timeliness is the task timeliness factor.

7. The real-time raster tile inference map service method according to claim 6, characterized in that, In step 303, the CPU rendering pipeline and the GPU inference pipeline process the same raw data block in parallel, specifically as follows: The CPU rendering pipeline performs dynamic resampling, radiometric correction and stretching, and slicing on the original data blocks sequentially to generate an 8-bit RGB visualization base map; the GPU inference pipeline performs tensor transformation and normalization on the original data blocks, inputs them into a preloaded deep learning model for forward inference, and outputs a probability map or feature map.

8. The real-time raster tile inference map service method according to claim 7, characterized in that, In step 305, the graded error handling strategy specifically includes: For communication errors, the exponential backoff algorithm is used for retransmission; If a node fails, perform a failover and route the task to a backup node; For data errors where the checksum fails, discard the message and trigger a retransmission.

9. The real-time raster tile inference map service method according to claim 8, characterized in that, In step 4, the visualization basemap generated by the CPU rendering pipeline and the intelligent interpretation result generated by the GPU inference pipeline are fused and returned as a service response, specifically: Using an overlay rendering mode, the intelligent interpretation results are overlaid as a semi-transparent layer on the visualization base map to synthesize a new image as a service response. Using a data separation mode, a response package containing multiple data parts is returned. The response package includes a visual base map and intelligent interpretation results, which are overlaid and rendered by the user.

10. A real-time raster tile inference map service system, applied to the real-time raster tile inference map service method according to any one of claims 1-9, characterized in that, include: The data storage unit, computing scheduling unit, tile rendering engine, intelligent inference engine, and service fusion unit are connected in sequence. The data storage unit is used for distributed storage of raw geospatial raster data and provides a data reading interface based on a vector pyramid index. The computation scheduling unit, as the system control center, is used to receive client tile requests, calculate the corresponding geospatial coordinate range based on the tile request, read the original data block covering the range from the data storage unit based on the geospatial coordinate range, and distribute the original data block to the tile rendering engine and the intelligent inference engine based on the MICP protocol; wherein, the MICP protocol is an application layer protocol specifically used for task distribution and state synchronization between the computation scheduling unit and the intelligent inference engine. The tile rendering engine, deployed on a CPU computing node, is used to dynamically visualize and render the received raw data blocks, generating a visual base map. The intelligent inference engine is deployed on a GPU computing node and embeds a deep learning pre-trained model to perform real-time feature extraction and interpretation of the received raw data blocks, generating intelligent interpretation results. The service fusion unit is used to fuse the visualization base map with the intelligent interpretation result and return it to the client as a service response.

Images