Building remote sensing interpretation method and device based on streaming processing and electronic equipment

Abstract

The invention relates to the technical field of data processing, and discloses a building remote sensing interpretation method and device based on streaming processing and electronic equipment, and the method comprises the steps: obtaining multi-source heterogeneous geographic data in a target region, dividing the multi-source heterogeneous geographic data into a plurality of target sub-rectangular regions based on a preset reference coordinate system, generating a multi-band sample of each target sub-rectangle in parallel; performing graph AI model reasoning and vector conversion operation on each target sub-rectangle multi-band sample in sequence to obtain a geometrically complete building contour vector; and performing attribute AI model reasoning based on a preset target attribute, and performing multi-stage streaming attribute fusion based on the geometrically complete building contour vector to generate a vector interpretation result with attributes in the target area. According to the method, the problems of two engineering bottlenecks of low reasoning speed and high storage occupation when massive high-resolution remote sensing data is processed in the prior art are effectively solved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and specifically to a method, apparatus, and electronic device for remote sensing interpretation of buildings based on streaming processing. Background Technology

[0002] Significant engineering performance bottlenecks arise when applying large-scale remote sensing models for nationwide high-resolution remote sensing interpretation. Taking nationwide building interpretation as an example (building location, roof, height, and use), it is necessary to create the samples required for the large-scale remote sensing model based on 350 million Level 18 remote sensing images and other heterogeneous multi-source data, perform inference, and then parse the vectors on the raster inference results to form the final analysis results.

[0003] In the process of interpreting and reasoning from this massive amount of remote sensing data, two main types of problems will arise: 1. Slow inference speed. Even with the purchase of 8 A6000 machines, on a remote sensing model with 100 million parameters in the ViT (VisionTransformer, computer vision) architecture, without performance optimization, it still takes more than 2 years for continuous inference, which is too slow and cannot meet the timeliness requirements of the business.

[0004] 2. High storage consumption. Even with image filtering, only 54% of the land area of ​​the target region can be inferred. The total sample size for inferring building locations, rooftops, heights, and uses will exceed hundreds of terabytes. Excessive storage consumption will consume excessive storage resources and require significant inter-server communication bandwidth, which is detrimental to practical deployment. Summary of the Invention

[0005] This invention provides a method, apparatus, and electronic device for interpreting building remote sensing data based on streaming processing, in order to solve the two major engineering bottlenecks of slow inference speed and high storage consumption faced by existing technologies when processing massive amounts of high-resolution remote sensing data.

[0006] In a first aspect, the present invention provides a building remote sensing interpretation method based on streaming processing, the method comprising: Acquire multi-source heterogeneous geographic data within the target area, and based on a preset reference coordinate system, divide the multi-source heterogeneous geographic data into multiple target sub-rectangular areas, and generate multi-band samples for each target sub-rectangular area in parallel. After performing graphical AI model inference and vectorization operations on each target sub-rectangle multi-band sample, a geometrically complete building outline vector is obtained. The system performs attribute AI model reasoning based on preset target attributes, and performs multi-stage streaming attribute fusion based on geometrically complete building outline vectors to generate vector interpretation results with attributes for the target area.

[0007] This invention provides a building remote sensing interpretation method based on streaming processing, which realizes asynchronous parallel processing of image parsing and multiple attribute assignment. It allows AI models with different attributes to perform inference and fusion independently and asynchronously based on computing resources, avoiding overall process blockage caused by delays in single attribute inference or data waiting. At the same time, the cross-stage attribute inheritance mechanism ensures accurate matching and lossless integration of multi-source attributes in the streaming process, thereby increasing the system throughput to several times that of the traditional synchronous mode while ensuring interpretation accuracy. This significantly shortens the cycle of nationwide data processing. By introducing a multi-stage streaming attribute fusion mechanism, it effectively solves the strong coupling bottleneck between attribute inference and geometric processing in large-scale remote sensing interpretation, and greatly reduces storage requirements.

[0008] In one optional implementation, based on a preset reference coordinate system, the multi-source heterogeneous geographic data is divided into multiple target sub-rectangular regions, and multi-band samples for each target sub-rectangular region are generated in parallel, including: Multi-source heterogeneous geographic data is uniformly aligned to a preset reference coordinate system, and the multi-source heterogeneous geographic data is divided into multiple independent data files according to a preset rectangle size for storage. Each data file is uniquely identified and indexed by its row and column numbers in the preset reference coordinate system. The boundaries of the target area are determined. Based on the input image size required by the AI ​​model, the boundaries are divided into a regular grid, where each regular grid cell is a target sub-rectangle. The row and column numbers of each target sub-rectangle in the preset reference coordinate system are calculated. Based on the row and column numbers of each target sub-rectangle, target sub-rectangles that intersect with the geographical boundary of the target area are selected from all grid cells to generate the final target sub-rectangle task list; The target sub-rectangle task list is traversed in parallel. Based on the row and column number of each target sub-rectangle, the required multi-source data files are retrieved and extracted. Through splicing and synthesis, the corresponding multi-band samples are generated.

[0009] This invention provides a building remote sensing interpretation method based on streaming processing, which significantly improves the sample generation efficiency under massive data through precise row and column number mapping and fast retrieval; combined with task list filtering and parallel processing, it ensures high utilization of computing resources, provides a stable and high-speed sample stream for subsequent AI model inference, and realizes efficient organization and parallel reading of multi-source heterogeneous geographic data, thereby supporting the high-performance execution of large-scale remote sensing interpretation tasks from the source.

[0010] In one optional implementation, the target sub-rectangle task list is traversed in parallel, and the required multi-source data files are retrieved and extracted based on the row and column numbers of each target sub-rectangle. These files are then spliced ​​and synthesized to generate corresponding multi-band samples, including: Based on the row and column numbers of each target sub-rectangle, all relevant data files are located from each source data, and all located data files are spliced ​​together to form the source data large rectangle corresponding to each source data. Calculate the coordinate offset of the target sub-rectangle relative to each source data rectangle, and based on the offset and the size of the target sub-rectangle, cut out the corresponding target area data block from each source data rectangle; The target area data blocks from different source data are spliced ​​and fused along the band dimension to generate the final multi-band sample of the corresponding target sub-rectangle.

[0011] This invention provides a building remote sensing interpretation method based on streaming processing. It reduces the I / O overhead of frequently reading small files by batch stitching large rectangles of source data; it achieves pixel-level precise positioning and cropping through a coordinate offset mechanism, ensuring the spatial accuracy of the samples; and it achieves seamless fusion of multi-source data in the band dimension, ensuring the integrity of the input data and perfectly adapting to the requirements of subsequent AI models for multi-band input formats. Thus, it supports the stable operation of large-scale streaming interpretation tasks in terms of both accuracy and efficiency. Through a three-level processing architecture of file-level stitching, precise coordinate positioning, and band fusion, it achieves efficient conversion of multi-source geographic data into standard samples.

[0012] In one optional implementation, after sequentially performing graphical AI model inference and vectorization operations on each target sub-rectangle multi-band sample, a geometrically complete building outline vector is obtained, including: Input the multi-band sample of each target sub-rectangle into the preset AI model to perform graphics task inference, and output the graphic raster corresponding to each target sub-rectangle; The graphic raster of each target sub-rectangle is converted into a vector graphic. The vector graphics located at the boundary of the target sub-rectangle are then merged to obtain a geometrically complete building outline vector. A unique identifier is assigned to each building outline vector.

[0013] The present invention provides a building remote sensing interpretation method based on streaming processing. After performing graphic AI model inference and vectorization operations on standardized samples using a preset model, a geometrically complete building outline vector is obtained. This step does not store intermediate files after graphic AI model inference, which greatly reduces server storage space and saves storage resources.

[0014] In one optional implementation, the graphic raster of each target sub-rectangle is converted into a vector graphic, the vector graphics located at the boundaries of the target sub-rectangles are merged to obtain a geometrically complete building outline vector, and a unique identifier is assigned to each building outline vector, including: Extract the graphic raster data of each target sub-rectangle, and filter it according to the preset invalid raster values ​​of building targets, deleting graphic rasters that do not contain valid building information; Perform a raster-to-vector conversion operation on all valid graphic rasters to generate an initial set of vector graphics; the initial set of vector graphics includes interior vectors that are completely inside the sub-rectangle, as well as boundary vectors that intersect the boundaries of the sub-rectangle; Identify and mark all boundary vectors, and perform vector intersection fusion operation on the boundary vectors of all target sub-rectangles to merge broken vectors that cross target sub-rectangles and belong to the same feature into a complete geometry; All internal vectors of each target sub-rectangle are merged with the fused boundary vectors to generate a geometrically complete building outline vector within the target sub-rectangle area, and a unique identifier is assigned to each building outline vector.

[0015] This invention provides a building remote sensing interpretation method based on streaming processing. Through a boundary vector fusion mechanism, it completely eliminates the problem of fragmented building outlines caused by block processing, ensuring the geometric integrity of the vector results. More importantly, by assigning a unique identifier to each building outline, a unified identity authentication system is established that runs through subsequent multi-stage processing. This allows attribute data from different sources and time series to be accurately associated with the corresponding buildings based on the identifier, thus strongly supporting the implementation of a streaming, asynchronous attribute fusion architecture. This step, through a processing chain of effective raster filtering, internal and external vector separation, boundary fusion, and global numbering, lays a crucial foundation for subsequent attribute fusion while achieving geometric integrity.

[0016] In one optional implementation, attribute AI model inference is performed based on preset target attributes, and multi-stage streaming attribute fusion is performed based on geometrically complete building outline vectors to generate vector interpretation results with attributes for the target area, including: Perform attribute AI model reasoning on the current target attribute to generate a current stage attribute raster covering the target area; Convert the attribute raster into attribute vectors, with each attribute vector carrying the attribute label for the current stage; Based on spatial location relationships and unique identifiers, a multi-stage streaming attribute fusion method is used to integrate attribute information from attribute vectors into the corresponding building outline vectors, generating attribute-bearing vector interpretation results for the target area.

[0017] This invention provides a streaming-based remote sensing interpretation method for buildings. Employing a phased streaming processing mechanism, it allows different attribute models to run independently and in parallel, completely eliminating the constraint of synchronous processing of multiple attributes required in traditional methods. This significantly improves system resource utilization and overall interpretation efficiency. Simultaneously, through a dual association mechanism of unique identifier numbers and spatial locations, it ensures accurate and reliable matching of massive vector data and multi-source attribute data. Ultimately, while guaranteeing the accuracy of attribute assignment, it achieves automated, pipelined fusion from geometric figures to multi-dimensional semantic information. This step, by constructing a streaming fusion pipeline of attribute reasoning, vectorization, and number association, enables efficient asynchronous integration of attribute information in large-scale remote sensing interpretation.

[0018] In one optional implementation, based on spatial location relationships and the unique identifier, a multi-stage streaming attribute fusion method is used to fuse attribute information from the attribute vector into the corresponding building outline vector, generating an attribute vector interpretation result for the target area, including: Establish a spatial index for the geometrically complete architectural outline vector; Traverse all target sub-rectangular regions corresponding to the current attribute grid, and perform spatial matching and attribute surface accumulation of the building outline vector and attribute vector within each target sub-rectangular region based on the building outline vector and the corresponding unique identifier; Merge the attribute surface accumulation results of all target sub-rectangular regions, and for the building outline vector corresponding to the same identifier number, determine the attribute category with the largest accumulated area as the final attribute of the building outline vector, and write the final attribute into the building outline vector corresponding to the unique identifier number. Each preset target attribute is fused sequentially until all target attributes are fused, resulting in a target region with attribute vector interpretation containing all attribute information.

[0019] This invention provides a building remote sensing interpretation method based on streaming processing. Utilizing a dual mechanism of spatial indexing and unique identifier numbering, it ensures rapid and error-free association between massive vector and attribute data, effectively solving the challenge of cross-regional attribute matching. Through a dominant attribute decision-making strategy based on area accumulation, it guarantees the rationality and accuracy of attribute assignment. Finally, leveraging a multi-stage streaming processing architecture, it achieves asynchronous, pipelined integration of various building attributes, significantly improving the conversion efficiency from geometric figures to complete semantic results and system throughput. This step, through a technical chain of accelerated retrieval via spatial indexing, precise matching via number association, area-dominant decision-making, and multi-stage iterative fusion, achieves efficient, accurate, and automated assignment of building attribute information over a wide range.

[0020] In one optional implementation, spatial matching and attribute surface accumulation of building outline vectors and attribute vectors within each target sub-rectangular area are performed based on the building outline vector and its corresponding unique identifier, including: Based on the spatial range of the target sub-rectangle, all building outline vectors that intersect with the current attribute vector are retrieved through spatial index, and the unique identifier number corresponding to each building outline vector is recorded. For each retrieved building outline vector, perform spatial intersection analysis with the attribute vector within the current target sub-rectangle, calculate the intersection area of ​​each intersection part, and group and accumulate the intersection areas according to attribute category and unique identifier number.

[0021] This invention provides a building remote sensing interpretation method based on streaming processing. It utilizes spatial indexing to quickly locate the outline of the building to be processed, greatly improving retrieval efficiency; it accurately associates the attribute contributions of the same building in different sub-regions through unique identifiers, ensuring accurate data attribution; and finally, based on the quantitative accumulation of intersecting areas, it provides objective and accurate numerical basis for subsequent attribute decisions, thus balancing processing speed and fusion accuracy in complex spatial data processing.

[0022] Secondly, the present invention provides a building remote sensing interpretation device based on streaming processing, the device comprising: The data acquisition and multi-band sample generation module is used to acquire multi-source heterogeneous geographic data within the target area, and based on a preset reference coordinate system, divide the multi-source heterogeneous geographic data into multiple target sub-rectangular areas, and generate multi-band samples for each target sub-rectangular area in parallel. The AI ​​inference and vector conversion module is used to sequentially perform graphical AI model inference and vector conversion operations on each target sub-rectangle multi-band sample to obtain a geometrically complete building outline vector. The attribute vector generation module is used to perform attribute AI model reasoning based on preset target attributes, and to perform multi-stage streaming attribute fusion based on geometrically complete building outline vectors to generate attribute vector interpretation results for the target area.

[0023] Thirdly, the present invention provides an electronic device, comprising: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the streaming-based building remote sensing interpretation method of the first aspect or any corresponding embodiment described above.

[0024] Fourthly, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to execute the streaming-based building remote sensing interpretation method of the first aspect or any corresponding embodiment described above.

[0025] Fifthly, the present invention provides a computer program product, including computer instructions for causing a computer to execute the streaming-based building remote sensing interpretation method of the first aspect or any corresponding embodiment described above. Attached Figure Description

[0026] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0027] Figure 1 This is a schematic diagram of the first process of a building remote sensing interpretation method based on streaming processing according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the second process of the building remote sensing interpretation method based on streaming processing according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the third process of the building remote sensing interpretation method based on streaming processing according to an embodiment of the present invention. Figure 4 This is a flowchart of multi-band sample generation in the building remote sensing interpretation method based on streaming processing according to an embodiment of the present invention; Figure 5 This is a flowchart of the multi-raster file contour vector generation in the building remote sensing interpretation method based on streaming processing according to an embodiment of the present invention; Figure 6 This is a schematic diagram of efficient generation of attribute-based vectors in a building remote sensing interpretation method based on streaming processing according to an embodiment of the present invention; Figure 7 This is a structural block diagram of a building remote sensing interpretation device based on streaming processing according to an embodiment of the present invention; Figure 8 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0029] It is understood that before using the technical solutions disclosed in the various embodiments of the present invention, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in the present invention and their authorization should be obtained in accordance with relevant laws and regulations through appropriate means.

[0030] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0031] Existing remote sensing interpretation schemes divide a large target area into many sub-rectangles of the same size as the AI ​​model's input image, based on the size of the input image. After inference within these smaller rectangles, the raster results are fused into a vector result. Finally, the vector results from multiple smaller rectangles are merged into a single large vector result, which is the interpretation result for the target area. Problems with existing schemes include: When merging multiple small rectangular vector results, the same vector (such as the roof of a building) is divided into multiple rectangles. No literature has clearly addressed how to fuse them back together during the merging process. Furthermore, some literature directly presents the unmerged results; although it represents a single building, the merged result is actually divided into multiple vector blocks, leading to slow inference speed and high storage consumption. Currently, there is no clear solution for vector merging after large-scale raster inference.

[0032] This invention provides a building remote sensing interpretation method based on streaming processing. By using a parallel sample generation mechanism under a unified coordinate system, a boundary-aware vector fusion strategy, and a synchronous attribute fusion technology designed for streaming scenes, it achieves an efficient and complete attribute-based vector streaming interpretation process without storing intermediate raster files. This significantly improves the processing efficiency of massive remote sensing data and effectively controls the consumption of storage resources.

[0033] According to an embodiment of the present invention, a method for remote sensing interpretation of buildings based on streaming processing is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0034] This embodiment provides a building remote sensing interpretation method based on streaming processing, which can be used in the aforementioned electronic devices. Figure 1 This is a flowchart of a building remote sensing interpretation method based on streaming processing according to an embodiment of the present invention, such as... Figure 1As shown, the process includes the following steps: Step S101: Obtain multi-source heterogeneous geographic data within the target area, and based on a preset reference coordinate system, divide the multi-source heterogeneous geographic data into multiple target sub-rectangular areas, and generate multi-band samples for each target sub-rectangular area in parallel.

[0035] Specifically, multi-source heterogeneous geographic data refers to geospatial data acquired from multiple different sources, differing in data format, structure, and physical meaning, used as input to an AI model to improve interpretation accuracy in order to complete architectural interpretation tasks. Specifically, it mainly includes the following two categories: Remote sensing imagery data: This is the core data source, consisting of high-resolution (sub-meter resolution) satellite or aerial remote sensing images covering the target area (such as the entire country). It provides visual, textural, and spectral information about ground features (such as buildings).

[0036] Other heterogeneous multi-source data: This is supplementary data introduced to assist AI models in making more accurate judgments, including: POI data, or Point of Interest data, contains location and type information for specific locations such as schools, shopping malls, and factories. It can help models infer the purpose of buildings (for example, if a building outline overlaps with multiple restaurant POIs, it is very likely to be a restaurant).

[0037] Probability maps / color temperature maps: These are raster data that have undergone preliminary processing or characterization of other features. For example, a probability map may be the probability of building presence derived from other models; a color temperature map may reflect the intensity and color of urban lights at night, indirectly indicating the commercial or industrial activity in a region.

[0038] The preset reference coordinate system can be the Web Mercator coordinate system (3857), GPS coordinate system (4326), etc. The purpose of the preset reference coordinate system is to provide a unified spatial benchmark for multi-source heterogeneous geographic data, thereby achieving accurate data alignment, efficient indexing, and seamless fusion, laying the foundation for subsequent parallel processing and result synthesis. Since AI models can only process a small, fixed-size image (e.g., 224x224 pixels) at a time, a complete "national map" cannot be processed. Therefore, the entire large multi-source heterogeneous geographic data is pre-cut into smaller pieces. Specifically, the operation is as follows: Dividing multi-source heterogeneous geographic data into multiple target sub-rectangular regions based on a preset reference coordinate system refers to using a regular grid to lay out a national map, dividing the entire target area (such as the whole country) into countless small, regular rectangular grids of identical size. Each small grid is a target sub-rectangular region. Simultaneously, multiple computing processes are launched to quickly retrieve the data required for the corresponding target sub-rectangular regions from different data source warehouses based on the previously established unified coordinate system and index, and then assemble them into the final multi-band sample.

[0039] In summary, by using a preset reference coordinate system to perform unified spatial benchmark transformation and gridded segmentation of multi-source heterogeneous geographic data, a standardized data index and task list are established. This step enables parallel reading of massive remote sensing data and rapid synthesis of multi-band samples, solving the problem of low data I / O efficiency at its source and providing a stable and high-speed input stream for subsequent processing.

[0040] Step S102: After performing graphical AI model inference and vector conversion operations on each target sub-rectangle multi-band sample, a geometrically complete building outline vector is obtained.

[0041] Specifically, the raster interpretation result refers to the preliminary recognition result in a pixel grid shape output by the AI ​​model after inference on the input multi-band samples. That is, after inputting standardized multi-band samples into the AI ​​model for graphic AI model inference and obtaining the corresponding graphic raster, a vector conversion operation is immediately performed, and the intermediate graphic raster is not stored.

[0042] A graphic raster is a binary or labeled raster image in which each pixel is assigned a value. For example, in building outline recognition, a pixel with a value of 1 might indicate that it is a building, while a pixel with a value of 0 might indicate that it is not a building (i.e., background). The graphic raster defines the spatial location, shape, and outline of features (such as buildings) and forms the basis for subsequent generation of vector boundaries.

[0043] The vector conversion operation refers to transforming the intermediate graphic raster into vector graphics. Vector graphics are data formats that represent geographic features using geometric shapes (such as points, lines, and polygons) defined by mathematical formulas. For a building, its vector graphic is typically a polygon, which is composed of a series of ordered vertices connected by geographic coordinates. The specific steps involved in its formation are as follows: The process identifies all connected pixel regions with the same values ​​(such as pixel blocks representing a single building) within the raster interpretation result of each target sub-rectangle. Using a specific algorithm, the boundaries of these pixel regions are extracted and converted into polygon vectors composed of a series of geographic coordinate points. Each pixel cluster identified as a building in the raster is transformed into an independent initial polygon vector graphic, providing the foundation for subsequent geometric blending and attribute assignment.

[0044] The initial vector graphics formed after vectorizing each target sub-rectangle are classified, and all boundary vectors intersecting the sub-rectangle boundaries are identified and marked. Simultaneously, internal vectors completely located within the sub-rectangles are confirmed. A vector fusion operation is specifically performed on all marked boundary vectors, merging fragmented polygons spanning different sub-rectangles but belonging to the same actual feature into a single, complete geometric entity. The seamless boundary vectors formed after fusion are then merged with the internal vectors in each sub-rectangle that do not require fusion, thereby outputting a geometrically complete building outline vector graphic covering the entire target area.

[0045] Step S103: Perform attribute AI model reasoning based on preset target attributes, and perform multi-stage streaming attribute fusion based on geometrically complete building outline vectors to generate vector interpretation results of the target area with attributes.

[0046] Specifically, multi-stage streaming attribute fusion is a "divide and conquer" pipeline operation mode that breaks down the complex task of interpreting building attributes into multiple stages of single attributes, and allows data to flow continuously and be processed asynchronously between stages like water.

[0047] For example: Phase 1: Specializing in processing and generating building outlines (geometry).

[0048] Phase Two: Specialized reasoning and integration of the first attribute (such as building use).

[0049] Phase 3: Specialized reasoning and integration of the second attribute (such as building height).

[0050] Once a sub-rectangle in the first stage completes its building outline generation, the generated vector can immediately enter the second stage for application fusion, without waiting for hundreds of millions of other sub-rectangles across the country to complete their outline generation. Similarly, the output of the second stage can immediately flow into the third stage. And so on, one stage for each attribute.

[0051] The attribute information (such as commercial) inferred in the current stage is accurately pasted onto the corresponding building outline vector through spatial calculation, thus completing the attribute fusion.

[0052] This step can be summarized as follows: Based on the building outline vector graphics, perform independent attribute fusion stages in sequence according to the preset attribute types; in each attribute fusion stage, perform model reasoning on the current target attribute, generate corresponding attribute data, and fuse the attribute data into the corresponding building outline vector through spatial position relationships; repeat this process until all preset attributes are fused, and finally output the target area vector interpretation result that integrates all the required attribute information.

[0053] The building remote sensing interpretation method based on streaming processing provided in this embodiment realizes asynchronous parallel processing of image parsing and multiple attribute assignments. It allows AI models with different attributes to perform inference and fusion independently and asynchronously based on computing resources, avoiding overall process blockage caused by single attribute inference delays or data waiting. Thus, while ensuring interpretation accuracy, the system throughput is increased to several times that of the traditional synchronous mode, significantly shortening the cycle of nationwide data processing. By introducing a multi-stage streaming attribute fusion mechanism, the strong coupling bottleneck between attribute inference and geometric processing in large-scale remote sensing interpretation is effectively solved, greatly reducing storage requirements.

[0054] This embodiment provides a building remote sensing interpretation method based on streaming processing, which can be used in the aforementioned electronic devices. Figure 2 This is a flowchart of a building remote sensing interpretation method based on streaming processing according to an embodiment of the present invention, such as... Figure 2 As shown, the process includes the following steps: Step S201: Obtain multi-source heterogeneous geographic data within the target area, and based on a preset reference coordinate system, divide the multi-source heterogeneous geographic data into multiple target sub-rectangular areas, and generate multi-band samples for each target sub-rectangular area in parallel.

[0055] Specifically, there are several geographic data sources, each stored in blocks of rectangles of varying sizes. These data sources need to be merged and aligned based on geographic location to quickly form new rectangles of multi-band target size. This ensures that even with streaming generation, there are sufficient samples for AI model inference on GPU cards, enabling rapid inference without waiting.

[0056] For example, in the national building use interpretation task, it is necessary to prepare remote sensing data input and POI (Place of Intetest) distance color temperature map data input. Both of these data inputs are rectangular rasters, but the rasters are different sizes. Therefore, it is necessary to generate multi-band samples. The flowchart for generating multi-band samples is as follows. Figure 4 As shown, step S201 above includes: Step S2011: Align the multi-source heterogeneous geographic data to a preset reference coordinate system, and divide the multi-source heterogeneous geographic data into multiple independent data files according to the preset rectangle size for storage. Each data file is uniquely identified and indexed by its row and column numbers in the preset reference coordinate system.

[0057] Specifically, select a reference coordinate system: coordinate origin, projected coordinate system selection, and resolution (pixel size). The selection of the reference coordinate system should be compatible with the coordinate characteristics of various source data and target output data, such as the Web Mercator coordinate system (3857) and the GPS coordinate system (4326).

[0058] Based on the selected region coordinate system, the files from each data source are aligned to form rectangular regions of appropriate size within that coordinate system. Each file's name indicates its row and column number within the selected coordinate system and rectangular size (facilitating subsequent fast indexing). Specifically, when the number of files is exceptionally large (e.g., hundreds of millions), they are stored in two-level directories, which can be based on either row or column numbers. The row number spacing is determined by the input image size; here, it is assumed that all images are of equal size, with no overlap or omissions.

[0059] Step S2012: Determine the boundaries of the target area. Based on the input image size required by the preset AI model, divide the boundaries into a regular grid, where each regular grid cell is a target sub-rectangle. Calculate the row and column number of each target sub-rectangle in the preset reference coordinate system.

[0060] Specifically, this step generates a list of target sub-rectangles of all target sizes based on the overall target area and reference coordinate system, and according to the input image size required by the AI ​​model, to facilitate parallel sample generation.

[0061] Determine the grid extent: First, obtain the circumscribed rectangle of the target area, i.e., its four boundaries (westernmost, easternmost, northernmost, and southernmost coordinates). This circumscribed rectangle defines the maximum spatial extent covered by the regular grid.

[0062] Calculate the row and column range: Based on the input image size required by the AI ​​model (i.e., the size of the target sub-rectangle), combined with the coordinates and resolution of the data source origin (usually the top left corner of the coordinate system), the calculation is performed using predefined mathematical relationships: Row number = ((source origin y-coordinate - current y-value) / target sub-rectangle height); Column number = rounded ((current x value - x coordinate of data source origin) / width of target sub-rectangle); Calculate the row and column numbers corresponding to the coordinates of the top left and bottom right corners of the rectangle that the target region is bounded by, thereby determining the range of row and column numbers for all potential grid cells covering the region.

[0063] Step S2013: Based on the row and column numbers of each target sub-rectangle, select the target sub-rectangles that intersect with the geographical boundary of the target area from all grid cells to generate the final target sub-rectangle task list.

[0064] Specifically, generate the initial grid: traverse the range of row and column numbers calculated in the previous step, generate the number (e.g., <row number, column number>) of each grid cell (i.e., target sub-rectangle) within that range, forming an initial, complete task grid list.

[0065] Valid task filtering: For each grid cell number, determine its actual geographic rectangle range and its spatial relationship with the precise geographic boundary of the target area (which may be irregular, such as national or provincial borders). Only retain grid cells whose geographic range is entirely within the target area or intersects with the target area's boundary. Grid cells completely outside the target area will be filtered out.

[0066] Step S2014: Parallel traverse the target sub-rectangle task list, retrieve and extract the required multi-source data files based on the row and column numbers of each target sub-rectangle, and generate corresponding multi-band samples through splicing and synthesis.

[0067] In an optional implementation, step S2014 includes: Step a1: Based on the row and column numbers of each target sub-rectangle, locate all relevant data files from each source data, and then stitch all the located data files together to form the source data large rectangle corresponding to each source data.

[0068] For each target sub-rectangle, the system calculates the range of source data files to be accessed for each type of source data (such as remote sensing imagery or POI map) based on its known row and column numbers. Specifically, it calculates the row and column numbers of the source data at the top left and bottom right corners of the target sub-rectangle, and combines the row and column number ranges corresponding to these two row and column numbers to obtain the file number to be retrieved for each data source.

[0069] Based on the calculated list of file numbers, the corresponding individual files are retrieved from each source data set. Within each source data set, these adjacent small files are concatenated in memory into a temporary, larger rectangle of source data with a wider coverage area (by concatenation). This step transforms numerous small file I / O operations into continuous matrix operations, preparing for subsequent precise pruning.

[0070] Step a2: Calculate the coordinate offset of the target sub-rectangle relative to each source data rectangle. Based on the offset and the size of the target sub-rectangle, cut out the corresponding target area data block from each source data rectangle.

[0071] For each large rectangle in the source data, calculate the pixel offset (x, y) of the top left corner of the target sub-rectangle relative to the top left corner of the large rectangle.

[0072] Based on the pixel offset (x, y) and the size of the target sub-rectangle, the target rectangular region can be quickly selected from the large rectangles found in the source data.

[0073] Specifically, after finding the pixel offset, the new pixel is found using the following calculation formula: new_matrix = old_matrix[x:x+L, y:y+L]; Where new_matrix represents the new pixel, old_matrix represents the original pixel, and L represents the pixel offset (i.e., length).

[0074] Based on this offset and the known dimensions (L, W) of the target sub-rectangle, data blocks that perfectly correspond to the target sub-rectangle region are quickly and accurately cropped from each source data rectangle using efficient array slicing operations.

[0075] Step a3 involves stitching and fusing target area data blocks from different source data along the band dimension to generate the final multi-band sample of the corresponding target sub-rectangle.

[0076] Pre-cropped data blocks from different sources (each block may already be multi-band, such as 3 bands in RGB imagery) are treated as different band groups and stitched and aligned along the band dimension. Ultimately, these data blocks are fused into a unified multi-channel two-dimensional matrix with more bands (e.g., 6 or more), which represents the multi-band samples required for the target sub-rectangle and can be directly input into the AI ​​model for inference.

[0077] As can be seen from the above process, the innovative design and combination of sample generation task decoupling and parallelism, fast source data file search algorithm under a unified coordinate system, and fast target area selection algorithm within a multi-source rectangular area have realized a fast sample generation algorithm, which can effectively ensure a sufficient supply of samples in the process of large-scale remote sensing interpretation.

[0078] Step S202: After performing graphical AI model inference and vector conversion operations on each target sub-rectangle multi-band sample, a geometrically complete building outline vector is obtained.

[0079] Specifically, step S202 includes: Step S2021: Input the multi-band sample of each target sub-rectangle into the preset AI model to perform graphics task inference, and output the graphic raster corresponding to each target sub-rectangle.

[0080] Specifically, each multi-band sample of the target sub-rectangle generated in parallel is used as input data and fed in batches into a pre-set AI model (such as a remote sensing large model based on the ViT architecture). Based on the knowledge obtained from its pre-training, the model performs forward inference on the input samples and simultaneously completes multiple pre-set tasks (such as building outline recognition, usage classification, height estimation, etc.).

[0081] The model outputs multi-class prediction results for each target sub-rectangle, represented as multiple two-dimensional arrays, where each pixel position contains the predicted score or regression value for different categories.

[0082] Step S2022: Convert the graphic raster of each target sub-rectangle into a vector graphic, merge the vector graphics located at the boundary of the target sub-rectangle to obtain a geometrically complete building outline vector, and assign a unique identifier number to each building outline vector.

[0083] Specifically, suppose there are several rectangular grids. Within each rectangular grid area, multiple internal vectors can be formed by aggregating pixels with the same value. The same surface object (such as a building roof) may be scattered across vector graphics converted from multiple sub-rectangular areas. How can we quickly and memory-efficiently merge the vector results from multiple rectangles?

[0084] There are already open-source code repositories that provide the function of merging overlapping vector graphics (such as the union_all interface of the Python Shapely library), but when doing large-scale inference, directly putting the vectors of all sub-rectangles (such as 300,000 sub-rectangles, millions of vector graphics) into the interface will generally result in insufficient memory on the machine and slow merging speed.

[0085] This embodiment divides vectors into internal rectangular vectors and rectangular boundary vectors by observing the characteristics of the data. Only on the rectangular boundary vectors will the above-mentioned problem of the same surface object being segmented in multiple vector graphics exist. It is only necessary to judge the vectors on the rectangular boundary and perform boundary intersection fusion.

[0086] The flowchart for generating multi-raster file outline vectors is as follows: Figure 5 As shown, in an optional implementation, step S2022 includes: Step b1: Extract the graphic raster data of each target sub-rectangle, and filter it according to the preset invalid raster values ​​of building targets, deleting graphic rasters that do not contain valid building information.

[0087] Specifically, the graphical raster data (such as building outline raster) of each target sub-rectangle is first extracted. Based on the preset invalid raster values ​​for building targets (for example, a value of 0 represents a non-building area), the raster is quickly scanned and filtered. Graphical rasters with all invalid raster values ​​(i.e., containing no valid building information) are identified, skipped, or deleted directly, thereby avoiding subsequent calculations on meaningless data and significantly improving processing efficiency.

[0088] For example, extract the graphic raster containing the roof, and first filter out the raster image without the roof (i.e., graphic raster with all raster values ​​of 0).

[0089] Step b2: Perform a raster-to-vector conversion operation on all valid graphic rasters to generate an initial vector graphic set; the initial vector graphic set includes internal vectors that are completely inside the sub-rectangle, and boundary vectors that intersect the boundary of the sub-rectangle.

[0090] Specifically, for all valid graphic rasters retained after filtering, a raster-to-vector conversion operation is performed. This operation aggregates adjacent pixels with the same value in the raster to form independent vector polygons, which together constitute the initial vector graphic set.

[0091] Simultaneously with generation, the system initially categorizes each vector polygon into two types based on its positional relationship with the boundary of the current target sub-rectangle: Internal vector: A vector that is completely located inside the sub-rectangle.

[0092] Boundary vector: The vector that intersects any boundary of the sub-rectangle.

[0093] Step b3: Identify and mark all boundary vectors, and perform a vector intersection fusion operation on the boundary vectors of all target sub-rectangles to merge broken vectors that cross target sub-rectangles and belong to the same feature into a complete geometry.

[0094] Specifically, all target sub-rectangles are traversed, and each boundary vector is identified and marked. Then, a vector intersection and fusion operation (such as union_all) is performed specifically on the set of boundary vectors of all sub-rectangles. The core purpose of this operation is to merge the fragmented boundary vectors that are divided into different target sub-rectangles due to area division, but actually belong to the same feature (such as a building), into a single, complete geometry.

[0095] Step b4: Merge all the internal vectors of each target sub-rectangle with the fused boundary vectors to generate a geometrically complete building outline vector within the target sub-rectangle area, and assign a unique identifier to each building outline vector.

[0096] Specifically, the internal vectors completely located within each target sub-rectangle are merged with the seamless boundary vectors formed after cross-rectangle fusion processing to generate a geometrically complete set of building outline vectors within the target sub-rectangle area. Subsequently, each building outline vector in this set is assigned a globally unique identifier, and this identifier is permanently stored as its core attribute, thereby giving each building entity a uniquely identifiable digital identity throughout the target area.

[0097] This step S2022 can also be summarized as the first stage of efficient generation of attributed vectors, such as... Figure 6 The section shown by the dotted line represents the first stage of work, which involves vectorizing and merging the graphic and attribute raster of each target rectangle sub-rectangle. After generating the graphic vector of the target area, a unique identifier is assigned to each vector graphic to facilitate subsequent merging of vector attributes within the sub-rectangle areas; for example... Figure 6The part shown under the dotted line represents Phase Two. Phase Two involves asynchronous streaming fusion of attribute information based on Phase One (geometric vector generation and numbering).

[0098] As can be seen, the above process enables fast and memory-efficient vector result generation, supporting the generation of streaming final vector results.

[0099] Step S203 involves performing attribute AI model inference based on preset target attributes, and then performing multi-stage streaming attribute fusion based on the geometrically complete building outline vector to generate a vector interpretation result with attributes for the target area. For details, please refer to [link to details]. Figure 1 Step S103 of the illustrated embodiment will not be described again here.

[0100] The building remote sensing interpretation method based on streaming processing provided in this embodiment uses a preset model to sequentially perform graphical AI model inference and vectorization operations on standardized samples to obtain geometrically complete building outline vectors. This step does not store intermediate files after graphical AI model inference, which greatly reduces server storage space and saves storage resources.

[0101] This embodiment provides a building remote sensing interpretation method based on streaming processing, which can be used in the aforementioned electronic devices. Figure 3 This is a flowchart of a building remote sensing interpretation method based on streaming processing according to an embodiment of the present invention, such as... Figure 3 As shown, the process includes the following steps: Step S301: Acquire multi-source heterogeneous geographic data within the target area, and based on a preset reference coordinate system, divide the multi-source heterogeneous geographic data into multiple target sub-rectangular areas, generating multi-band samples for each target sub-rectangular area in parallel. For details, please refer to [link to details]. Figure 2 Step S201 of the illustrated embodiment will not be described again here.

[0102] Step S302 involves sequentially performing graphical AI model inference and vectorization operations on each target sub-rectangle multi-band sample to obtain a geometrically complete building outline vector. For details, please refer to [link to relevant documentation]. Figure 2 Step S202 of the illustrated embodiment will not be described again here.

[0103] Step S303: Perform attribute AI model reasoning based on preset target attributes, and perform multi-stage streaming attribute fusion based on geometrically complete building outline vectors to generate vector interpretation results of the target area with attributes.

[0104] Specifically, this step belongs to the stage of efficient generation of attributed vectors, such as... Figure 6 As shown, step S303 above includes: Step S3031: Perform attribute AI model reasoning on the current target attribute to generate a current stage attribute raster covering the target area.

[0105] Specifically, the target attribute interpretation model preset for the current stage (such as an AI model specifically used to identify building uses) is scheduled. Subsequently, all target sub-rectangle multi-band samples that have been generated in step S301 and cover the target area are used as the input data stream for this model.

[0106] Next, distributed model inference is performed: the attribute interpretation model performs parallel forward inference on the input multi-band samples. Based on the features it has learned, the model analyzes each pixel in the sample and outputs the predicted value of the current target attribute corresponding to that pixel (e.g., the probability of each category for a classification task, and a specific estimated value for a regression task).

[0107] Subsequently, stage attribute rasters are generated: the model inference results of all target sub-rectangles are collected and organized and stitched together according to their original geographical locations, ultimately forming one or more current stage attribute rasters covering the entire target area. Each pixel value in each raster file represents the interpretation result of that geographical location for the current target attribute.

[0108] Step S3032: Convert the attribute raster into attribute vectors, with each attribute vector carrying the attribute label for the current stage.

[0109] Specifically, this step belongs to the data preparation stage of attribute fusion. Its core operation is to convert the target attribute information in raster format into spatial objects in vector format. The specific process is as follows: The building attribute raster of each target sub-rectangle (e.g., a raster map where each pixel value represents residential, commercial, or industrial) is read. Using a vectorization algorithm, pixels with the same attribute values ​​and spatially adjacent pixels are aggregated to form continuous, uniform polygon vectors. Each such generated vector polygon carries a specific attribute label to record the attribute category it represents.

[0110] Step S3033: Based on spatial location relationships and unique identifiers, multi-stage streaming attribute fusion is used to fuse the attribute information in the attribute vector into the corresponding building outline vector, generating a vector interpretation result with attributes for the target area.

[0111] In an optional implementation, step S3033 includes: Step c1: Establish a spatial index for the geometrically complete building outline vector.

[0112] Specifically, a spatial index is established for the geometrically complete building outline vector generated in Phase 1, which facilitates the rapid retrieval of vectors that intersect with the target sub-rectangular region.

[0113] The aforementioned spatial index refers to establishing a spatial query structure similar to a "book catalog" for all geometrically complete building outline vector polygons. This allows for the quick selection of a few attribute vectors that may intersect with the building outline vectors based on their positions, without having to traverse all the data.

[0114] Step c2: Traverse all target sub-rectangular regions corresponding to the current attribute grid, and perform spatial matching and attribute surface accumulation of the building outline vector and attribute vector within each target sub-rectangular region based on the building outline vector and the corresponding unique identifier.

[0115] Specifically, step c2 above includes: Step c21: Based on the spatial range of the target sub-rectangle, retrieve all building outline vectors that intersect with the current attribute vector through spatial index, and record the unique identifier number corresponding to each building outline vector.

[0116] Step c22: Perform spatial intersection analysis on each retrieved building outline vector and the attribute vector within the current target sub-rectangle, calculate the intersection area of ​​each intersection part, and group and accumulate the intersection areas according to attribute category and unique identifier number.

[0117] Specifically, iterate through all attribute grid target sub-rectangular regions and perform the following operations: (1) Based on the current target sub-rectangle envelope, retrieve the building outline vector generated in stage one that intersects with the target sub-rectangle (in the case of being within the target sub-rectangle area).

[0118] (2) Copy the retrieved building outline vector (copying is mainly for streaming parallelism, ensuring that multiple programs that retrieve vectors from the previous stage are read-only at the same time, avoiding write conflicts).

[0119] (3) Establish a spatial index for the attribute vector to facilitate quick calculation of the overlap between the retrieved building outline vector and the attribute vector.

[0120] (4) Traverse the copied building outline vector, find all attribute vectors that overlap (intersect) with it based on the spatial index, and perform the following operations: ① The area of ​​each overlapping attribute vector on the building outline vector is obtained based on the intersection area.

[0121] ②Calculate the intersection area corresponding to the same attribute, and obtain the area of ​​each intersecting attribute value on the building outline vector.

[0122] ③ Assign the summed attribute area value to the building outline vector, i.e., step c3 below.

[0123] Step c3: Merge the attribute surface accumulation results of all target sub-rectangular regions. For the building outline vector corresponding to the same identifier number, determine the attribute category with the largest accumulated area as the final attribute of the building outline vector, and write the final attribute into the building outline vector corresponding to the unique identifier number.

[0124] Specifically, for each uniquely identified building outline vector, all intersecting attribute vectors, their intersection areas, and corresponding attribute labels are calculated. All intersection areas are grouped and summed according to their attribute labels. For example, the total area of ​​the commercial attribute is obtained by summing the intersection areas of all attributes labeled "commercial"; similarly, the total area of ​​attributes such as residential and industrial is calculated. These summed total areas are compared, and a decision is made based on the principle of area dominance. That is, the attribute label with the largest summed area is formally determined as the final attribute of the building outline vector.

[0125] Step c4: Perform fusion processing on each preset target attribute in sequence until all target attributes are fused to obtain the target region with attribute vector interpretation result containing all attribute information.

[0126] Specifically, this step involves merging the attribute merging results of all attribute raster target sub-rectangles, including: (1) Distribute the building outline vector graphic data evenly into several subsets based on the unique identifier number; (2) Merge the building outline vectors with duplicate numbers in each subset in parallel, and perform the following operations: ①Statistical subset of the building outline vector corresponding to each number; ②If there is only one building outline vector for that number in the subset, retain that building outline vector; ③ If there are multiple building outline vectors under this number in the subset, the merged attribute area is obtained by accumulating the attribute values, and the geometry of one of the vectors is retained as the geometry of the merged vector. The attribute value of one of the vectors in the previous stage is retained as the existing attribute value of the merged vector.

[0127] (3) Directly merge the merging results of each sub-geometry to form the final attribute fusion result.

[0128] In summary, the duplication of building outline vector numbers during copying means that the same surface vector object is scattered across multiple sub-rectangular regions, while the elimination of duplicate building outline vector numbers means that the attributes of the same surface vector in multiple sub-rectangular regions are merged.

[0129] Meanwhile, step S303 is decoupled and parallelized, and the execution logic is efficient and fast, which can quickly support streaming vector interpretation and attribute fusion after streaming inference.

[0130] The building remote sensing interpretation method based on streaming processing provided in this embodiment adopts a staged streaming processing mechanism, allowing different attribute models to run independently and in parallel. This completely eliminates the constraint of synchronous processing of multiple attributes in traditional methods, greatly improving system resource utilization and overall interpretation efficiency. Simultaneously, through a dual association mechanism of unique identifier numbers and spatial locations, it ensures accurate and reliable matching of massive vector data and multi-source attribute data. Ultimately, while ensuring the accuracy of attribute assignment, it achieves automated, pipelined fusion from geometric figures to multi-dimensional semantic information. This step, by constructing a streaming fusion pipeline of attribute reasoning-vectorization-number association, realizes efficient asynchronous integration of attribute information in large-scale remote sensing interpretation.

[0131] This embodiment also provides a building remote sensing interpretation device based on streaming processing, which is used to implement the above embodiments and preferred embodiments, and will not be repeated as already described. As used below, the term "module" can be a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0132] This embodiment provides a building remote sensing interpretation device based on streaming processing, such as Figure 7 As shown, it includes: The data acquisition and multi-band sample generation module 701 is used to acquire multi-source heterogeneous geographic data within the target area, and based on a preset reference coordinate system, divide the multi-source heterogeneous geographic data into multiple target sub-rectangular areas, and generate multi-band samples for each target sub-rectangular area in parallel.

[0133] The AI ​​inference and vector conversion module 702 is used to sequentially perform graphical AI model inference and vector conversion operations on each target sub-rectangle multi-band sample to obtain a geometrically complete building outline vector.

[0134] The attribute vector generation module 703 is used to perform attribute AI model reasoning based on preset target attributes, and to perform multi-stage streaming attribute fusion based on geometrically complete building outline vectors to generate attribute vector interpretation results for the target area.

[0135] In some alternative implementations, the data acquisition and multi-band sample generation module 701 includes: The data preprocessing unit is used to uniformly align multi-source heterogeneous geographic data to a preset reference coordinate system, and divide the multi-source heterogeneous geographic data into multiple independent data files according to a preset rectangle size for storage. Each data file is uniquely identified and indexed by its row and column numbers in the preset reference coordinate system.

[0136] The range determination and target sub-rectangle division unit is used to determine the four boundaries of the target area. Based on the input image size required by the AI ​​model, the four boundaries are divided into regular grids, where each regular grid unit is a target sub-rectangle, and the row and column numbers of each target sub-rectangle in the preset reference coordinate system are calculated.

[0137] The task list generation unit is used to filter out the target sub-rectangles that intersect with the geographic boundary of the target area from all grid cells based on the row and column numbers of each target sub-rectangle, and generate the final target sub-rectangle task list.

[0138] The multi-band sample generation unit is used to traverse the target sub-rectangle task list in parallel, retrieve and extract the required multi-source data files based on the row and column numbers of each target sub-rectangle, and generate the corresponding multi-band samples through splicing and synthesis.

[0139] In some optional implementations, the multi-band sample generation unit includes: The data source file splicing sub-unit is used to locate all relevant data files from each source data based on the row and column numbers of each target sub-rectangle, and then splice all the located data files to form the source data large rectangle corresponding to each source data.

[0140] The target region data block generation sub-unit is used to calculate the coordinate offset of the target sub-rectangle relative to each source data large rectangle. Based on the offset and the size of the target sub-rectangle, the corresponding target region data block is cut out from each source data large rectangle.

[0141] The multi-band sample generation subunit is used to stitch and fuse target area data blocks from different source data along the band dimension to generate the final multi-band sample of the corresponding target sub-rectangle.

[0142] In some alternative implementations, the AI ​​inference and vector conversion module 702 includes: The AI ​​inference unit is used to input multi-band samples of each target sub-rectangle into a preset AI model to perform graphics task inference and output the graphic raster corresponding to each target sub-rectangle.

[0143] The building outline vector generation unit is used to convert the graphic raster of each target sub-rectangle into a vector graphic, merge the vector graphics located at the boundary of the target sub-rectangle to obtain a geometrically complete building outline vector, and assign a unique identifier number to each building outline vector.

[0144] In some alternative implementations, the building outline vector generation unit includes: The invalid raster filtering unit is used to extract the graphic raster data of each target sub-rectangle and filter it according to the preset invalid raster value of the building target, deleting graphic rasters that do not contain valid building information.

[0145] The raster-to-vector unit is used to perform raster-to-vector conversion operations on all valid graphic rasters, generating an initial vector graphic set; the initial vector graphic set includes internal vectors that are completely inside the sub-rectangle, and boundary vectors that intersect the boundary of the sub-rectangle. The boundary vector fusion unit is used to identify and mark all boundary vectors, and perform vector intersection fusion operation on the boundary vectors of all target sub-rectangles, merging broken vectors that cross target sub-rectangles and belong to the same feature into a complete geometry.

[0146] The merging and numbering assignment unit is used to merge all the internal vectors of each target sub-rectangle with the fused boundary vectors to generate a geometrically complete building outline vector within the target sub-rectangle area, and assign a unique identifier number to each building outline vector.

[0147] In some alternative implementations, the attribute vector generation module 703 includes: The attribute reasoning unit is used to perform attribute AI model reasoning on the current target attribute and generate a current stage attribute raster covering the target area.

[0148] The attribute vectorization unit is used to convert an attribute raster into attribute vectors, each of which carries the attribute label for the current stage.

[0149] The multi-stage streaming attribute fusion unit is used to fuse attribute information from attribute vectors into corresponding building outline vectors based on spatial location relationships and unique identifiers, thereby generating attribute-bearing vector interpretation results for the target area.

[0150] In some alternative implementations, the multi-stage streaming attribute fusion unit includes: Spatial indexing creates sub-units, used to create spatial indexes for geometrically complete building outline vectors.

[0151] The attribute surface accumulation sub-unit is used to traverse all target sub-rectangular regions corresponding to the current attribute raster. Based on the building outline vector and the corresponding unique identifier, it performs spatial matching and attribute surface accumulation of the building outline vector and attribute vector within each target sub-rectangular region.

[0152] The block attribute fusion sub-unit is used to merge the attribute surface accumulation results of all target sub-rectangular regions. For the building outline vector corresponding to the same identifier number, the attribute category with the largest accumulated area is determined as the final attribute of the building outline vector, and the final attribute is written into the building outline vector corresponding to the unique identifier number.

[0153] The global attribute merging sub-unit is used to sequentially perform fusion processing on each preset target attribute until all target attributes are fused, resulting in the target region with attribute vector interpretation results containing all attribute information.

[0154] In some optional implementations, spatial matching and attribute surface accumulation of building outline vectors and attribute vectors within each target sub-rectangular area are performed based on the building outline vector and its corresponding unique identifier, including: Based on the spatial range of the target sub-rectangle, all building outline vectors that intersect with the current attribute vector are retrieved through spatial index, and the unique identifier number corresponding to each building outline vector is recorded. For each retrieved building outline vector, perform spatial intersection analysis with the attribute vector within the current target sub-rectangle, calculate the intersection area of ​​each intersection part, and group and accumulate the intersection areas according to attribute category and unique identifier number.

[0155] The building remote sensing interpretation device based on streaming processing provided in this embodiment of the invention can execute the building remote sensing interpretation method based on streaming processing provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects for executing the method. Further functional descriptions of the above modules and units are the same as those in the corresponding embodiments described above, and will not be repeated here.

[0156] Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention.

[0157] The following is a detailed reference. Figure 8 This diagram illustrates a suitable structural schematic for implementing an electronic device according to embodiments of the present invention. The electronic device may include a processor (e.g., a central processing unit, graphics processor, etc.) 801, which can perform various appropriate actions and processes based on a program stored in read-only memory (ROM) 802 or a program loaded from memory 808 into random access memory (RAM) 803. The RAM 803 also stores various programs and data required for the operation of the electronic device. The processor 801, ROM 802, and RAM 803 are interconnected via a bus 804. An input / output (I / O) interface 805 is also connected to the bus 804.

[0158] Typically, the following devices can be connected to I / O interface 805: input devices 806 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 807 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; memory devices 808 including, for example, magnetic tapes, hard disks, etc.; and communication devices 809. Communication device 809 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 8Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown, and more or fewer devices may be implemented or have instead.

[0159] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 809, or installed from a memory 808, or installed from a ROM 802. When the computer program is executed by the processor 801, it performs the functions defined in the streaming-based building remote sensing interpretation method of the present invention.

[0160] Figure 8 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0161] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the streaming-based building remote sensing interpretation method shown in the above embodiments is implemented.

[0162] A portion of this invention can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to the invention through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0163] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A method for remote sensing interpretation of buildings based on streaming processing, characterized in that, The method includes: Acquire multi-source heterogeneous geographic data within the target area, and based on a preset reference coordinate system, divide the multi-source heterogeneous geographic data into multiple target sub-rectangular areas, and generate multi-band samples for each target sub-rectangular area in parallel. After performing graphical AI model inference and vectorization operations on each target sub-rectangle multi-band sample, a geometrically complete building outline vector is obtained. The system performs attribute AI model reasoning based on preset target attributes, and performs multi-stage streaming attribute fusion based on geometrically complete building outline vectors to generate vector interpretation results with attributes for the target area.

2. The building remote sensing interpretation method based on streaming processing according to claim 1, characterized in that, Based on a preset reference coordinate system, the multi-source heterogeneous geographic data is divided into multiple target sub-rectangular regions, and multi-band samples for each target sub-rectangle are generated in parallel, including: The multi-source heterogeneous geographic data is uniformly aligned to a preset reference coordinate system, and the multi-source heterogeneous geographic data is divided into multiple independent data files according to a preset rectangle size for storage. Each data file is uniquely identified and indexed by its row and column numbers in the preset reference coordinate system. The boundaries of the target area are determined. Based on the input image size required by the AI ​​model, the boundaries are divided into a regular grid, where each regular grid cell is a target sub-rectangle. The row and column numbers of each target sub-rectangle in the preset reference coordinate system are calculated. Based on the row and column numbers of each target sub-rectangle, target sub-rectangles that intersect with the geographical boundary of the target area are selected from all grid cells to generate the final target sub-rectangle task list; The target sub-rectangle task list is traversed in parallel. Based on the row and column number of each target sub-rectangle, the required multi-source data files are retrieved and extracted. Through splicing and synthesis, the corresponding multi-band samples are generated.

3. The building remote sensing interpretation method based on streaming processing according to claim 2, characterized in that, The parallel traversal of the target sub-rectangle task list, based on the row and column numbers of each target sub-rectangle, retrieves and extracts the required multi-source data files, and generates corresponding multi-band samples through splicing and synthesis, including: Based on the row and column numbers of each target sub-rectangle, all relevant data files are located from each source data, and all located data files are spliced ​​together to form the source data large rectangle corresponding to each source data. Calculate the coordinate offset of the target sub-rectangle relative to each source data rectangle, and based on the offset and the size of the target sub-rectangle, cut out the corresponding target area data block from each source data rectangle; The target region data blocks from different source data are spliced ​​and fused along the band dimension to generate the final multi-band sample of the corresponding target sub-rectangle.

4. The building remote sensing interpretation method based on streaming processing according to claim 1, characterized in that, After performing graphical AI model inference and vectorization operations on each target sub-rectangle multi-band sample, a geometrically complete building outline vector is obtained, including: Input the multi-band sample of each target sub-rectangle into the preset AI model to perform graphics task inference, and output the graphic raster corresponding to each target sub-rectangle; The graphic raster of each target sub-rectangle is converted into a vector graphic. The vector graphics located at the boundary of the target sub-rectangle are then merged to obtain a geometrically complete building outline vector. A unique identifier is assigned to each building outline vector.

5. The building remote sensing interpretation method based on streaming processing according to claim 4, characterized in that, The process involves converting the graphic raster of each target sub-rectangle into a vector graphic, fusing the vector graphics located at the boundaries of the target sub-rectangles to obtain a geometrically complete building outline vector, and assigning a unique identifier to each building outline vector, including: Extract the graphic raster data of each target sub-rectangle, and filter it according to the preset invalid raster values ​​of building targets, deleting graphic rasters that do not contain valid building information; Perform a raster-to-vector conversion operation on all valid graphic rasters to generate an initial vector graphic set; the initial vector graphic set includes internal vectors that are completely inside the sub-rectangle, and boundary vectors that intersect the boundary of the sub-rectangle; Identify and label all the boundary vectors, and perform a vector intersection fusion operation on the boundary vectors of all target sub-rectangles to merge the broken vectors that cross the target sub-rectangles and belong to the same feature into a complete geometry; All internal vectors of each target sub-rectangle are merged with the fused boundary vectors to generate a geometrically complete building outline vector within the target sub-rectangle area, and a unique identifier is assigned to each building outline vector.

6. The building remote sensing interpretation method based on streaming processing according to claim 5, characterized in that, Attribute AI model inference is performed based on preset target attributes, and multi-stage streaming attribute fusion is carried out based on geometrically complete building outline vectors to generate vector interpretation results with attributes for the target area, including: Perform attribute AI model reasoning on the current preset target attributes to generate a current stage attribute raster covering the target area; The attribute raster is converted into attribute vectors, and each attribute vector carries the attribute label for the current stage; Based on the spatial location relationship and the unique identifier number, a multi-stage streaming attribute fusion method is used to fuse the attribute information in the attribute vector into the corresponding building outline vector, generating a vector interpretation result with attributes for the target area.

7. The building remote sensing interpretation method based on streaming processing according to claim 6, characterized in that, Based on spatial location relationships and the unique identifier, a multi-stage streaming attribute fusion method is used to fuse the attribute information in the attribute vector into the corresponding building outline vector, generating an attribute vector interpretation result for the target area, including: Establish a spatial index for the geometrically complete building outline vector; Traverse all target sub-rectangular regions corresponding to the current attribute grid, and perform spatial matching and attribute surface accumulation of the building outline vector and attribute vector within each target sub-rectangular region based on the building outline vector and the corresponding unique identifier; Merge the attribute surface accumulation results of all target sub-rectangular regions, and for the building outline vector corresponding to the same identifier number, determine the attribute category with the largest accumulated area as the final attribute of the building outline vector, and write the final attribute into the building outline vector corresponding to the unique identifier number. Each preset target attribute is fused sequentially until all target attributes are fused, resulting in a target region with attribute vector interpretation containing all attribute information.

8. The building remote sensing interpretation method based on streaming processing according to claim 7, characterized in that, Based on the building outline vector and its corresponding unique identifier, spatial matching and attribute surface accumulation are performed between the building outline vector and attribute vector within each target sub-rectangular area, including: Based on the spatial range of the target sub-rectangle, all building outline vectors that intersect with the current attribute vector are retrieved through the spatial index, and the unique identifier number corresponding to each building outline vector is recorded. For each retrieved building outline vector, perform spatial intersection analysis with the attribute vector within the current target sub-rectangle, calculate the intersection area of ​​each intersection part, and group and accumulate the intersection areas according to attribute category and unique identifier number.

9. A building remote sensing interpretation device based on streaming processing, characterized in that, The device includes: The data acquisition and multi-band sample generation module is used to acquire multi-source heterogeneous geographic data within the target area, and based on a preset reference coordinate system, divide the multi-source heterogeneous geographic data into multiple target sub-rectangular areas, and generate multi-band samples for each target sub-rectangular area in parallel. The AI ​​inference and vector conversion module is used to sequentially perform graphical AI model inference and vector conversion operations on each target sub-rectangle multi-band sample to obtain a geometrically complete building outline vector. The attribute vector generation module is used to perform attribute AI model inference based on preset target attributes, and to perform multi-stage streaming attribute fusion based on geometrically complete building outline vectors to generate attribute vector interpretation results for the target area.

10. An electronic device, characterized in that, include: A memory and a processor are communicatively connected, the memory stores computer instructions, and the processor executes the computer instructions to perform the building remote sensing interpretation method based on streaming processing as described in any one of claims 1 to 8.

Images