Method and system for identifying and analyzing electromagnetic and acoustic environmental sensitive targets of power transmission lines based on artificial intelligence and geographic information system

Abstract

The application provides a method and system for power transmission line electromagnetic and acoustic environment sensitive target identification and analysis based on artificial intelligence and geographic information system. The method takes fused multi-source images as input data; and realizes automatic, accurate and high-resolution sensitive target identification and spatial analysis by using an improved SAM3 model, a polygon structure network and parameter efficient fine-tuning, so as to realize a closed loop and avoid manual operation. The system based on the method has rich functions, can be compatible with multi-system and multi-platform operation, and has convenience and generalization ability.

Description

Technical Field

[0001] This invention belongs to the field of power engineering environmental protection and artificial intelligence application technology, specifically involving an integrated method and system for identifying electromagnetic and acoustic environmental sensitive targets along power transmission lines, conducting on-site collaborative investigations, and automatically generating charts based on visual artificial intelligence models and geographic information systems (GIS). Background Technology

[0002] The power transmission and transformation project has a wide distribution of electromagnetic and acoustic sensitive targets such as residential buildings, schools, and hospitals along its route. According to the requirements of the environmental impact assessment technical guidelines, sensitive targets along the line need to be investigated, identified, and assessed during the construction and operation phases of the power transmission line in order to determine their spatial relationship with the power transmission line, their electromagnetic and acoustic environmental impact, and to provide a basis for the formulation of subsequent environmental protection measures. In actual projects, the environmental impact assessment site investigation has the following characteristics: (1) The investigation scope is wide, involving multiple types of buildings, such as buildings with functions such as residence, factory, work, and study; (2) The investigation scenarios are complex, including terrain with elevation differences, dense residential areas, and irregular building groups; (3) The resulting maps and reports are required to be unified, standardized, and traceable, and must have high spatial accuracy and consistency.

[0003] However, traditional environmental impact assessment (EIA) sites have many shortcomings. For example, relying on manual reconnaissance and manual compilation of sensitive point maps and reports suffers from pain points such as long cycles, insufficient accuracy and consistency, and time-consuming internal compilation. Although there have been attempts in the industry to use convolutional networks such as U-Net and FCN for building segmentation, their recognition accuracy and generalization ability are insufficient in complex terrain, low-contrast imagery, and multi-source data fusion scenarios. The recognition results are disconnected from report compilation, making it difficult to meet the automated application requirements of the entire "identification-analysis-reporting" process in EIA investigations. There are also problems such as insufficient radiometric consistency and fusion of multi-source imagery (satellite / drone), affecting model transfer and stability; pixel masks often exhibit blurred outlines and jagged edges at complex boundaries, making them difficult to use directly as compliant polygonal elements in GIS. In addition, existing systems mostly remain at the image recognition stage, failing to form a closed loop with GIS spatial analysis and report generation.

[0004] In summary, there is an urgent need in this field for a method that integrates multi-source images to achieve high-precision and high-recall segmentation in situations where small targets are densely packed and have large morphological differences. This would enable closed-loop automation of the identification, analysis, and reporting of electromagnetic and acoustic sensitive targets in power transmission lines, thereby improving the interpretability, traceability, and engineering usability of the results. Summary of the Invention

[0005] The purpose of this invention is to provide an integrated method and system for identifying electromagnetic and acoustic environmental sensitive targets along power transmission lines, conducting on-site collaborative investigations, and automatically generating charts based on visual artificial intelligence models and geographic information systems (GIS).

[0006] In a first aspect, the present invention provides a method for constructing a polygonal structure recognition model, the method comprising the steps of:

[0007] (1) Provides a variety of image data;

[0008] (2) Preprocess the various image data to obtain fused image data;

[0009] (3) Processing the fused image data using artificial intelligence, and fine-tuning the processing using Parameter Efficient Fine-Tuning (PEFT); including the following steps:

[0010] (3.1) The image segmentation model encoder with fixed input parameters of the fused image data is fine-tuned using low-rank adaptation (LoRA) in the parameter efficient fine-tuning to generate multi-scale image feature embeddings; the multi-scale image feature embeddings are decoded using the image segmentation model decoder to generate an initial mask;

[0011] (3.2) Convert the initial mask into an initial polygonal outline; convert the initial polygonal outline into a vertex sequence representation by normalization;

[0012] (3.3) The contour optimization model is trained using the vertex sequence representation; the contour optimization model is fine-tuned using the parameter effectiveness learning in the parameter efficient fine-tuning, and the training is constrained using the loss function;

[0013] (4) When the training results reach the predetermined termination condition, the model training is terminated and the polygon structure recognition model is obtained.

[0014] In another preferred embodiment, the various image data are remote sensing image data.

[0015] In another preferred embodiment, the remote sensing image data includes high-resolution satellite imagery, UAV orthophotos, oblique photogrammetry, and aerial photogrammetry.

[0016] In another preferred embodiment, the remote sensing image data is high-resolution satellite imagery and UAV orthophotos.

[0017] In another preferred embodiment, the various image data include building segmentation masks.

[0018] In another preferred embodiment, the number of building segmentation masks is ≥10,000.

[0019] In another preferred embodiment, the preprocessing includes the steps of:

[0020] (a1) The various image data are processed including geometric correction, spectral normalization, radiometric consistency, size normalization, pixel value standardization and color space adjustment to construct the input dataset;

[0021] (a2) Slice the input data in the input dataset to form pyramid tiles; incorporate the pyramid tiles into the tile cache; and

[0022] (a3) Vectorize and annotate the building segmentation mask in the input dataset to obtain fused image data.

[0023] In another preferred embodiment, the size normalization refers to adjusting the image data to a predetermined resolution so that it can be input into the downstream encoder.

[0024] In another preferred embodiment, the pixel normalization refers to processing the image data as follows:

[0025] (x1) Scale the RGB channel values ​​of the image data to [0,1]; or

[0026] (x2) Scale the RGB channel values ​​of the image data to the standard normal distribution range.

[0027] In another preferred embodiment, step (a1) further includes data augmentation of the various image data to increase the amount of data in the various image data.

[0028] In another preferred embodiment, the data augmentation includes multi-temporal brightness recalibration, random scale / rotation / color perturbation, and pseudo-lighting shadow simulation.

[0029] In another preferred embodiment, the image segmentation model includes the SAM3 model, SAM model, HRNet, U-Net, and DeepLabV3+.

[0030] In another preferred embodiment, the image segmentation model is the SAM3 model.

[0031] In another preferred embodiment, the SAM3 model is an improved SAM3 model.

[0032] In another preferred embodiment, the contour optimization model includes a polygon optimization network (PRN), spline curves, Bézier curves, and a raster contour method based on Marching Squares.

[0033] In another preferred embodiment, the contour optimization model is a polygon optimization network (PRN).

[0034] In another preferred embodiment, step (3.1) specifically includes the following steps: using the low-rank adaptation (LoRA) in parameter efficient fine-tuning of the image segmentation model encoder with fixed input parameters of the polygonal contour to perform low-rank decomposition of the attention and / or convolution weights of the image segmentation model, train the low-rank components, thereby generating multi-scale image feature embeddings; and using the image segmentation model decoder to decode the multi-scale image feature embeddings, thereby generating an initial mask.

[0035] In another preferred embodiment, step (3.1) specifically includes the following steps: using the low-rank adaptation (LoRA) in the parameter efficient fine-tuning of the improved SAM3 model encoder with fixed input parameters of the polygonal contour to perform low-rank decomposition of the attention weights of the improved SAM3 model, thereby generating multi-scale image feature embeddings; and using the improved SAM3 model mask generation head to decode the multi-scale image feature embeddings, thereby generating an initial mask.

[0036] In another preferred embodiment, step (3.2) specifically includes the steps of: converting the initial mask into an initial polygonal contour using an edge detection and contour fitting algorithm; and converting the initial polygonal contour into a vertex sequence representation using geometric normalization.

[0037] In another preferred embodiment, step (3.3) specifically includes the following steps: training the Polygon Optimization Network (PRN) using the vertex sequence representation; fine-tuning the Transformer-based sequence regression structure in the PRN using Adapter-Tuning in the parameter efficient fine-tuning to form a fine-tuned Transformer-based sequence regression structure; the fine-tuned Transformer-based sequence regression structure finely adjusts the position of each vertex in the vertex sequence; and constraining the training using a loss function.

[0038] In another preferred embodiment, the loss function is calculated as L = L_vertex + L_smooth + L_IoU.

[0039] In another preferred embodiment, the fine-tuned Transformer-based sequence regression structure contains a lightweight adapter layer.

[0040] In another preferred embodiment, the fine-tuned Transformer-based sequence regression structure is used to improve cross-domain generalization.

[0041] In another preferred embodiment, parameter validity learning in the parameter efficient fine-tuning is used to fine-tune a small number of bias or cue encoder parameters, reducing memory and overfitting risks.

[0042] In another preferred embodiment, the parameter validity learning includes BitFit and Prompt-Tuning.

[0043] In another preferred embodiment, the method further includes the step of: (5) verifying the contour optimization model obtained in step (4).

[0044] In another preferred embodiment, the verification will use one or more of the following as evaluation metrics:

[0045] (b1) Recall and / or precision per building;

[0046] (b2) Boundary intersection-union ratio (bIoU);

[0047] (b3) 95% Hausdorff distance (95% HD);

[0048] (b4) Contour-F1;

[0049] (b5) Vector topology validity rate.

[0050] In another preferred embodiment, the recall of the contour optimization model obtained in step (4) is ≥0.9.

[0051] In another preferred embodiment, the accuracy of the contour optimization model obtained in step (4) is ≥0.8.

[0052] In another preferred embodiment, the vector topology validity of the contour optimization model obtained in step (4) is >95%, such as >96%, >97%, >98%, >99%, preferably >99.9%.

[0053] A second aspect of the present invention provides a method for generating vectorized polygon data, the method comprising the steps of:

[0054] (S1) Provides various image data;

[0055] (S2) Preprocessing the various image data includes the following steps:

[0056] (S2.1) The various image data are processed including geometric correction, spectral normalization, radiometric consistency, size normalization, pixel value standardization and color space adjustment to construct the input dataset;

[0057] (S2.2) Slice the input data in the input dataset to form pyramid tiles; incorporate the pyramid tiles into the tile cache; and

[0058] (S2.3) Vectorize and annotate the input dataset to obtain fused image data;

[0059] (S3) The fused image data is processed using a polygon structure recognition model to generate vectorized polygon data; the polygon structure recognition model is constructed using the method described in the first aspect of the present invention.

[0060] In another preferred embodiment, the various image data are remote sensing image data.

[0061] In another preferred embodiment, the remote sensing image data includes high-resolution satellite imagery, UAV orthophotos, oblique photogrammetry, and aerial photogrammetry.

[0062] In another preferred embodiment, the output format of the vectorized polygon data is selected from the group consisting of WKT format, WKB format, or GeoJSON format.

[0063] A third aspect of the present invention provides a method for spatial analysis of sensitive points of transmission lines based on GIS, the method comprising the steps of:

[0064] (I) Provide data related to power transmission lines, various image data containing building images, field survey data, and administrative division and building point of interest (POI) data; the data related to power transmission lines includes centerline data and conductor crossarm parameters of the power transmission lines; the field survey data provides sensitive point information;

[0065] (II) Based on the centerline data and crossarm parameters of the transmission line, obtain the spatial geometric position data of the side conductors of the transmission line; process the various image data using the method described in the second aspect of the present invention to obtain building vector data;

[0066] (III) Associate and merge the vector data of one or more buildings belonging to the same sensitive point range to form a sensitive point object;

[0067] (IV) Analyze the sensitive point objects, including:

[0068] (ivi) Match the sensitive point object with the administrative division and building POI data to obtain the attributes of the sensitive point object;

[0069] (ivii) In each of the sensitive point objects, based on the spatial geometric position data of the edge guide, analyze the vector data of each building to obtain the building closest to the edge guide, the shortest distance, and the orientation information;

[0070] Steps (ivi) and (ivii) can be interchanged, performed sequentially, or performed simultaneously; thus enabling spatial analysis of sensitive points on transmission lines.

[0071] In another preferred embodiment, the building vector data includes: building polygonal features and building planar features.

[0072] In another preferred embodiment, the attributes of the sensitive point object include: the administrative region information described by the sensitive point object, the sensitive point name, and the sensitive point function.

[0073] A fourth aspect of the present invention provides a system for target identification and analysis based on artificial intelligence and GIS, the system comprising:

[0074] (U1) Input unit, configured to input data, the data including various image data containing building images, transmission line related data, field survey data, and administrative division and building POI data; the transmission line related data includes transmission line centerline data and conductor crossarm parameters; the field survey data provides sensitive point information;

[0075] (U2) Identification unit, the identification unit being configured to perform the following operations:

[0076] (u2.1) Preprocess the various image data to obtain fused image data;

[0077] (u2.2) The fused image data is processed using a polygon structure recognition model to generate building vector data; the polygon structure recognition model is constructed using the method described in the first aspect of the present invention;

[0078] (U3) Analysis unit, the analysis unit being configured to perform the following operations:

[0079] (u3.1) Based on the centerline data of the line and the crossarm parameters of the conductor, obtain the spatial geometric position data of the side conductor of the transmission line;

[0080] (u3.2) Associate and merge one or more building vector data belonging to the same sensitive point range to form a sensitive point object;

[0081] (u3.3) Analyzing the sensitive point objects includes: matching the sensitive point objects with the administrative division and building POI data to obtain the attributes of the sensitive point objects; in each sensitive point object, analyzing the vector data of each building based on the spatial geometric position data of the edge guide to obtain the building closest to the edge guide, the shortest distance, and the orientation information; thereby obtaining the analysis results;

[0082] (U4) Output unit, which is configured to output the analysis results of the analysis unit.

[0083] In another preferred embodiment, the system is built on a reasoning framework.

[0084] In another preferred embodiment, the inference framework is selected from: ONNX Runtime or TensorRT / TVM.

[0085] In another preferred embodiment, the various image data include: UAV orthophotos, high-resolution satellite images, oblique photography images, and aerial photography images, preferably UAV orthophotos.

[0086] In another preferred embodiment, the centerline data of the transmission line is in KML format.

[0087] In another preferred embodiment, the centerline data of the transmission line can be re-entered or repeatedly.

[0088] In another preferred embodiment, the data may also include one or more parameters selected from the following: task number for identification, analysis and evaluation, scope of identification, analysis and evaluation, line type (e.g., overhead line or cable), voltage level, crossarm length (e.g., crossarm length of overhead line), and gallery radius (e.g., gallery radius of overhead line).

[0089] In another preferred embodiment, the analysis results include a sensitivity survey and / or a sensitivity atlas.

[0090] In another preferred embodiment, the sensitive point survey form and sensitive point atlas are used in the environmental impact assessment report.

[0091] In another preferred embodiment, the sensitive point survey form and sensitive point atlas are in Word format.

[0092] In another preferred embodiment, the sensitive point survey form includes: information on the administrative region of the sensitive point, the name of the sensitive point, the function of the sensitive point, the number of households within the evaluation range, the structure of buildings within the evaluation range, the structure of the nearest building to the transmission line conductor, directional information, environmental impact factors, noise environmental protection requirements, platforms within the evaluation range, and erection methods.

[0093] In another preferred embodiment, the sensitive point atlas includes: the relative position of the sensitive point to the transmission line, the distance between the sensitive point and the transmission line, the name of the sensitive point, building information, and on-site photos.

[0094] In another preferred embodiment, the analysis results also include spatial vector data of the side conductors, demolition range, and evaluation range of the selected line type.

[0095] It should be understood that, within the scope of this invention, the above-described technical features of this invention and the technical features specifically described below (such as in the embodiments) can be combined with each other to form new or preferred technical solutions. Due to space limitations, they will not be described in detail here. Attached Figure Description

[0096] Figure 1 The diagram shows the overall process flow of the algorithm data processing and AI recognition.

[0097] Figure 2 The diagram shows a flowchart of polygon structure modeling and efficient parameter fine-tuning.

[0098] Figure 3 The diagram shows the effect of the algorithm in recognizing vectorized polygons.

[0099] Figure 4 The diagram shows a schematic of the sensitive points created during the on-site investigation.

[0100] Figure 5 A schematic diagram of sensitive points in the electromagnetic and acoustic environment is shown.

[0101] Figure 6 A schematic diagram showing a partial survey form and a partial atlas of sensitive points is displayed. Detailed Implementation

[0102] Through extensive and in-depth research, the inventors have developed for the first time an integrated method and system for identifying electromagnetic and acoustic sensitive targets along power transmission lines, conducting on-site collaborative investigations, and automatically generating charts, using multi-source imagery. This system achieves high-precision, high-recall segmentation of small, densely packed targets with significant morphological differences. Specifically, the method standardizes and corrects images from different sources to maintain the structural stability of the input data. By improving the SAM model and introducing a polygon structure network for polygon optimization and efficient parameter fine-tuning, automatic, accurate, and high-resolution image identification is achieved. The system based on this method supports functions including creating analysis schemes, automatically identifying and calculating shortest distances, and is compatible with multiple systems and platforms, offering rich functionality and ease of use. This invention is based on this foundation.

[0103] It should be understood that the specific methods and experimental conditions of the invention described below in varying degrees of detail are intended to provide a substantive understanding of the invention. Definitions of certain terms used in this specification are provided below. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0104] the term

[0105] As used herein, the terms “containing” or “including (comprise)” can be open-ended, semi-closed, or closed-ended. In other words, the terms also include “consistently made of” or “made of”.

[0106] As used herein, the term “and / or” refers to and covers any and all possible combinations of one or more of the related listed items.

[0107] As used in this article, the term "significant" means that, in a hypothesis test, the observed effect (such as the difference between the experimental and control groups) is unlikely to be caused solely by random error. A hypothesis test includes: the null hypothesis (H0), which assumes that the observed effect does not exist (such as no difference between the experimental and control groups); the p-value, which is the probability of observing the current or more extreme effect when H0 is true; and the significance threshold (α). The significance threshold is typically used to determine whether a hypothesis test is significant. Generally, the significance threshold is 0.05. If the p-value ≤ α, then H0 is rejected, meaning the observed effect exists, and the result is called "significant."

[0108] As used in this article, the terms "multiple image data" and "multi-source image data" refer to image data from multiple imaging devices or multiple time periods, typically including satellite imagery, UAV imagery, oblique photogrammetry, aerial photogrammetry, etc. This terminology emphasizes the diversity of image types and the differences in resolution and spectral characteristics.

[0109] As used in this article, the term "pyramid tile" refers to the structure of a pyramid model. A pyramid model is a multi-resolution hierarchical model that, under a unified spatial reference, stores and displays data at different resolutions according to user needs, forming a pyramid structure—the "tiles"—with resolutions ranging from coarse to fine and data volume from small to large, while maintaining the same geographical scope. The deeper layers of the pyramid represent more detailed map information and have larger scales.

[0110] As used in this paper, the term "parameter efficiency learning" is a method to maximize training time and space efficiency while ensuring that model performance is not significantly affected. Specifically, in machine learning or deep learning models, it involves training or efficiently updating only a small subset of key parameters while keeping most of the original parameters frozen. Parameter efficiency learning includes Adapter-Tuning, Prefix-Tuning, BitFit, and Prompt-Tuning. "Adapter-Tuning" refers to adding an Adapter module between each (or some) of the attention / convolutional network layers in the frozen model body (i.e., the pre-trained model). Each Adapter module consists of two feedforward sub-layers. By limiting the size of the input dimension, the number of parameters in the Adapter module is controlled, significantly reducing the number of trainable parameters, saving training time, and still allowing the model to achieve the effect of training with all parameters.

[0111] As used in this paper, the term "improved SAM3 model" refers to a SAM3 model whose backbone is retained for feature extraction, and whose output mask generator is transformed by a lightweight structure transformation module to convert the input image into a polygonal outline.

[0112] As used in this paper, the terms “boundary intersection-union ratio”, “bIoU” and “boundary bIoU” are used interchangeably. It is a term used to evaluate the degree of overlap of bounding boxes in object detection. It is calculated by maximizing the overlap between the predicted box and the ground truth box.

[0113] As used in this paper, the term "95% Hausdorff distance (95%HD)" is used to describe the maximum error statistic between the predicted profile and the true profile, and the value of the 95th percentile reduces the sensitivity to extreme points.

[0114] As used in this article, the term "Contour-F1" is a combined precision and recall metric based on contours, used to measure the degree of matching of predicted polygon contours at the boundary level, and is an important supplementary metric for building boundary assessment.

[0115] As used in this article, the term "vector topology validity rate" refers to the percentage of output polygons that are valid under GIS topology rules, including no self-intersections, no overlaps, no gaps, polygon closure, and correct orientation, which are necessary conditions for polygons to be used as GIS elements.

[0116] As used in this article, the terms "point of interest" and "POI" are used interchangeably. A POI is a data unit used in a geographic information system to mark a specific location. POIs typically include shops, hospitals, schools, gas stations, etc., and contain information such as the name, category, coordinates, and address of these locations.

[0117] As used herein, the term "sensitive point" refers to an object used for spatial analysis in this invention. A sensitive point can be a point of interest; a sensitive point object may include one or more points of interest.

[0118] As used herein, the term "computer system" refers to a system comprising at least one processor and readable storage, which executes the program instructions of the present invention to implement the method steps. Although the flowchart shows a fixed order, in fact the order can be flexibly adjusted, parallelized, or steps can be omitted depending on the implementation, as long as the technical solution of the claims is satisfied.

[0119] The computer system is equipped with at least one processor and a memory. The processor invokes a sequence of computer-executable instructions stored in the memory to implement the evaluation process defined in the claims. Although the flowchart describes the operation steps in a specific logical order, in actual execution, the steps may be processed in parallel, their order adjusted, or partially omitted in some cases. As long as such adjustments do not deviate from the core features of the technical solution described in the claims and achieve the same technical effect, they all fall within the scope of protection of this invention. This flexibility in execution order is determined by the programmable nature of computer instructions.

[0120] The main advantages of this invention include:

[0121] (1) The polygonal structure modeling in this invention significantly improves the edge quality of complex boundaries and fine-grained targets, while maintaining engineering-grade throughput, and possesses both contour accuracy and efficiency.

[0122] (2) The method and system of the present invention realize the closed loop of the whole chain of “identification-vectorization-analysis-mapping”. The output vector results can be directly put into the database and drive spatial analysis and automatic map generation, reducing manual editing breakpoints and improving consistency.

[0123] (3) The method and system of the present invention only update a small number of parameters, reduce migration costs and memory requirements, support rapid cross-regional adaptation, and have good generalization ability and convenient size.

[0124] (4) The system of the present invention allows ≥1000 concurrent users, load balancing, fault tolerance and automated deployment, and can be used in domestic OS / database environment, meeting the requirements for group-level implementation.

[0125] The present invention will be further illustrated below with reference to specific embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. It is also understood that the purpose of describing the present invention in conjunction with the embodiments is to cover other options or modifications that may be derived based on the claims of the present invention. To provide a thorough understanding of the invention, numerous specific details will be included in the following description.

[0126] This study achieves a closed-loop quality control process throughout the entire process, including test planning, test cases, independent environment setup, issue tracking, regression analysis, and test summary; acceptance testing does not allow for any issues of grade B or higher.

[0127] For risk management, this study develops strategies such as prototype review / POC first, enhanced data collection, and pre-locking of computing resources to address changes in requirements, algorithm performance, and fluctuations in GPU resources.

[0128] Example 1: Data Processing Layer

[0129] This study employed a multi-source image consistency fusion method for data processing; the workflow is described below. Figure 1 Specifically, this includes:

[0130] (1) Geometric / radiative correction and spectral normalization: Geometric correction, spectral normalization and radiometric consistency processing are performed on high-resolution satellite and UAV orthophotos to construct a unified input dataset; uniform preprocessing operations are performed on the input images to ensure that the model has a consistent feature distribution under different data sources.

[0131] Before entering the model, the input image undergoes steps such as size normalization, pixel value standardization, and color space adjustment. Size normalization typically adjusts the image to a fixed resolution to match the input requirements of the SAM encoder; pixel normalization scales the RGB channel values ​​to [0,1] or the standard normal distribution range. These preprocessing operations ensure that the input data maintains structural stability before entering the backbone network, providing a unified semantic basis for subsequent feature extraction.

[0132] (2) Slicing and caching: Pyramid tiles are generated during the preprocessing stage and incorporated into the tile cache to improve the I / O efficiency of visualization and inference on the Web side.

[0133] (3) Data labeling assets and expansion: Accumulate ≥10,000 house segmentation masks on satellite and aerial images, use vectorized labeling and maintain metadata dictionary to form vectorized training labels, thereby ensuring stable iteration of training and evaluation.

[0134] Example 2: AI Recognition Layer

[0135] The AI ​​recognition layer mainly consists of an improved SAM3, polygonal structure modeling, and efficient parameter fine-tuning. Specifically, it includes:

[0136] (1) Improved SAM3: In the whole system, the backbone of the SAM3 model is retained for feature extraction, and its output mask is converted into polygonal contours through a lightweight structure transformation module.

[0137] (2) Polygon structure modeling: A specially designed polygon optimization network corrects the initial contour described above, making the position of each vertex more closely match the true boundary of the target. This method ensures that the optimized polygon is both accurate and continuous by introducing geometric constraints and a boundary smoothing loss function. Figure 2 ).

[0138] 2.1 Initial Mask Extraction: In the first stage, the input image is fed into the frozen SAM encoder to generate multi-scale image feature embeddings. Subsequently, the initial mask image is generated using the SAM mask generator.

[0139] 2.2 Adaptive Vertex Sampling: In the second stage, the mask is converted into an initial polygonal contour through edge detection and contour fitting algorithms, and a series of vertex sequence representations are generated through geometric normalization operations.

[0140] 2.3 Vertex Regression and Refinement: In the third stage, a polygon optimization network (PRN) was introduced. Taking the vertex sequence as input, it uses a Transformer-based sequence regression structure to finely adjust the position of each vertex. The optimization results are trained by combining geometric error loss (L_vertex), smoothness constraint loss (L_smooth), and IoU matching loss (L_IoU). The resulting polygons achieve high levels of vertex continuity, boundary smoothness, and region matching.

[0141] 2.4 Differentiable Loss Combination: IoU is used in the pixel domain, and vertex geometric consistency loss and boundary smoothing loss are introduced in the geometric domain, so that the model can balance local edge accuracy and overall contour consistency during optimization. The entire training process adopts a multi-task joint optimization strategy to dynamically balance the different loss terms during training;

[0142] 2.5 GIS-friendly output: Output directly in WKT / WKB or GeoJSON, maintaining coordinate reference for easy data import (PostGIS / GeoPackage) and subsequent spatial analysis.

[0143] (3) Parameter-Efficient Fine-Tuning (PEFT): Furthermore, to reduce training costs and improve model transferability in new tasks, this study employs a parameter-efficient fine-tuning mechanism based on low-rank adaptation (LoRA), adjusting only a small number of key layer parameters while freezing the remaining weights of the backbone network. This design significantly reduces the number of training parameters, achieving improved computational efficiency while maintaining segmentation performance. Figure 2 ).

[0144] 3.1 LoRA: Perform low-rank decomposition on the attention / convolution weights in SAM / HRNet, and train only the low-rank components;

[0145] 3.2 Adapter: Insert a lightweight adapter layer after the Transformer / high-branch features to improve cross-domain generalization;

[0146] 3.3 BitFit / Prompt-Tuning: Fine-tunes a small number of bias or prompt encoder parameters to reduce the risk of memory and overfitting;

[0147] 3.4 Freezing Strategy: Freeze most of the parameters of the backbone and train only in the polygon head and the adaptation layer to balance efficiency and stability.

[0148] (4) Training and Validation: The AI ​​recognition layer was trained and validated on an A800×4 / FP16 or equivalent data center GPU. Data augmentation was performed using methods such as multi-temporal brightness recalibration, random scale / rotation / color perturbation, and pseudo-lighting shadow simulation. Recall / precision per unit of buildings, boundary bIoU, 95% Hausdorff distance (95%HD), and Contour-F1 were used as evaluation indicators, with a focus on boundary quality.

[0149] The training metrics are considered met when the recall rate per building is ≥0.95, the precision is ≥0.90, and the vector topology validity rate is close to 100%. A schematic diagram of the final output vectorized polygon is shown below. Figure 3 As shown.

[0150] (5) Reasoning and service orientation:

[0151] 5.1 Inference Framework: ONNX Runtime deploys FastAPI / Flask services, asynchronous task queues, and flow control;

[0152] 5.2 Fusion Strategy: Pyramid resolution and sliding window stitching; overlapping polygons are merged based on vector IoU / NMS.

[0153] 5.3 Quality Control: Automatically correct or filter abnormally small / abnormally slender / self-intersecting polygons, while retaining confidence level and geometric quality score;

[0154] 5.4 Hardware Recommendations: NVIDIA 4090 or higher or equivalent data center GPU to meet engineering throughput and latency requirements.

[0155] (6) Structured output and clustering oriented towards sensitive points:

[0156] 6.1 Constructing line geometry: Obtain the line centerline data and corresponding conductor crossarm parameters of the transmission line. Based on the line centerline and the conductor crossarm parameters, generate the spatial geometric position data of the side conductors of the transmission line.

[0157] 6.2 Obtaining Building Vectors: Based on the algorithm, the image data is identified to obtain building vector data, which are polygonal or planar features of individual buildings; avoiding the accumulation of errors in post-processing from raster to vector.

[0158] 6.3 Sensitive Point Creation and Association: Based on the APP's on-site survey results, multiple building vectors belonging to the same sensitive point range are associated and merged to form sensitive point objects; each sensitive point object is associated with at least one or more building vector elements. A schematic diagram of sensitive point creation based on the on-site survey is shown below. Figure 4 As shown.

[0159] 6.4 Sensitive Point Administrative Region and Name Association: Spatially overlay or spatially match the sensitive point objects with administrative division and building POI data to automatically determine and associate the administrative region information, sensitive point name, and sensitive point function of the sensitive point object, and support manual modification.

[0160] 6.5 Nearest Building Distance and Orientation Analysis: For each sensitive point object, calculate the shortest distance between its associated buildings and the spatial geometric position data of the edge guide, determine the target building closest to the edge guide, and automatically output the shortest distance and the orientation information of the target building relative to the edge guide.

[0161] 6.6 Result Labeling: The administrative region information, shortest distance to the nearest building, location information, and corresponding building of the sensitive point object are labeled in the GIS analysis results to realize the spatial analysis and display of the sensitive points of the power transmission line.

[0162] Example 3: Business Application Layer (GIS Spatial Analysis and One-Click Reporting)

[0163] (1) Intelligent analysis of sensitive points: Create analysis schemes based on administrative divisions / evaluation scope / whether the line is parallel, etc.; automatically identify the building closest to the line among sensitive points and calculate the shortest distance; visualize the list and details of sensitive points / buildings, and support export and saving. Figure 5 ).

[0164] Users can re-import the KML file of the line centerline into the system and create evaluation and analysis tasks based on the composition of the transmission line project. This includes generating task numbers, selecting the line as an overhead line or cable, selecting the voltage level, filling in the length of the overhead line crossarm or the radius of the pipe gallery, and setting the line relocation scope and evaluation or investigation scope. The system automatically generates the corresponding spatial vector data of the line edge conductor, relocation scope and evaluation scope based on the input parameters.

[0165] (2) Automatic chart generation: Generate a list of sensitive points, distribution map, distance analysis map, and on-site survey image layout (automatically avoid legends and key targets).

[0166] After users create sensitive points, the system can automatically generate sensitive point charts and supports batch generation; the system can automatically generate Word versions of sensitive point survey forms and sensitive point atlases according to task number order and the selected analysis task and site survey number, ensuring consistent chart numbering, which can be directly used in environmental impact assessment reports. Figure 6 The sensitive point survey form includes information such as administrative region, sensitive point name, function, number of households within the evaluation scope, building structure within the evaluation scope, structure of the nearest building, location relationship with the project, environmental impact factor, noise environmental protection requirements, platform within the evaluation scope, and erection method; the sensitive point map centrally expresses the relative location relationship and distance between the sensitive point and the line, sensitive point name, building information, and on-site photos.

[0167] Example 5: Mobile Collaboration Layer and System and Data Management

[0168] Based on the above embodiments, a corresponding application product (APP) has been built, which can realize field verification and offline data transmission. The APP is compatible with Android / iOS and mobile phones / tablets, supports engineering project visualization, point / box marking, photo upload, offline caching and offline data transmission; sensitive points, buildings, monitoring points and marked points can be added / verified directly on site; and lightweight regulatory query is provided, with data synchronized with the PC in real time.

[0169] In addition, the app provides general capabilities such as organization / role / user / log / data dictionary / layer management. The system can be deployed on Linux / domestic OS (Kylin) and domestic databases (DM / Kingbase), meeting the requirements of the domestic IT innovation environment and possessing security features. The system supports ≥1000 concurrent users, load balancing, fault tolerance, automated deployment, and monitoring and early warning.

[0170] Example 6: System Performance Analysis

[0171] Interaction performance: 99% of interfaces respond ≤1s, the rest ≤1.5s; complex statistics ≤4s;

[0172] Algorithm metrics: Recall ≥ 0.95 and precision ≥ 0.90 per building; bIoU / 95%HD / Contour-F1 are used as supplementary evaluation criteria for boundary quality.

[0173] Finished charts: The accuracy of the sensitive point reports automatically generated by the system in typical projects is >95%, verified by manual sampling.

[0174] All documents mentioned in this invention are incorporated herein by reference as if each document were individually incorporated by reference. Furthermore, it should be understood that after reading the foregoing teachings of this invention, those skilled in the art can make various alterations or modifications to this invention, and these equivalent forms also fall within the scope defined by the appended claims.

Claims

1. A method of constructing a polygonal structure recognition model, characterized by, The method includes the following steps: (1) Provides a variety of image data; (2) Preprocess the various image data to obtain fused image data; (3) The fused image data is processed using artificial intelligence, and the processing is fine-tuned using parameter fine-tuning. include step: (3.1) The image segmentation model encoder with fixed input parameters of the fused image data is used to fine-tune the image segmentation model by using low-rank adaptation in the parameter fine-tuning, thereby generating multi-scale image feature embeddings; The image feature embeddings at multiple scales are decoded using an image segmentation model decoder to generate an initial mask. (3.2) Convert the initial mask into an initial polygonal outline; The initial polygonal outline is converted into a vertex sequence representation by normalization; (3.3) The polygon optimization network is trained using the vertex sequence representation; the Transformer-based sequence regression structure in the polygon optimization network is fine-tuned using the Adapter-Tuning in the parameter efficient fine-tuning to form a fine-tuned Transformer-based sequence regression structure; the fine-tuned Transformer-based sequence regression structure finely adjusts the position of each vertex in the vertex sequence; and the training is constrained using a loss function. The loss function is calculated as L = L_vertex + L_smooth + L_IoU; where L_vertex is the geometric error loss. L_smooth is the smoothing constraint loss; L_IoU is the matching loss; (4) When the training results reach the predetermined termination condition, the model training is terminated and the polygon structure recognition model is obtained.

2. The method as described in claim 1, characterized in that, The image segmentation models include SAM3 model, SAM model, HRNet, U-Net, and DeepLabV3+.

3. The method as described in claim 1, characterized in that, In step (3.1), the specific steps include: inputting the fused image data into a SAM3 model encoder with fixed parameters, and using low-rank adaptation in parameter efficient fine-tuning to perform low-rank decomposition on the attention weights of the SAM3 model, thereby generating multi-scale image feature embeddings; The SAM3 model mask generation head is used to decode the multi-scale image feature embeddings to generate an initial mask.

4. The method as described in claim 1, characterized in that, Step (3.2) specifically includes the following steps: converting the initial mask into an initial polygonal contour using an edge detection and contour fitting algorithm; and converting the initial polygonal contour into a vertex sequence representation using geometric normalization.

5. A method for generating vectorized polygon data, characterized in that, The method includes the following steps: (S1) Provides various image data; (S2) Preprocessing the various image data includes the following steps: (S2.1) The various image data are processed including geometric correction, spectral normalization, radiometric consistency, size normalization, pixel value standardization and color space adjustment to construct the input dataset; (S2.2) Slice the input data in the input dataset to form pyramid tiles; incorporate the pyramid tiles into the tile cache; and (S2.3) Vectorize and annotate the input dataset to obtain fused image data; (S3) The fused image data is processed using a polygon structure recognition model to generate vectorized polygon data; the polygon structure recognition model is constructed using the method described in claim 1.

6. A method for spatial analysis of sensitive points of transmission lines based on GIS, characterized in that, The method includes the following steps: (I) Provides data related to power transmission lines, various image data containing building images, field survey data, and administrative division and building point of interest (POI) data; the data related to power transmission lines includes centerline data and conductor crossarm parameters; the field survey data provides sensitive point information; (II) Based on the centerline data and crossarm parameters of the transmission line, obtain the spatial geometric position data of the side conductors of the transmission line; The various image data are processed using the method described in claim 5 to obtain building vector data; (III) Associate and merge the vector data of one or more buildings belonging to the same sensitive point range to form a sensitive point object; (IV) Analyze the sensitive point objects, including: (ivi) Match the sensitive point object with the administrative division and building POI data to obtain the attributes of the sensitive point object; (ivii) In each of the sensitive point objects, based on the spatial geometric position data of the edge guide, analyze the vector data of each building to obtain the building closest to the edge guide, the shortest distance, and the orientation information; Steps (ivi) and (ivii) are interchanged or performed simultaneously; thus, spatial analysis of sensitive points on transmission lines is performed.

7. A system for target recognition and analysis based on artificial intelligence and GIS, characterized in that, The system includes: (U1) Input unit, configured to input data, the data including various image data containing building images, transmission line related data, field survey data, and administrative division and building POI data; the transmission line related data includes transmission line centerline data and conductor crossarm parameters; the field survey data provides sensitive point information; (U2) Identification unit, the identification unit being configured to perform the following operations: (u2.1) Preprocess the various image data to obtain fused image data; (u2.2) The fused image data is processed using a polygon structure recognition model to generate building vector data; the polygon structure recognition model is constructed using the method described in claim 1; (U3) Analysis unit, the analysis unit being configured to perform the following operations: (u3.1) Based on the centerline data of the line and the crossarm parameters of the conductor, obtain the spatial geometric position data of the side conductor of the transmission line; (u3.2) Associate and merge one or more building vector data belonging to the same sensitive point range to form a sensitive point object; (u3.3) Analyzing the sensitive point objects includes: matching the sensitive point objects with the administrative division and building POI data to obtain the attributes of the sensitive point objects; in each sensitive point object, analyzing the vector data of each building based on the spatial geometric position data of the edge guide to obtain the building closest to the edge guide, the shortest distance, and the orientation information; thereby obtaining the analysis results; (U4) Output unit, which is configured to output the analysis results of the analysis unit.

8. The system as described in claim 7, characterized in that, The data also includes one or more parameters selected from the following: task number for identification, analysis and evaluation, scope of identification, analysis and evaluation, line type, voltage level, crossarm length, and pipe rack radius.

9. The system as described in claim 7, characterized in that, The analysis results include a sensitive point questionnaire and / or a sensitive point atlas; The sensitive point survey form includes: administrative region information of the sensitive point object, name of the sensitive point, function of the sensitive point, number of households within the evaluation range, building structure within the evaluation range, structure of the nearest building to the transmission line conductor, direction information, environmental impact factor, noise environmental protection requirements, platform within the evaluation range, and erection form. The sensitive point atlas includes: the relative position of the sensitive point to the transmission line, the distance between the sensitive point and the transmission line, the name of the sensitive point, building information, and on-site photos; The analysis results also include spatial vector data of the side conductors, demolition range, and evaluation range for the selected line type.

Images