Geodetic survey area measurement method, device, equipment and medium

Abstract

The application relates to a land survey area measurement method, device, equipment and medium. The method comprises the following steps: pre-processing and scene complexity evaluation are performed on surveying and mapping images, a comprehensive state vector is generated in combination with a computing resource state and an aircraft state; the comprehensive state vector is input into a deep reinforcement learning decision model to obtain an inference configuration action; the surveying and mapping images are cropped and / or scaled according to the inference configuration action to generate adaptive input images, and a semantic segmentation model is used to perform forward inference on the adaptive input images to obtain a semantic segmentation result image; area results and geographical boundary information are generated and cached based on geographical reference information and the semantic segmentation result image; spatial fusion and deduplication processing are performed on all the surveying and mapping images based on the geographical boundary information and the area results, a target land class continuous distribution map is generated, and the total area of the target land class in the surveying and mapping area is calculated based on the target land class continuous distribution map. The method can improve the task completion rate and the overall surveying and mapping efficiency.

Description

Technical Field

[0001] This invention belongs to the field of surface surveying technology, and in particular relates to a method, apparatus, equipment and medium for determining the area of ​​a geodetic survey. Background Technology

[0002] With the development of the field of geodetic survey, UAV remote sensing technology has emerged. By carrying visible light or multispectral cameras, it can acquire ultra-high resolution images of target areas as needed, becoming a tool for quickly determining the area of ​​specific land surface types. It has the characteristics of being mobile and flexible, having high resolution, and relatively low cost.

[0003] The application process of traditional UAV remote sensing technology revolves around image acquisition, real-time processing, and area calculation. First, the visible light or multispectral camera on the UAV platform is used to collect ultra-high resolution images of the target area. Then, semantic segmentation models are deployed using airborne edge computing equipment to segment specific land surface types in the images. Based on the segmentation results, the area of ​​the target land surface type is calculated, enabling efficient on-demand mapping.

[0004] However, in remote areas with no or weak network coverage, the above methods are limited by the computing power, memory and battery life of airborne edge computing devices. Deploying high-precision deep learning models can lead to data accumulation and loss, and shorten flight time. On the other hand, using lightweight models cannot guarantee segmentation accuracy, resulting in excessive area measurement errors, which are difficult to meet the needs of practical applications. Summary of the Invention

[0005] Therefore, it is necessary to provide a method, apparatus, equipment, and medium for determining the area of ​​a land survey that can optimally balance processing speed and processing accuracy under limited resource constraints, in order to address the above-mentioned technical problems.

[0006] Firstly, this application provides a method for determining the area of ​​a land survey, including:

[0007] During the real-time acquisition of UAV flight mapping, the computing resource status of the airborne edge computing device, the aircraft status, and the currently captured mapping images are collected. The mapping images are preprocessed and the scene complexity is evaluated. A comprehensive state vector is generated by combining the computing resource status and the aircraft status.

[0008] The comprehensive state vector is input into a pre-trained deep reinforcement learning decision model to obtain inference and configuration actions; the deep reinforcement learning decision model is deployed on an airborne edge computing device.

[0009] Based on the input resolution specified by the inference configuration action, the preprocessed mapping image is cropped and / or scaled to generate an adapted input image. The semantic segmentation model corresponding to the inference configuration action is dynamically loaded to perform forward inference on the adapted input image to obtain the semantic segmentation result map of the target land class.

[0010] Based on the georeferenced information and semantic segmentation results of the surveying and mapping image, calculate the area of ​​the target land type within the current surveying and mapping image, and generate and cache the area results and geographic boundary information;

[0011] After the UAV completes the mapping flight along the predetermined route, based on the cached area results and geographic boundary information, all mapping images are spatially fused and deduplicated to generate a continuous distribution map of target land types for the entire mapping area. The total area of ​​target land types in the mapping area is then calculated based on the continuous distribution map of target land types.

[0012] In one embodiment, the deep reinforcement learning decision model obtains the inference configuration action through the following method:

[0013] The integrated state vector is analyzed to obtain several normalized parameters, including CPU utilization, GPU utilization, chip temperature, remaining power estimate, ground speed, and scene complexity.

[0014] Multiple normalized parameters are concatenated into a one-dimensional state tensor, and the policy network of the deep reinforcement learning decision model is used to propagate the one-dimensional state tensor forward to obtain the action value corresponding to each inference configuration action in the pre-constructed inference configuration action space; the action value is the expected long-term cumulative reward for selecting each inference configuration action in the current drone state.

[0015] The inference configuration action with the highest action value is selected as the output; the inference configuration action includes the selected semantic segmentation model index, input resolution, and post-processing switch flag.

[0016] In one embodiment, the mapping imagery is preprocessed and scene complexity is assessed. A comprehensive state vector is generated by combining the computational resource status and the aircraft status, including:

[0017] Radiometric and geometric distortion corrections are performed on the survey images to obtain corrected images. Based on coarse georeferenced information, the corrected images are scaled to a uniform reference size to obtain preprocessed survey images.

[0018] The preprocessed mapping images are input into a pre-trained mini convolutional neural network to obtain scene category probability distribution vectors.

[0019] The gray-level co-occurrence matrix texture entropy is calculated from the preprocessed survey image, and the scene complexity vector is obtained by combining the scene category probability distribution vector;

[0020] Feature extraction is performed on the computational resource state and the aircraft state respectively to obtain the resource state vector and the flight state vector;

[0021] A comprehensive state vector is constructed by fusing the scenario complexity vector, resource state vector, and flight state vector.

[0022] In one embodiment, the preprocessed mapping image is cropped and / or scaled according to the input resolution specified by the inference configuration action to generate an adapted input image. The semantic segmentation model corresponding to the inference configuration action is then dynamically loaded to perform forward inference on the adapted input image, resulting in a semantic segmentation map of the target land class, including:

[0023] Perform bilinear interpolation scaling and / or center cropping on the preprocessed mapping image to obtain an adapted input image that matches the input resolution;

[0024] Load the semantic segmentation model with the corresponding semantic segmentation model index from the multi-precision model library;

[0025] The adapted input image is fed into the semantic segmentation model for forward inference to obtain the initial segmentation probability map;

[0026] If the post-processing switch flag is enabled, the initial segmentation probability map is optimized by the conditional random field algorithm to obtain the semantic segmentation result map.

[0027] If the post-processing switch flag is off, the initial segmentation probability map is thresholded to obtain the semantic segmentation result map.

[0028] In one embodiment, after the UAV completes its predetermined mapping flight, based on the cached area results and geographic boundary information, spatial fusion and deduplication are performed on all mapping images to generate a continuous distribution map of target land types for the entire mapping area. The total area of ​​target land types in the mapping area is then calculated based on this continuous distribution map, including:

[0029] Based on the bounding boxes and geographic coordinates of the area results and geographic boundary information, the corresponding semantic segmentation result maps of all survey images are projected onto a unified global geographic coordinate system to obtain the projection results.

[0030] Based on the projection results, for any two overlapping mapping images corresponding to semantic segmentation result maps, a weighted fusion is performed within the overlapping area based on the credibility of the semantic segmentation results of each pixel in different mapping images to obtain a global probability map; the credibility is determined by the accuracy level and input resolution of the semantic segmentation model associated with the semantic segmentation result map.

[0031] Threshold segmentation is performed on the global probability map to obtain a binarized global target land class distribution map, and the global target land class distribution map is transformed into a continuous distribution map of target land classes through a connected component analysis algorithm.

[0032] Based on the geographic coordinates of each polygon in the continuous distribution map of the target land type, the area is calculated for each polygon using the area calculation formula in the global geographic coordinate system, and the areas are summed to obtain the total area of ​​the target land type.

[0033] Secondly, this application also provides a geodetic survey area measuring device, comprising:

[0034] The status module is used to collect the computing resource status of the airborne edge computing device, the aircraft status, and the currently captured mapping images during the UAV flight mapping process in real time. It also performs preprocessing and scene complexity assessment on the mapping images and generates a comprehensive status vector by combining the computing resource status and the aircraft status.

[0035] The resource allocation module is used to input the comprehensive state vector into the pre-trained deep reinforcement learning decision model to obtain the inference configuration action; the deep reinforcement learning decision model is deployed on the airborne edge computing device.

[0036] The segmentation module is used to crop and / or scale the preprocessed mapping image according to the input resolution specified by the inference configuration action, generate an adapted input image, and dynamically load the semantic segmentation model corresponding to the inference configuration action to perform forward inference on the adapted input image to obtain the semantic segmentation result map of the target land class.

[0037] The analysis module is used to calculate the area of ​​the target land class within the current mapping image based on the georeferenced information and semantic segmentation results of the mapping image, and to generate and cache the area results and geographic boundary information.

[0038] The measurement results module is used to perform spatial fusion and deduplication of all mapping images based on cached area results and geographic boundary information after the UAV completes the mapping flight along the predetermined route, generate a continuous distribution map of target land types in the entire mapping area, and calculate the total area of ​​target land types in the mapping area based on the continuous distribution map of target land types.

[0039] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the above-described geodetic survey area determination methods.

[0040] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the above-described geodetic survey area determination methods.

[0041] The aforementioned method, apparatus, equipment, and medium for determining the area of ​​a land survey utilize real-time acquisition of the resource status, aircraft status, and mapping images from an UAV-borne edge computing device. It assesses the complexity of the image scene to generate a comprehensive state vector, which is then input into a pre-trained deep reinforcement learning decision model deployed on the airborne device. This model dynamically determines the optimal semantic segmentation model and its inference configuration. Based on the configuration, it adaptively processes the images to obtain segmentation results and calculates the area of ​​each frame for caching. Finally, after the flight, it spatially fuses and deduplicates all cached results to generate a continuous distribution map and calculates the total area. This achieves a dynamic adaptive processing closed loop based on real-time state perception and reinforcement learning decision-making in an offline environment. This allows the UAV to intelligently allocate its limited computing power according to its own resources, flight status, and scene content. While ensuring overall measurement accuracy, it maximizes the effective mapping area and mission reliability of a single flight, fundamentally resolving the acute contradiction between the limited computing power of airborne edge devices and the need for real-time high-precision mapping in remote areas without network access. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 This is a schematic flowchart of the method for determining the area of ​​a land survey according to the present invention;

[0044] Figure 2 This is a flowchart illustrating the steps of step S102.

[0045] Figure 3 This is a flowchart illustrating the steps of step S105.

[0046] Figure 4 This is a structural diagram of the geodetic survey area measuring device of the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0048] In one embodiment, such as Figure 1As shown, a method for determining the area of ​​a land survey is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, and further to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0049] S101. Real-time acquisition of the computing resource status of the airborne edge computing device, the aircraft status, and the currently captured mapping images during the UAV flight mapping process, and preprocessing and scene complexity assessment of the mapping images, and generating a comprehensive state vector by combining the computing resource status and the aircraft status.

[0050] In a schematic manner, the real-time data acquisition process relies on the UAV's onboard sensing and computing system to capture multi-dimensional data. The acquisition of computing resource status is achieved through the system monitoring interface of the onboard edge computing device, specifically covering CPU utilization, GPU utilization, available memory capacity, chip core temperature, real-time power consumption, and remaining battery power. The acquisition of aircraft status is based on the bus interface of the UAV flight control system, extracting key information such as ground flight speed, relative altitude to the ground, attitude angle, and flight path progress. The flight control system outputs data in real time through the GPS module and inertial measurement unit (IMU), aligning it with the computing resource status data via a data synchronization protocol. Mapping images are acquired by the onboard visible light or multispectral camera according to a preset flight path overlap rate. Camera acquisition parameters are automatically adjusted based on ambient light intensity, and the raw images are stored in a lossless format in a temporary buffer.

[0051] Furthermore, radiometric correction, lens distortion correction, and georeferenced initial correction are sequentially performed on the mapping image. Radiometric correction employs a linear correction method based on ground control points to eliminate the impact of illumination variations on image grayscale values. Lens distortion correction corrects the image pixel coordinates using the camera intrinsic parameter matrix, compensating for radial and tangential distortion. Georeferenced initial correction combines real-time position information provided by the flight control system with camera extrinsic parameters to initially map the image pixel coordinates to the geodetic coordinate system. Scene complexity assessment employs a multi-feature fusion strategy; specifically, it calculates the image's texture entropy and local standard deviation, where the texture entropy is... ,in, For texture entropy, This represents the total number of gray levels in the image. grayscale value The probability of occurrence, this index reflects the complexity of image texture; the local standard deviation is... ,in, For local standard deviation, This refers to the size of a local window. Coordinates within the window The pixel grayscale value at that location, This is the mean of the pixel grayscale values ​​within the window, used to quantify the severity of local grayscale changes in the image. The texture entropy, local standard deviation, and the acquired computational resource status and aircraft status data are normalized; for example, the min-max normalization method is used. ,in, For the normalized data, The original data, The minimum value of this type of data. The maximum value of this type of data is used to finally concatenate all normalized parameters in a preset order to generate a comprehensive state vector with fixed dimensions.

[0052] S102. Input the comprehensive state vector into the pre-trained deep reinforcement learning decision model to obtain the inference configuration action; the deep reinforcement learning decision model is deployed on an airborne edge computing device.

[0053] Optionally, the deep reinforcement learning decision model is an edge-optimized model based on offline training. Its training process is completed in a simulation environment on a high-performance ground workstation. The simulation environment accurately simulates the computational load characteristics of the airborne edge computing device, changes in the aircraft's flight state, and image data features under different scenarios. The model adopts a deep Q-network (DQN) architecture. The input layer dimension is consistent with the dimension of the integrated state vector, the hidden layer contains three fully connected layers, the activation function is the ReLU function, and the output layer dimension matches the number of actions in the preset inference configuration. The output result is the Q-value of each action, representing the expected cumulative reward of that action in the current state. Furthermore, during the model training phase, a multi-objective reward function guides the agent to learn the optimal decision strategy. The multi-objective reward function is... ,in, For the first Instant rewards for each moment , , , The weighting coefficients for accuracy reward, speed reward, power consumption penalty, and temperature penalty are pre-configured according to the priority requirements of the geodetic survey task. For the first The accuracy evaluation value of the semantic segmentation result at each time step is calculated by a lightweight evaluation network; For the first The speed reward value at any given moment is positive when the image processing time is less than or equal to the acquisition interval, and negative otherwise. The reward value is positively correlated with the difference in processing time. For the first The power consumption penalty value at any given moment is negatively correlated with the real-time power consumption of the airborne edge computing device; For the first The temperature penalty value increases significantly when the chip temperature exceeds a preset threshold. For example, after model training, INT8 quantization compression technology is used to optimize network parameters, reducing model storage and inference time. The quantization process determines the quantization range of parameters through a calibration dataset, ensuring that the loss of decision accuracy after quantization is controlled within an acceptable range. The compressed model is deployed in the memory of an onboard edge computing device. When the integrated state vector is input, the model quickly calculates the Q-value of each inference configuration action through forward propagation, and uses a greedy strategy to select the action with the largest Q-value as the output. This inference configuration action includes three core configuration items: semantic segmentation model selection, input resolution parameters, and post-processing switch state, which are decoded into executable operation instructions through preset encoding rules.

[0054] S103. Based on the input resolution specified by the inference configuration action, the preprocessed mapping image is cropped and / or scaled to generate an adapted input image, and the semantic segmentation model corresponding to the inference configuration action is dynamically loaded to perform forward inference on the adapted input image to obtain the semantic segmentation result map of the target land class.

[0055] Based on the input resolution specified by the inference configuration action, cropping or scaling operations are performed on the preprocessed mapping image. Specifically, when the input resolution is less than the resolution of the preprocessed image, a sliding window-based cropping method is used, with the window size consistent with the input resolution and the window step size set according to the flight path overlap rate to ensure that there is a preset proportion of overlapping area between adjacent cropped blocks, avoiding boundary breaks during subsequent stitching. When the input resolution is greater than the resolution of the preprocessed image, a bicubic interpolation algorithm is used for scaling, that is, the target pixel value is calculated by weighted averaging of the gray values ​​of 16 adjacent pixels, balancing scaling accuracy and computational efficiency. When the input resolution is consistent with the resolution of the preprocessed image, the preprocessed image is directly output as the adapted input image.

[0056] Dynamic loading of semantic segmentation models is achieved through the model management module. The model library pre-stores semantic segmentation models of different accuracy levels, including high-precision models, balanced models, and lightweight models. All models have been pre-trained and converged using the target land class dataset. Based on the model selection instruction in the inference configuration action, the corresponding model parameters are loaded from the storage medium to the GPU memory of the onboard edge computing device using memory mapping technology. If other models are already loaded, the GPU memory resources occupied by those models are released first to avoid memory leaks. After the model is loaded, the adapted input image is normalized according to the model requirements, mapping pixel grayscale values ​​to the [0,1] interval, and then input into the model for forward inference. Illustratively, the model's encoder extracts features from the adapted input image, and the decoder performs semantic segmentation based on the extracted feature map, outputting a probability map of each pixel belonging to the target land class. If the inference configuration action specifies enabling post-processing, Conditional Random Field (CRF) optimization is performed on the probability map. CRF optimizes the spatial consistency of the segmentation results by defining the potential function between pixels, and its energy function is... ,in, The total energy of CRF, This is the semantic segmentation result image. , pixels , Category tags, This is the potential energy term for a single pixel, determined by the probability map output by the model. This represents the paired pixel potential term, which is related to the spatial distance and grayscale similarity between pixels. After forward inference and post-processing, the semantic segmentation result image of the target land class is obtained. This image is a single-channel image, and the pixel value is the category label, i.e., the target land class or the background.

[0057] S104. Based on the geographic reference information and semantic segmentation results of the surveying and mapping image, calculate the area of ​​the target land type within the current surveying and mapping image, and generate and cache the area results and geographic boundary information.

[0058] Schematic representation: Georeferenced Information (GRI) includes geodetic coordinate system parameters of the mapping imagery, camera extrinsic parameters, and flight altitude data. The geodetic coordinate system parameters use a nationally unified geodetic datum, and the camera extrinsic parameters include the camera's position and attitude parameters, jointly provided by the flight control system and camera calibration data. Based on the GRI, the ground area corresponding to a single pixel is calculated. Specifically, the ground sampling distance (GSD) is calculated using the camera focal length, flight altitude, and sensor pixel size. ,in, This is the ground sampling distance, which is the actual length corresponding to a single pixel. For camera focal length, The drone's altitude relative to the ground. This represents the physical size of a single pixel on the sensor. The actual area of ​​a single pixel is... ,in, This represents the actual area corresponding to a single pixel.

[0059] Furthermore, the area of ​​the target land class is calculated by counting the number of pixels of the target land class in the semantic segmentation result image. A connected component analysis algorithm is used to traverse the segmentation result image, and the pixel value is counted as the total number of pixels of the target land class label. Then the area of ​​the target land type within the current mapping image is ,in, This represents the area of ​​the target land cover type corresponding to the current frame image.

[0060] Optionally, the area results and geographic boundary information are cached using a key-value pair storage structure. High-speed solid-state drives (SSDs) are selected as the storage medium to ensure read and write speeds. The key is a combination of the image acquisition timestamp and frame sequence number to ensure data uniqueness; the value contains... The geographic bounding box coordinates and segmentation result map are compressed. The geographic bounding box coordinates are obtained by mapping the pixel coordinates of the four corner pixels of the image to the geodetic coordinate system and stored in latitude and longitude format. The segmentation result map is compressed using a lossless compression algorithm before storage to reduce storage space.

[0061] S105. After the UAV completes the mapping flight along the predetermined route, based on the cached area results and geographic boundary information, all mapping images are spatially fused and deduplicated to generate a continuous distribution map of target land types for the entire mapping area. The total area of ​​target land types in the mapping area is calculated based on the continuous distribution map of target land types.

[0062] Furthermore, after the UAV completes the pre-determined flight path mapping, it initiates a spatial fusion and deduplication process. This involves reading the area results and geographic boundary information of all mapped images from the cache, constructing a spatial index based on the geographic boundary coordinates, and using an R-tree index structure to quickly retrieve overlapping regions of adjacent images. For processing overlapping regions, a confidence-weighted fusion strategy is employed. Specifically, the confidence map, converted from the probability map output by the model, is extracted from the semantic segmentation result map corresponding to each overlapping image. The weighted average confidence of each pixel within the overlapping region is calculated, with the weight being the segmentation confidence of the corresponding image. ,in, Coordinates within the overlapping region Confidence after fusion at the point, The number of images covering this pixel, For the first The confidence level of an image at that pixel. For the first The weighting coefficients of each image are determined by the evaluation results of the image acquisition quality and segmentation accuracy.

[0063] Furthermore, based on the fused confidence map, an adaptive threshold segmentation algorithm is used to determine the pixel range of the target land class. The threshold is automatically calculated using the maximum inter-class variance method to ensure the accuracy of the segmentation results. A polygon fitting algorithm is used to convert the pixel regions of the target land class into vector boundaries. The Douglas-Peucker algorithm is then used to simplify the vector boundaries, removing redundant vertices and improving data storage and computational efficiency. For overlapping vector regions, a polygon Boolean merging operation is performed to eliminate redundantly calculated areas, generating a continuous distribution map of the target land class across the entire surveyed area. This distribution map is stored in a vector polygon dataset format and contains complete boundary and attribute information of the target land class.

[0064] The total area of ​​the target land types in the surveyed area is calculated by summing the vector areas of the continuous distribution map. For example, the Gauss-Kruger projection is used to convert the geodetic coordinates corresponding to the vector boundaries into planar coordinates, and the Shoelace formula is used to calculate the area of ​​each vector surface. ,in, The area of ​​a single vector surface. The number of vertices on the vector face. For the first The planar coordinates of the vertices, Sum the areas of all vector surfaces to obtain the total area of ​​the target land type in the surveyed area. It generates a comprehensive report containing total area data, continuous distribution maps, and the original results for each frame.

[0065] In the aforementioned method for determining the area of ​​a land survey, the computing resource status, aircraft status, and currently captured survey images of the airborne edge computing device are collected in real time during the UAV flight surveying process. The survey images are preprocessed and scene complexity is assessed. A comprehensive state vector is generated by combining the aforementioned computing resource status and aircraft status. This comprehensive state vector is input into a pre-trained deep reinforcement learning decision model deployed on the airborne edge computing device to obtain inference configuration actions. Based on the input resolution specified by the action, the preprocessed survey images are cropped and / or scaled to generate adapted input images. The corresponding semantic segmentation model is dynamically loaded to perform forward inference on the adapted input images to obtain the semantics of the target land class. The segmentation result map, based on the georeferenced information and semantic segmentation result map of the surveying image, calculates the area of ​​the target land class within the current image and caches the area results and geographic boundary information. After the UAV completes the predetermined route surveying, all cached area results and geographic boundary information are spatially fused and deduplicated to generate a continuous distribution map of the target land class in the entire surveying area, and the total area of ​​the target land class is calculated accordingly. This achieves online adaptive and highly intelligent autonomous decision-making for geodetic survey area determination under harsh offline conditions, resolving the contradiction between static model deployment and dynamic environment mismatch, improving system robustness, task completion rate, and overall surveying efficiency, while also possessing the characteristics of continuous optimization and flexible expansion.

[0066] In one embodiment, such as Figure 2 As shown, the deep reinforcement learning decision model obtains the inference configuration action through the following method:

[0067] S201. Analyze the integrated state vector to obtain multiple normalized parameters; the normalized parameters include CPU utilization, GPU utilization, chip temperature, remaining power estimate, ground speed, and scene complexity.

[0068] Indicatively, the integrated state vector is a fixed-dimensional one-dimensional tensor, with its internal parameters arranged in a preset standardized order. The parsing process is implemented through the preprocessing module of the model input layer, which has parameter index mapping and data verification functions. During parsing, the preprocessing module extracts the values ​​of the corresponding dimensions from the integrated state vector one by one according to the preset parameter position index table. The extracted normalized parameters cover CPU utilization, GPU utilization, chip temperature, remaining power estimation, ground velocity, and scene complexity.

[0069] The normalized values ​​of CPU and GPU utilization directly reflect the current computational load level of the airborne edge computing device. Their values ​​are mapped to the [0,1] interval, with values ​​closer to 1 indicating more efficient use of computational resources. The normalized value of chip temperature corresponds to the thermal state of the device's core chip. It is obtained by normalizing the raw temperature data collected by the device's temperature sensor using a min-max normalization process, representing the relative relationship between the chip's current temperature and the safe temperature threshold. The normalized value of the remaining battery power estimate is obtained by normalizing the remaining battery power percentage provided by the battery management system, directly reflecting the drone's current endurance. The normalized value of ground velocity is derived from real-time ground velocity data collected by the flight control system, reflecting the degree of matching between the drone's flight efficiency and image acquisition frequency. The normalized value of scene complexity integrates the comprehensive evaluation results of image texture entropy and local standard deviation, obtained through weighted summation and normalization, comprehensively reflecting the complexity of the surface features in the current mapping image. Optionally, during the parsing process, the preprocessing module verifies the validity of each extracted normalized parameter. If the parameter value exceeds the [0,1] range or there is an abnormal jump, the parameter value from the previous moment is used to replace it, ensuring the stability and reliability of the parsing results.

[0070] S202. Concatenate multiple normalized parameters into a one-dimensional state tensor, and use the policy network of the deep reinforcement learning decision model to perform forward propagation on the one-dimensional state tensor to obtain the action value corresponding to each inference configuration action in the pre-constructed inference configuration action space; the action value is the expected long-term cumulative reward for selecting each inference configuration action in the current drone state.

[0071] Furthermore, the concatenation of multiple normalized parameters is achieved through tensor dimensionality expansion and sequential combination. Each normalized parameter is treated as a one-dimensional scalar tensor, and these parameters are sequentially concatenated according to a fixed order: CPU utilization, GPU utilization, chip temperature, remaining battery power estimate, ground speed, and scene complexity, forming a tensor with a dimension of... A one-dimensional state tensor, wherein, This represents the total number of normalization parameters. The splicing process is supported by a tensor operations library, which ensures the continuity of data storage for the one-dimensional state tensor through contiguous memory allocation, reducing data read latency during subsequent forward propagation.

[0072] The policy network of the deep reinforcement learning decision model adopts a deep Q-network architecture, and its forward propagation process sequentially passes through the input layer, hidden layer, and output layer. Specifically, the input layer receives a one-dimensional state tensor and converts it into a feature vector that the network can process. The dimension of the feature vector is the same as the number of neurons in the input layer. The hidden layer contains three fully connected layers. The first fully connected layer maps the input feature vector to a high-dimensional feature space, and the mapping process follows... ,in, This is the output feature vector of the first hidden layer. This is the weight matrix of the first fully connected layer. The feature vector output by the input layer. This represents the bias vector of the first fully connected layer. The ReLU activation function is used to introduce non-linear feature transformation. The second and third fully connected layers sequentially extract deep features from the feature vectors output by the previous layer, following the same computational logic as the first layer. Through linear transformations of the weight matrix and bias vector combined with the ReLU activation function, the feature representation capability is progressively optimized. The output layer is a fully connected layer, with the number of neurons equal to the number of actions in the pre-constructed inference configuration action space. A linear transformation maps the high-dimensional feature vector of the last hidden layer to the action value corresponding to each inference configuration action. The mapping formula is: ,in, Action value, i.e., the action selected for inference configuration in the current state. The expected long-term cumulative reward The weight matrix of the output layer. This is the output feature vector of the last hidden layer. This is the bias vector for the output layer. This is the one-dimensional state tensor corresponding to the current state of the drone.

[0073] Optionally, the action value is the weighted sum of the expected rewards at all future moments after the action is selected, i.e. ,in, Represents the mathematical expectation. This is a discount factor, with a value range of [0,1], used to weigh the importance of current immediate rewards against future long-term rewards. Indexing future time steps For the first Instant rewards for each moment For the first Real-time drone status For the first The inference configuration action selected at each moment. During the forward propagation, the network quickly calculates the action value corresponding to each inference configuration action using the weight parameters and bias parameters obtained during pre-training, completing the mapping from state to action value.

[0074] S203. Select the inference configuration action with the highest action value as the output; the inference configuration action includes the selected semantic segmentation model index, input resolution, and post-processing switch flag.

[0075] For example, the comparison of action values ​​is implemented using the argmax function, which iterates through the action values ​​of all neurons in the output layer and selects the index corresponding to the action value with the largest value. ,in, Configure actions for the optimal inference output. Configure an action space for pre-built inference, containing all possible combinations of inference configuration actions. The value of each action is defined as its corresponding action value. During the selection process, if multiple inference configuration actions have the same action value and are all the maximum value, the action with the highest priority is selected according to the preset action priority order in the action space to ensure the uniqueness of the selection result.

[0076] Furthermore, the decoding of inference configuration actions is based on a preset action encoding table, which stores the mapping relationship between action indices and specific configuration items. The selected semantic segmentation model index corresponds to a preset semantic segmentation model with different precision levels in the model library. Each index uniquely points to a pre-trained semantic segmentation model, and the model library quickly locates and calls the corresponding model file through the index. The input resolution is a preset discrete resolution level, and each resolution level is adapted to the model's input requirements and the computing power of the onboard equipment to ensure that the model can efficiently process image data of the corresponding resolution. The post-processing switch flag is a binary value used to control whether to perform post-processing operations on the probability map output by the model. When the flag is 1, post-processing is enabled; when it is 0, post-processing is disabled. The type and parameters of the post-processing operation have been pre-configured during the model training phase. After decoding, the selected semantic segmentation model index, input resolution, and post-processing switch flag are combined into an executable inference configuration action instruction.

[0077] In one embodiment, the mapping imagery is preprocessed and scene complexity is assessed. A comprehensive state vector is generated by combining the computational resource status and the aircraft status, including:

[0078] S11. Perform radiometric and geometric distortion correction on the survey image to obtain a corrected image. Based on coarse geographic reference information, scale the corrected image to a uniform reference size to obtain a preprocessed survey image.

[0079] Radiometric correction employs a linear correction method based on the sensor's radiometric response model. This method eliminates the influence of illumination variations and sensor noise by establishing a mapping relationship between the input grayscale value and the true radiance. The correction formula is as follows: ,in, This is the corrected radiance value. The sensor gain coefficient. The original grayscale value of the image. This represents the dark current offset of the sensor. and All parameters were pre-obtained through sensor calibration experiments and stored in the system parameter library. Geometric distortion correction addresses the radial and tangential distortions of the camera lens. Based on a preset camera intrinsic parameter matrix and distortion coefficients, it is achieved through pixel coordinate transformation, using the following formula: ,in, These are the pixel coordinates of the distorted image. These are the corrected pixel coordinates. , The radial distortion coefficient is... , The tangential distortion coefficient is... The distance from the pixel to the center of the image is the square of the distance. All distortion parameters are predetermined through the camera calibration process.

[0080] Furthermore, coarse georeferenced information is provided in real time by the UAV flight control system, including the UAV's current geodetic coordinates, flight altitude, and attitude angle data. Based on this information, combined with camera focal length and field of view parameters, the ground coverage and scale of the image are calculated. The scaling of the corrected image employs a bilinear interpolation algorithm, determining the target pixel value by weighted averaging of the grayscale values ​​of four adjacent pixels. This ensures a balance between detail preservation and computational efficiency in the scaled image. During scaling, the scale requirements corresponding to the coarse georeferenced information are strictly followed, uniformly scaling all corrected images to a preset reference size to form a pre-processed mapping image with consistent specifications.

[0081] S12. Input the preprocessed mapping image into the pre-trained micro convolutional neural network to obtain the scene category probability distribution vector.

[0082] The illustrative example shows a pre-trained miniature convolutional neural network, a lightweight network architecture designed specifically for airborne edge computing devices. The total number of network parameters is kept within a low overhead range to ensure millisecond-level inference speed. The network structure consists of an input layer, three convolutional layers, two pooling layers, and an output layer. The input layer receives pre-processed mapping images, with its channel count matching the camera's acquisition mode. The first convolutional layer uses small-sized convolutional kernels for feature extraction, and the convolution operation formula is... ,in, The output feature map of the convolutional layer in coordinates First The characteristic values ​​of each channel, The number of channels in the input image. The kernel size is the convolution kernel size. For the convolution kernel at position Input Channel With output channel The corresponding weight value, The input image contains pixel values ​​at the corresponding positions and channels. For the first The bias term for each output channel. The ReLU activation function is used. Convolutional layers and pooling layers alternate. The pooling layers use max pooling to compress the feature map size, reducing computation while preserving key features. The output layer is a fully connected layer with the same number of neurons as the preset number of scene categories. The softmax activation function converts the output of the fully connected layer into a scene category probability distribution vector. ,in, For the first The probability value of the scenario type, For the fully connected network of the output layer, the first The original output value of the scenario. This represents the total number of scene categories. This serves as an index for scene categories. The pre-training process of this mini convolutional neural network is based on an annotated image dataset containing various typical scenes from land surveys, such as farmland, forest land, construction land, and water areas. The training objective is to minimize the cross-entropy loss function to ensure that the network can accurately output the probability distributions for various scenes. The final output scene category probability distribution vector has a dimension of [missing information]. Each element in the vector corresponds to the probability of a certain scenario occurring, and the sum of all elements is 1.

[0083] S13. Calculate the gray-level co-occurrence matrix texture entropy from the preprocessed mapping image, and combine it with the scene category probability distribution vector to obtain the scene complexity vector.

[0084] Optionally, the gray-level co-occurrence matrix is ​​constructed by performing gray-level compression on the preprocessed mapping image, mapping the original gray values ​​to a preset number of discrete gray levels, thus reducing computational complexity. At specified directions, such as 0°, 45°, 90°, and 135°, and with a step size, the co-occurrence frequency of gray values ​​of adjacent pixels in the image is statistically analyzed to form the gray-level co-occurrence matrix. Matrix elements Represents grayscale value With grayscale value The number of times adjacent occurrences occur under a specified direction and step size. Texture entropy is calculated based on the gray-level co-occurrence matrix (GLCM), which reflects the complexity of the image texture. The calculation process first normalizes the GLCM to obtain the probability matrix. The normalization formula is ,in, These are the elements of the normalized probability matrix. This represents the number of gray levels after compression. The elements are the original gray-level co-occurrence matrix, and the denominator is the sum of all elements in the gray-level co-occurrence matrix. The texture entropy is... ,in, The texture entropy of the gray-level co-occurrence matrix. For the elements of the normalized probability matrix, when When, define This avoids generating meaningless values ​​in logarithmic operations.

[0085] Furthermore, the scene complexity vector is constructed using a feature concatenation strategy. The calculated gray-level co-occurrence matrix texture entropy is used as a single feature value and concatenated dimensionally with the scene category probability distribution vector to form the scene complexity vector. The first element of the scene complexity vector is the texture entropy. The subsequent elements are the probability values ​​of the scene category probability distribution vector, and the final scene complexity vector has a dimension of . , This represents the total number of scene categories.

[0086] S14. Extract features from the computational resource state and the aircraft state respectively to obtain the resource state vector and the flight state vector.

[0087] Indicatively, the computational resource state vector includes several raw monitoring parameters such as CPU utilization, GPU utilization, available memory capacity, chip core temperature, real-time power consumption, and remaining battery power. The feature extraction process first standardizes these raw parameters using the z-score standardization formula. ,in, These are the standardized eigenvalues. These are the original parameter values. This is the average value of the parameter in historical monitoring data. This represents the standard deviation of the parameter in historical monitoring data. Standardization eliminates the influence of differences in the dimensions of different parameters, ensuring a balanced weighting of each feature in subsequent fusion.

[0088] Core resource status information is extracted through feature aggregation. A sliding window statistical method is used to calculate the mean and rate of change of each standardized parameter within a preset time window. The mean reflects the stable state of the parameter, while the rate of change reflects its dynamic trend. , ,in, To standardize the mean of the parameter over the time window, The number of sampling points within the time window. For the first Standardized parameter values ​​for each sampling point To standardize the rate of change of parameters within a time window, the mean and rate of change of all parameters are combined in a preset order to form a resource status vector. This vector comprehensively reflects the current resource load and dynamic trend of the airborne edge computing device.

[0089] The aircraft's status includes raw data such as ground speed, flight altitude, attitude angles, and flight path progress. Feature extraction involves performing a trigonometric function transformation on the attitude angle data, converting the angle values ​​into sine and cosine values ​​to eliminate feature discontinuities caused by angle periodicity. The transformation formula is as follows: , ,in, The original attitude angle, and These are the transformed feature values. All raw data of the aircraft state and the transformed attitude angle features are standardized, and the instantaneous values ​​of each parameter and the smoothed values ​​obtained by moving average filtering are calculated. The instantaneous values ​​reflect the real-time state, and the smoothed values ​​reflect the trend state. All standardized instantaneous values ​​and smoothed values ​​are combined in a preset order to form a flight state vector. This vector completely represents the flight state and motion trend of the UAV.

[0090] S15. Construct a comprehensive state vector by fusing the scenario complexity vector, resource state vector, and flight state vector.

[0091] Optionally, the construction of the comprehensive state vector adopts a feature concatenation and fusion strategy. This involves verifying the dimensions of the scene complexity vector, resource state vector, and flight state vector separately to ensure that the dimensions of each vector are consistent with the preset dimensions, avoiding dimensional mismatch issues during fusion. Following a fixed order, the three vectors are concatenated to form a one-dimensional comprehensive state vector. The concatenation process is implemented using tensor operations, linking the three vectors end-to-end sequentially to ensure continuous data storage in memory. The dimension of the concatenated comprehensive state vector is the sum of the dimensions of the scene complexity vector, resource state vector, and flight state vector. To further optimize the representational capability of the comprehensive state vector, it is normalized after concatenation. The final comprehensive state vector contains all key information about scene environment features, airborne equipment resource features, and UAV flight state features. Each feature dimension is arranged in a fixed order, providing comprehensive and standardized input to the deep reinforcement learning decision model, supporting the model in making optimal inference and configuration action decisions.

[0092] In one embodiment, the preprocessed mapping image is cropped and / or scaled according to the input resolution specified by the inference configuration action to generate an adapted input image. The semantic segmentation model corresponding to the inference configuration action is then dynamically loaded to perform forward inference on the adapted input image, resulting in a semantic segmentation map of the target land class, including:

[0093] S21. Perform bilinear interpolation scaling and / or center cropping on the preprocessed survey image to obtain an adapted input image that matches the input resolution.

[0094] The preprocessed mapping image already possesses a unified reference size and standardized format. The generation of the adapted input image must strictly adhere to the input resolution specified by the inference configuration action. Based on the relationship between the reference size and the target input resolution, a processing method of bilinear interpolation scaling, center cropping, or a combination of both is dynamically selected. Illustratively, when the target input resolution is larger than the reference size, a bilinear interpolation algorithm is used for scaling. This algorithm constructs a linear interpolation model of pixel values, using the grayscale information of four adjacent pixels to calculate the target pixel value, ensuring the continuity of detail and smooth grayscale transition in the scaled image. The scaling formula is as follows: ,in, To adapt the scaled input image to the coordinate system Pixel value at that location, , These represent the fractional part of the target pixel coordinates relative to the original pixel coordinates, with values ​​ranging from [0,1]. , , , These are the values ​​of the four original pixels adjacent to the target pixel in the preprocessed image. , These are the integer coordinates of adjacent original pixels.

[0095] When the target input resolution is smaller than the reference size, a center cropping strategy is adopted. The image block of the central region is cropped according to the target resolution size, based on the geometric center of the preprocessed image. This ensures that the cropped image retains the core scene information of the original image and avoids invalid information in the edge region from interfering with subsequent segmentation inference.

[0096] When the aspect ratio of the target input resolution is inconsistent with that of the reference size, the preprocessed image is first adjusted to an intermediate size consistent with the aspect ratio of the target resolution using bilinear interpolation scaling. Then, the intermediate-sized image is center-cropped to obtain a finally adapted input image that meets the input resolution requirements. All processing steps must retain the georeferenced information of the image to ensure the accuracy of coordinate mapping in subsequent area calculations. At the same time, the processed adapted input image must meet the preset requirements of the semantic segmentation model for the number of input channels and data types.

[0097] S22. Load the semantic segmentation model with the corresponding semantic segmentation model index from the multi-precision model library.

[0098] The multi-precision model library is a collection of models pre-built and stored in the storage medium of the airborne edge computing device. It includes three levels of semantic segmentation models: high precision, balanced, and lightweight. Each model corresponds to a unique semantic segmentation model index, and the mapping relationship between the index and the model is predefined through the model configuration file.

[0099] As an illustration, the model management module parses the semantic segmentation model index in the inference configuration action, and uses the index to query the corresponding model storage path, model structure parameters, and inference dependencies in the model configuration file. During loading, the model management module uses memory mapping technology to load the parameter file of the target model from the storage medium to the GPU memory. If other semantic segmentation models are already loaded in the current GPU memory, a GPU memory release operation is performed first to reclaim the GPU memory resources occupied by the model, avoiding memory leaks and resource conflicts. After the model is loaded, the model's inference engine is automatically initialized, completing inter-layer connection verification, weight parameter loading, and inference interface binding to ensure that the model is in a ready state. During the loading process, GPU memory usage and loading time are monitored in real time. If insufficient GPU memory or loading timeout occurs, a model degradation mechanism is triggered, automatically loading the low-precision alternative model corresponding to the index to ensure the continuity of the processing flow.

[0100] S23. Input the adapted input image into the semantic segmentation model for forward inference to obtain the initial segmentation probability map.

[0101] Optionally, the input image is preprocessed with data normalization before being input into the semantic segmentation model, mapping the pixel grayscale values ​​to the standard input range used during model training. ,in, These are the normalized pixel values. To adapt the input image in coordinates The original pixel value at that location, The pixel mean of the training dataset for the model. The standard deviation of pixels in the model training dataset. and All are stored in the model configuration file.

[0102] The semantic segmentation model employs a lightweight encoder-decoder architecture. The encoder consists of multiple depthwise separable convolutional layers and bottleneck layers, extracting multi-scale semantic features from the image through progressive downsampling. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function to accelerate model convergence and enhance feature representation capabilities. The decoder uses transposed convolutional layers to upsample the feature maps and performs skip connections with the multi-scale feature maps output by the encoder, restoring image resolution while preserving detailed information. Specifically, during forward inference, the normalized adapted input image is first input into the encoder, where multiple rounds of convolution and downsampling operations generate a high-dimensional semantic feature map. After the feature map is input into the decoder, it undergoes upsampling and feature fusion to generate an initial segmentation probability map with the same size as the adapted input image. The initial segmentation probability map is a multi-channel image, with the number of channels matching the total number of target and background land cover categories. Each channel corresponds to a land cover category, and the pixel value represents the probability that the location belongs to the corresponding land cover category, ranging from [0,1]. The inference process is accelerated through GPU parallel computing to ensure that the inference time meets real-time processing requirements.

[0103] S24. If the post-processing switch flag is enabled, the initial segmentation probability map is optimized by the conditional random field algorithm to obtain the semantic segmentation result map.

[0104] When the post-processing switch flag is enabled, the Conditional Random Field (CRF) algorithm is activated to optimize the boundaries of the initial segmentation probability map. This algorithm improves segmentation accuracy by modeling spatial dependencies and gray-level similarities between pixels, correcting boundary blurring and local misclassification issues in the initial segmentation results. ,in, This represents the total energy value. A set of pixel category labels, , pixels , Category tags, This is the single-pixel potential energy term, calculated from the initial segmentation probability map, representing the pixel. Category The prior probability is , For pixels in the initial segmentation probability map Category The probability of; These are paired pixel potential terms used to measure pixel potential energy. and Category label compatibility, for ,in, For category compatibility coefficient, when The value is positive when the time condition is met, and negative otherwise. , For pixels , spatial coordinates, The square of the spatial distance. , For pixels , grayscale value, For grayscale similarity, , These are the smoothing coefficients for spatial distance and gray-level similarity, respectively. All parameters were optimized and determined using the validation dataset during the model training phase.

[0105] Minimize the total energy function through iterative optimization. The optimal category label assignment result is obtained, and a semantic segmentation result image with clear boundaries and consistent categories is generated. This image is a single-channel image, and the pixel value is the category label of the corresponding land class.

[0106] S25. If the post-processing switch flag is off, the initial segmentation probability map is thresholded to obtain the semantic segmentation result map.

[0107] When the post-processing switch flag is off, a thresholding method is used to convert the initial segmentation probability map into a semantic segmentation result map. Thresholding refers to distinguishing target land classes from background pixels using a preset threshold. The threshold is determined using an adaptive thresholding strategy, dynamically calculating the optimal threshold based on the probability distribution characteristics of the initial segmentation probability map. That is, by statistically analyzing the probability histogram of the target land class channels in the initial segmentation probability map, the probability value that maximizes the variance between the foreground and background land class is found as the threshold. The variance between classes is ,in, For inter-class variance, For candidate thresholds, Threshold The following pixel percentage represents the background percentage. Threshold The percentage of the above pixels represents the percentage of the target land type. Threshold The average probability value of the following pixels, Threshold The average probability value of the above pixels. Iterate through all candidate thresholds and select the one that makes... The largest As the final threshold.

[0108] Optionally, for each pixel in the target land class channel of the initial segmentation probability map, if its probability value is greater than or equal to a threshold... If the probability value is less than the threshold, it is determined to be a pixel of the target land class and marked as the target land class label in the semantic segmentation result image; If a pixel is not found to be a background pixel, it is labeled as a background category label. After thresholding, a single-channel semantic segmentation result image is generated. This image has the same size as the adapted input image, and the pixel values ​​are discrete category labels, clearly distinguishing the target land class from the background region.

[0109] In one embodiment, such as Figure 3 As shown, after the UAV completes its predetermined mapping flight, based on the cached area results and geographic boundary information, spatial fusion and deduplication are performed on all mapping images to generate a continuous distribution map of target land types for the entire mapping area. Based on this continuous distribution map, the total area of ​​target land types in the mapping area is calculated, including:

[0110] S301. Based on the bounding boxes and geographic coordinates of the area results and geographic boundary information, project the corresponding semantic segmentation result maps of all survey images onto a unified global geographic coordinate system to obtain the projection results.

[0111] The schematic diagram illustrates that the geographic boundary information associated with the area results includes the bounding box coordinates and the geodetic coordinates of the four corner points for each surveying image. The bounding box coordinates define the spatial extent of the image in the original acquisition coordinate system, while the geodetic coordinates of the four corner points are calculated based on the real-time position information recorded by the UAV flight control system and camera extrinsic parameters, representing the absolute position of the image in the geodetic coordinate system. The global geographic coordinate system adopts the nationally unified geodetic datum framework. According to the geographical location and extent of the surveying area, the corresponding projection type is preset, and the projection parameters all conform to the national geodetic surveying standards and are pre-stored in the system's projection parameter library.

[0112] Specifically, the projection transformation process uses each semantic segmentation result image as a processing unit, extracts the geographic boundary information of the image, and converts the geodetic coordinates of the four corner points into planar coordinates in the global geographic coordinate system through a coordinate transformation model, establishing a mapping relationship between the original image coordinates and the global planar coordinates. The coordinate transformation model is as follows: ,in, , These are planar coordinates in the global geographic coordinate system. Longitude, Latitude of the earth , These are projection transformation functions from longitude and latitude to planar coordinates, respectively, the specific form of which is determined by the preset projection type. Based on the projection results of the four corner points, a bilinear interpolation algorithm is used to construct the coordinate mapping relationship of the entire semantic segmentation result image. For each pixel in the segmentation result image, its planar coordinates in the global geographic coordinate system are calculated through the mapping relationship, realizing the coordinate projection of the entire image. After projection, all pixels in each semantic segmentation result image correspond to unique planar coordinates in the global coordinate system, forming a projection result with a unified spatial reference.

[0113] S302. Based on the projection results, for any two overlapping mapping images corresponding to semantic segmentation result maps, within the overlapping area, a weighted fusion is performed based on the credibility of the semantic segmentation results of each pixel in different mapping images to obtain a global probability map; the credibility is determined by the accuracy level and input resolution of the semantic segmentation model associated with the semantic segmentation result map.

[0114] Furthermore, the determination of overlapping regions is based on spatial topological analysis of the bounding boxes of the projected semantic segmentation result images. By calculating the spatial intersection of the bounding boxes of any two images, the extent of the overlapping region is determined, and the set of pixels within the intersection is the object to be weighted fusion. The reliability calculation employs a two-factor weighted model, comprehensively considering the impact of the semantic segmentation model's accuracy level and the input resolution on the reliability of the segmentation results. ,in, For the first The credibility weight of each image This is a balancing coefficient, with a value range of [0,1]. It is used to adjust the weight ratio between the model accuracy level and the input resolution, and is determined through optimization using a validation dataset. For the first The accuracy weights of the semantic segmentation model used for each image are assigned according to the model's preset accuracy level, namely high accuracy, balanced, and lightweight. The higher the accuracy level, the larger the weight value. For the first The weights of the image input resolution are obtained by normalizing the input resolution to the [0,1] interval. The higher the resolution, the larger the weight value.

[0115] The weighted fusion process calculates the weighted average of the segmentation probabilities of each pixel within the overlapping region across all images covering that pixel. This average is used as the probability value of that pixel in the global probability map. The fusion formula is as follows: ,in, Coordinates in the global probability graph The probability value at that location. The number of images covering this pixel, For the first The semantic segmentation probability value of this pixel in the image. For the first The credibility weights of the images are determined. For pixels in non-overlapping regions, their segmentation probability values ​​in the corresponding images are directly used as the pixel values ​​in the global probability map. Through the fusion strategy, the segmentation results of overlapping regions can fully utilize the complementary information of multiple images, reduce the impact of segmentation errors in a single image, and improve the reliability and accuracy of the global probability map.

[0116] S303. Threshold segmentation is performed on the global probability map to obtain a binary global target land class distribution map, and the global target land class distribution map is transformed into a continuous distribution map of target land classes through a connected component analysis algorithm.

[0117] Threshold segmentation employs an adaptive threshold determination method, calculating the optimal segmentation threshold based on the probability distribution characteristics of the global probability graph. This threshold maximizes the separation between the target land class and the background. The threshold calculation employs the maximum inter-class variance method. By statistically analyzing the probability histogram of the global probability map, all possible candidate thresholds are traversed, the inter-class variance is calculated, and the candidate threshold with the largest inter-class variance is selected as the optimal threshold. Similarly, the formula for calculating inter-class variance is: ,in, For inter-class variance, For candidate thresholds, For probability values ​​greater than or equal to The percentage of pixels (the percentage of target land type). The probability value is less than The percentage of pixels (background percentage). For probability values ​​greater than or equal to The average probability value of pixels, The probability value is less than The average probability value of pixels.

[0118] Based on the optimal threshold Binarization is performed; specifically, in the global probability graph, probability values ​​greater than or equal to... The pixel is marked as the target location type, with a pixel value of 1 and a probability value less than 1. Pixels are marked as background with a value of 0 to generate a binarized global target land cover distribution map. Connectivity analysis uses the 8-neighborhood connectivity criterion, traversing the binarized distribution map pixel by pixel, marking each set of interconnected target land cover pixels as a connected component. For each connected component, a polygon fitting algorithm is used to extract its boundary contour. The Douglas-Peucker algorithm is then used to simplify the boundary contour, removing redundant vertices and filtering out small connected components with areas smaller than a preset minimum threshold. The simplified boundary contours of all valid connected components are combined to form a continuous target land cover distribution map covering the entire survey area. This map is stored in vector polygon dataset format, with each vector polygon corresponding to a complete target land cover region, and the boundaries are continuous and non-overlapping.

[0119] S304. Based on the geographic coordinates of each polygon in the continuous distribution map of the target land type, use the area calculation formula under the global geographic coordinate system to calculate the area of ​​each polygon and sum the areas to obtain the total area of ​​the target land type.

[0120] Furthermore, before area calculation, it is ensured that the vertex coordinates of each polygon in the continuous distribution map of the target land type are planar coordinates in the global geographic coordinate system. If the global geographic coordinate system is a latitude and longitude coordinate system, the vertex latitude and longitude coordinates are first converted to planar coordinates using the aforementioned projection transformation model to ensure coordinate consistency. The area of ​​each polygon is calculated using the shoelace formula, which calculates the area of ​​the closed figure formed by the vertices by traversing the vertex coordinates of the polygon. ,in, This represents the area of ​​a single polygon. This represents the number of vertices of the polygon. For the first The planar coordinates of the vertices, That is, the last vertex coincides with the first vertex, ensuring that the shape is closed.

[0121] The area accumulation process involves sequentially traversing all polygons in the continuous distribution map of the target land types, calculating the area of ​​each polygon one by one, and summing the results to obtain the total area of ​​the target land types in the surveyed area. ,in, The total area of ​​the target land type, The total number of polygons in the continuous distribution map of the target land type. For the first The area of ​​each polygon is calculated. During the accumulation process, spatial topology verification is used to ensure that there are no overlapping areas among the polygons, avoiding duplicate area calculations. The final output total area result includes an accuracy evaluation index for the area calculation, such as the relative error range. This index is calculated based on projection accuracy, segmentation accuracy, and fusion error, comprehensively reflecting the reliability of the total area result.

[0122] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0123] Based on the same inventive concept, this application also provides a geodetic survey area measuring device for implementing the geodetic survey area measuring method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations in one or more geodetic survey area measuring device embodiments provided below can be found in the limitations of the geodetic survey area measuring method described above, and will not be repeated here.

[0124] In one exemplary embodiment, such as Figure 4 As shown, a geodetic survey area measuring device is provided, comprising:

[0125] The status module 401 is used to collect the computing resource status of the airborne edge computing device, the aircraft status, and the currently captured mapping images during the UAV flight mapping process in real time, and to preprocess the mapping images and evaluate the scene complexity, and generate a comprehensive status vector by combining the computing resource status and the aircraft status.

[0126] The resource allocation module 402 is used to input the comprehensive state vector into the pre-trained deep reinforcement learning decision model to obtain the inference configuration action; the deep reinforcement learning decision model is deployed on the airborne edge computing device.

[0127] The segmentation module 403 is used to crop and / or scale the preprocessed mapping image according to the input resolution specified by the inference configuration action, generate an adapted input image, and dynamically load the semantic segmentation model corresponding to the inference configuration action to perform forward inference on the adapted input image to obtain the semantic segmentation result map of the target land class.

[0128] Analysis module 404 is used to calculate the area of ​​the target land class within the current mapping image based on the georeferenced information and semantic segmentation result map of the mapping image, and to generate and cache the area result and geographic boundary information.

[0129] The measurement results module 405 is used to perform spatial fusion and deduplication processing on all survey images based on the cached area results and geographic boundary information after the UAV completes the survey flight along the predetermined route, generate a continuous distribution map of target land types in the entire survey area, and calculate the total area of ​​target land types in the survey area based on the continuous distribution map of target land types.

[0130] In one embodiment, the resource allocation module 402 is further configured to:

[0131] The integrated state vector is analyzed to obtain several normalized parameters, including CPU utilization, GPU utilization, chip temperature, remaining power estimate, ground speed, and scene complexity.

[0132] Multiple normalized parameters are concatenated into a one-dimensional state tensor, and the policy network of the deep reinforcement learning decision model is used to propagate the one-dimensional state tensor forward to obtain the action value corresponding to each inference configuration action in the pre-constructed inference configuration action space; the action value is the expected long-term cumulative reward for selecting each inference configuration action in the current drone state.

[0133] The inference configuration action with the highest action value is selected as the output; the inference configuration action includes the selected semantic segmentation model index, input resolution, and post-processing switch flag.

[0134] In one embodiment, the status module 401 is further configured to:

[0135] Radiometric and geometric distortion corrections are performed on the survey images to obtain corrected images. Based on coarse georeferenced information, the corrected images are scaled to a uniform reference size to obtain preprocessed survey images.

[0136] The preprocessed mapping images are input into a pre-trained mini convolutional neural network to obtain scene category probability distribution vectors.

[0137] The gray-level co-occurrence matrix texture entropy is calculated from the preprocessed survey image, and the scene complexity vector is obtained by combining the scene category probability distribution vector;

[0138] Feature extraction is performed on the computational resource state and the aircraft state respectively to obtain the resource state vector and the flight state vector;

[0139] A comprehensive state vector is constructed by fusing the scenario complexity vector, resource state vector, and flight state vector.

[0140] In one embodiment, the segmentation module 403 is further configured to:

[0141] Perform bilinear interpolation scaling and / or center cropping on the preprocessed mapping image to obtain an adapted input image that matches the input resolution;

[0142] Load the semantic segmentation model with the corresponding semantic segmentation model index from the multi-precision model library;

[0143] The adapted input image is fed into the semantic segmentation model for forward inference to obtain the initial segmentation probability map;

[0144] If the post-processing switch flag is enabled, the initial segmentation probability map is optimized by the conditional random field algorithm to obtain the semantic segmentation result map.

[0145] If the post-processing switch flag is off, the initial segmentation probability map is thresholded to obtain the semantic segmentation result map.

[0146] In one embodiment, the measurement result module 405 is further configured to:

[0147] Based on the bounding boxes and geographic coordinates of the area results and geographic boundary information, the corresponding semantic segmentation result maps of all survey images are projected onto a unified global geographic coordinate system to obtain the projection results.

[0148] Based on the projection results, for any two overlapping mapping images corresponding to semantic segmentation result maps, a weighted fusion is performed within the overlapping area based on the credibility of the semantic segmentation results of each pixel in different mapping images to obtain a global probability map; the credibility is determined by the accuracy level and input resolution of the semantic segmentation model associated with the semantic segmentation result map.

[0149] Threshold segmentation is performed on the global probability map to obtain a binarized global target land class distribution map, and the global target land class distribution map is transformed into a continuous distribution map of target land classes through a connected component analysis algorithm.

[0150] Based on the geographic coordinates of each polygon in the continuous distribution map of the target land type, the area is calculated for each polygon using the area calculation formula in the global geographic coordinate system, and the areas are summed to obtain the total area of ​​the target land type.

[0151] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps in the above method embodiments.

[0152] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0153] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0154] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for determining the area of ​​a geodetic survey, characterized in that, The method includes: During the real-time acquisition of UAV flight mapping, the computing resource status of the airborne edge computing device, the aircraft status, and the currently captured mapping images are collected. The mapping images are preprocessed and the scene complexity is evaluated. A comprehensive state vector is generated by combining the computing resource status and the aircraft status. The integrated state vector is input into a pre-trained deep reinforcement learning decision model to obtain inference configuration actions; the deep reinforcement learning decision model is deployed on the airborne edge computing device. Based on the input resolution specified by the inference configuration action, the preprocessed mapping image is cropped and / or scaled to generate an adapted input image, and the semantic segmentation model corresponding to the inference configuration action is dynamically loaded to perform forward inference on the adapted input image to obtain the semantic segmentation result map of the target land class. Based on the geographic reference information of the surveying and mapping image and the semantic segmentation result map, calculate the area of ​​the target land type in the current surveying and mapping image, and generate and cache the area result and geographic boundary information; After the UAV completes the mapping flight along the predetermined route, based on the cached area results and the geographic boundary information, spatial fusion and deduplication processing are performed on all mapping images to generate a continuous distribution map of target land types for the entire mapping area, and the total area of ​​target land types in the mapping area is calculated based on the continuous distribution map of target land types.

2. The method according to claim 1, characterized in that, The deep reinforcement learning decision model obtains the inference configuration action through the following method: The comprehensive state vector is analyzed to obtain multiple normalized parameters; the normalized parameters include CPU utilization, GPU utilization, chip temperature, remaining power estimate, ground speed, and scene complexity. The normalized parameters are concatenated into a one-dimensional state tensor, and the policy network of the deep reinforcement learning decision model is used to forward propagate the one-dimensional state tensor to obtain the action value corresponding to each inference configuration action in the pre-constructed inference configuration action space; the action value is the expected long-term cumulative reward for selecting each inference configuration action in the current UAV state. The inference configuration action with the highest action value is selected as the output; the inference configuration action includes the selected semantic segmentation model index, input resolution, and post-processing switch flag.

3. The method according to claim 1, characterized in that, The preprocessing and scene complexity assessment of the mapping imagery, combined with the computing resource status and the aircraft status to generate a comprehensive state vector, includes: Radiometric and geometric distortion corrections are performed on the survey image to obtain a corrected image. The corrected image is then scaled to a uniform reference size based on coarse georeferenced information to obtain the preprocessed survey image. The preprocessed mapping image is input into a pre-trained micro convolutional neural network to obtain a scene category probability distribution vector. The gray-level co-occurrence matrix texture entropy is calculated from the preprocessed mapping image, and the scene complexity vector is obtained by combining the scene category probability distribution vector. Feature extraction is performed on the computational resource state and the aircraft state respectively to obtain resource state vector and flight state vector; The comprehensive state vector is constructed by fusing the scenario complexity vector, the resource state vector, and the flight state vector.

4. The method according to claim 2, characterized in that, The step involves cropping and / or scaling the preprocessed mapping image according to the input resolution specified by the inference configuration action to generate an adapted input image, and dynamically loading the semantic segmentation model corresponding to the inference configuration action to perform forward inference on the adapted input image to obtain a semantic segmentation result map of the target land class, including: The preprocessed mapping image is subjected to bilinear interpolation scaling and / or center cropping to obtain the adapted input image that conforms to the input resolution; Load the semantic segmentation model corresponding to the semantic segmentation model index from the multi-precision model library; The adapted input image is input into the semantic segmentation model for forward inference to obtain an initial segmentation probability map; If the post-processing switch flag is enabled, the initial segmentation probability map is optimized by the conditional random field algorithm to obtain the semantic segmentation result map. If the post-processing switch flag is off, the initial segmentation probability map is thresholded to obtain the semantic segmentation result map.

5. The method according to claim 2, characterized in that, After the UAV completes its predetermined mapping flight, based on the cached area results and geographic boundary information, spatial fusion and deduplication processing are performed on all mapping images to generate a continuous distribution map of target land types for the entire mapping area. The total area of ​​target land types in the mapping area is then calculated based on this continuous distribution map, including: Based on the area result and the bounding box and geographic coordinates of the geographic boundary information, the corresponding semantic segmentation result maps of all survey images are projected onto a unified global geographic coordinate system to obtain the projection result; Based on the projection results, for any two overlapping mapping images corresponding to semantic segmentation result maps, within the overlapping area, a weighted fusion is performed based on the credibility of the semantic segmentation results of each pixel in different mapping images to obtain a global probability map; the credibility is determined by the accuracy level of the semantic segmentation model associated with the semantic segmentation result map and the input resolution. The global probability map is segmented by threshold to obtain a binarized global target land class distribution map, and the global target land class distribution map is transformed into a continuous distribution map of target land classes by a connected component analysis algorithm; Based on the geographic coordinates of each polygon in the continuous distribution map of the target land type, the area is calculated for each polygon using the area calculation formula in the global geographic coordinate system, and the areas are accumulated to obtain the total area of ​​the target land type.

6. A device for measuring the area of ​​a land survey, characterized in that, The device includes: The status module is used to collect the computing resource status of the airborne edge computing device, the aircraft status, and the currently captured mapping images in real time during the UAV flight mapping process. It also performs preprocessing and scene complexity assessment on the mapping images and generates a comprehensive status vector by combining the computing resource status and the aircraft status. The resource allocation module is used to input the comprehensive state vector into a pre-trained deep reinforcement learning decision model to obtain inference configuration actions; the deep reinforcement learning decision model is deployed on the airborne edge computing device. The segmentation module is used to crop and / or scale the preprocessed mapping image according to the input resolution specified by the inference configuration action, generate an adapted input image, and dynamically load the semantic segmentation model corresponding to the inference configuration action to perform forward inference on the adapted input image to obtain the semantic segmentation result map of the target land class. The analysis module is used to calculate the area of ​​the target land class within the current surveying image based on the geographic reference information of the surveying image and the semantic segmentation result map, and to generate and cache the area result and geographic boundary information; The measurement results module is used to perform spatial fusion and deduplication processing on all survey images based on the cached area results and geographic boundary information after the UAV completes the survey flight along the predetermined route, generate a continuous distribution map of target land types for the entire survey area, and calculate the total area of ​​target land types in the survey area based on the continuous distribution map of target land types.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 5.

Images