3D ground surface modeling method and system for farmland scene

Abstract

The application discloses a 3D GS modeling method and system for farmland scenes, and belongs to the technical field of virtual modeling. The method comprises the following steps: acquiring an RGB image set of a target farmland carrying camera pose information, performing feature extraction and semantic segmentation on the RGB image set, and generating an initial sparse point cloud and a target mask graph; a multi-view voting mechanism is constructed, semantic information of the target mask graph is mapped into the initial sparse point cloud, a covering label is given to each three-dimensional point, a sparse point cloud set is generated, and the sparse point cloud set is divided into a plurality of semantic subsets; a layered Gaussian initialization strategy is configured, Gaussian merging is performed on different semantic subsets, a layered initialized 3D Gaussian set is generated, a control threshold and an optimization weight are configured, regional parallel optimization is performed, and a layered farmland model is generated; a monitoring area is divided for the target farmland, local supplementary collection is performed, a change area is identified and locally optimized, and thus the layered farmland model is locally updated.

Description

Technical Field

[0001] This invention belongs to the field of virtual modeling technology, and relates to a 3DGS modeling method and system for farmland scenes. Background Technology

[0002] Digital agriculture and smart farmland management systems are being rapidly implemented on a large scale. High-precision 3D reconstruction of farmland serves as the fundamental spatial data carrier supporting dynamic monitoring of crop growth, quantitative yield prediction, field disaster identification, and autonomous operation of intelligent agricultural machinery. 3D Gaussian Splatting (3DGS), relying on photorealistic rendering effects and real-time visualization capabilities, has already seen initial adoption in the field of 3D digital reconstruction of agriculture and forestry.

[0003] A patent application with publication number CN120689517A discloses a method and apparatus for 3D tree modeling based on 3DGS. The method includes: for the 3D tree modeling task, determining that the modeling components include the trunk, branches, and leaves; acquiring tree-view image data collected by a UAV; performing camera pose estimation and obtaining sparse point clouds through sparse reconstruction; inputting a 3D Gaussian splash 3DGS model to generate a 3D tree point cloud; selecting the corresponding model according to the segmentation object to generate 2D masks from different perspectives; wherein, extracting the complete tree selects the arbitrary segmentation model SAM model, and extracting tree components selects the semantic segmentation network MFAANet model based on the encoding and decoding structure; generating corresponding 2D masks on the input images from different perspectives; and generating the final point cloud classification result based on the multi-view masks obtained from segmentation through a multi-view mask voting mechanism.

[0004] While existing technologies can achieve high-fidelity 3D reconstruction of individual trees or small forest patches, meeting the basic needs of forestry resource surveys, they cannot adapt to the differentiated reconstruction requirements of different features within farmland scenarios. The use of a uniform global Gaussian density control rule throughout the process results in low overall model optimization efficiency. Specifically, it is impossible to implement differentiated reconstruction control for different semantic features within farmland. The globally uniform Gaussian density parameter generates a large number of useless Gaussian ellipsoids in large areas of homogeneous soil, causing a serious waste of GPU memory resources and a significant increase in model training cycle. Furthermore, core areas such as crop canopies, which require high detail reproduction, are constrained by a globally fixed density threshold, making it difficult to generate sufficient Gaussian units. This makes it difficult to reproduce the outlines, surface textures, and spatial arrangement details of crop leaves, significantly reducing the imaging accuracy and field operational adaptability of farmland 3D models. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a 3DGS modeling method and system for farmland scenarios. It achieves differentiated control throughout the entire process through semantic tags, adopts hierarchical Gaussian initialization and regional parallel optimization to resolve the contradiction between soil redundancy Gaussian and insufficient crop details, and combines incremental updates to improve the practicality of the model.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] 3DGS modeling methods for farmland scenes include:

[0008] Obtain an RGB image set of the target farmland equipped with camera pose information, perform feature extraction and semantic segmentation on the RGB image set, and generate an initial sparse point cloud and target mask map;

[0009] A multi-view voting mechanism is constructed to map the semantic information of the target mask image to the initial sparse point cloud, assign a cover label to each 3D point, generate a sparse point cloud set, and divide the sparse point cloud set into multiple semantic subsets based on the cover label;

[0010] Using the configured hierarchical Gaussian initialization strategy, Gaussian merging is performed on different semantic subsets to generate a hierarchically initialized 3D Gaussian set. Control thresholds and optimization weights are configured to perform regional parallel optimization and generate a hierarchical farmland model.

[0011] The target farmland is divided into monitoring areas, and supplementary data is collected locally within these areas. Using an incremental update mechanism, the changes in the farmland are identified and locally optimized by comparing the supplementary images with the corresponding viewpoint images of the layered farmland model at the pixel level, thereby locally updating the layered farmland model.

[0012] Specifically, the steps for generating the initial sparse point cloud include:

[0013] Based on a single image in the RGB image set, the ORB algorithm is used to extract corner feature points and their corresponding binary feature descriptors from the image.

[0014] The FLANN nearest neighbor search algorithm is used to perform initial feature matching on feature descriptors between images from different viewpoints to generate a matching set.

[0015] For any pair of matching points in the matching set, the triangulation algorithm is used to calculate the coordinates of the three-dimensional spatial points corresponding to the matching points, and a three-dimensional point set of the target farmland is generated.

[0016] Using the random sampling consensus algorithm, the ground plane equation is fitted based on the three-dimensional point set;

[0017] The three-dimensional point set and all camera extrinsic matrices are used as variables to be optimized. Projection error term and distance error term are set to construct the optimization objective function.

[0018] The optimization objective function is minimized using the Levenberg-Marquardt algorithm to obtain the set of optimization points;

[0019] Based on the statistical outlier removal method, outlier filtering is performed on the optimized point cloud to obtain the initial sparse point cloud.

[0020] Specifically, the steps for semantic segmentation of the images in the RGB image set include:

[0021] Based on the encoder-decoder architecture and the MobileNetV2 backbone network, a semantic segmentation network is constructed, including an input layer, an encoding layer, a bottleneck layer, a decoding layer, and an output layer.

[0022] The semantic segmentation network is trained by constructing a segmentation dataset through manual annotation.

[0023] Images from the RGB image set are sequentially input into the semantic segmentation network to obtain the predicted coverage category of each pixel. An initial mask image is generated by combining the numerical identifier configured for each coverage category.

[0024] For the initial mask image with the coverage category of crop plants, a crop binary image is constructed. The Hough transform is used to detect the straight line segments in the crop binary image, the crop row straight lines are filtered, and the crop row constraint is corrected.

[0025] Connectivity analysis was performed on the soil background in the initial mask image to filter out isolated soil domains.

[0026] The cover category with the highest proportion in the eight neighboring areas surrounding the isolated soil domain is used to replace the cover category in the isolated soil domain.

[0027] Morphological opening and closing operations are used to process the mask of each coverage category. By merging the category mask rows, the target mask map of the corresponding image is generated.

[0028] Specifically, the multi-perspective voting mechanism includes:

[0029] For any three-dimensional point in the initial sparse point cloud Using a pinhole camera projection, three-dimensional points are projected onto the first... From the plane of the image, obtain the corresponding pixel coordinates. ;in, , This represents the number of images in the RGB image set.

[0030] Get the The effective range of the image, and the pixel coordinates Conduct effective verification;

[0031] Calculate 3D points in response to valid coordinates. To the The distance from the camera's optical center to the image, and the distance from the camera's optical center to the 3D point. Given the ray direction, traverse all relevant 3D points on the ray and calculate the distance from each relevant 3D point to the camera optical center;

[0032] 3D points are filtered based on viewpoint occlusion. Effective observation perspectives are identified, and a list of observation perspectives is constructed.

[0033] Specifically, the multi-perspective voting mechanism also includes:

[0034] For a three-dimensional point For each observation viewpoint, obtain the corresponding camera's optical axis direction vector and 3D point. The surface normal vector is used to calculate the viewpoint weight by performing a cosine operation on the angle between the vectors.

[0035] Obtain 3D points The number of successful feature point matchings in the corresponding image is used to calculate the feature matching weight through logarithmic operations.

[0036] The observation weight of the observation viewpoint is calculated by multiplying the viewpoint weight by the feature matching weight.

[0037] For a three-dimensional point All observation weights are normalized, and the normalized observation weights are then filled into the observation view list.

[0038] For a three-dimensional point For each observation viewpoint, the coverage category is read from the corresponding target mask image based on the pixel coordinates obtained from the projection;

[0039] Based on the coverage category, for 3D points The observation perspectives are classified, and within the same coverage category, the observation weights are summed to serve as the weights of the coverage category.

[0040] The coverage type corresponding to the maximum weight is used as the 3D point. The overlay tag.

[0041] Specifically, the hierarchical Gaussian initialization strategy includes:

[0042] Based on the coverage labels of three-dimensional points, the sparse point cloud is divided into a crop plant subset, a soil background subset, an irrigation canal subset, a field ridge subset, and a farmland facility subset.

[0043] The K-means spatial clustering algorithm is used to perform cluster analysis on the soil background subset, generating multiple spatially continuous clusters;

[0044] Merge the 3D points in each cluster into an initial 3D Gaussian and initialize the position and color parameters;

[0045] For the crop plant subset, each 3D point is initialized as an initial 3D Gaussian, and its position, shape, color, and opacity are also initialized.

[0046] Cluster analysis was performed on the subsets of irrigation canals, field ridges, and farmland facilities. Each cluster was merged into an initial 3D Gaussian, and its position, shape, color, and opacity were initialized.

[0047] Integrate the initial Gaussians of all semantic subsets to generate a hierarchically initialized 3D Gaussian set.

[0048] Specifically, the steps for constructing a stratified farmland model include:

[0049] Configure independent Gaussian density control parameters for each semantic subset;

[0050] The importance level of the semantic subsets is reconstructed and assigned a value. The values ​​of each semantic subset are then normalized, and the optimized weight of the corresponding semantic subset is calculated.

[0051] The 3DGS differentiable renderer renders based on the Gaussian set of the current semantic subset to generate a rendered image.

[0052] Based on the pixel mean square error between the rendered image and the corresponding original RGB image, the reprojection loss of the semantic subset is obtained;

[0053] The loss term for the corresponding semantic subset is calculated based on the product of the optimized weights and the reprojection loss.

[0054] Based on the GPU's hardware specifications, a computational flow with the same number of semantic subsets is created, and the Gaussian data, density control parameters, and corresponding reprojection loss of the semantic subsets are bound together.

[0055] Specifically, the steps for constructing a stratified farmland model also include:

[0056] All computational flows are triggered in parallel for optimization loops, including: differentiable rendering, loss calculation, gradient backpropagation, parameter update, and density control.

[0057] After each computational flow completes one iteration, an intermediate Gaussian set of the corresponding semantic subset is generated;

[0058] For any semantic subset, calculate the loss reduction rate of the last 5 iterations;

[0059] If the loss reduction rate is less than a preset convergence threshold, the semantic subset is marked as converged; otherwise, it is marked as non-converged.

[0060] Configure a resource scheduling mechanism, traverse the convergence state of each semantic subset, allocate the minimum computing resources to the converged subset, and distribute the remaining computing resources evenly to the non-converged subset.

[0061] The optimization loop and resource scheduling are iteratively calculated until all semantic subsets reach a convergent state or reach the preset maximum number of iterations, thereby generating the optimized Gaussian set of each semantic subset.

[0062] By merging the optimized Gaussian set of all semantic subsets, a hierarchical farmland model is constructed.

[0063] Specifically, the incremental update mechanism includes:

[0064] Based on local images of the monitored area, a local image set is constructed, and a baseline image consistent with the local images is rendered based on a hierarchical farmland model.

[0065] Calculate the pixel-level grayscale difference between the local image and the reference image to generate a difference heatmap;

[0066] A change threshold is set based on the gray standard deviation of unchanged areas in historical change data.

[0067] Pixels with a grayscale difference greater than the change threshold are marked as changed pixels. A two-dimensional change region is constructed using a region growing algorithm. The two-dimensional change region is then projected onto a three-dimensional space using camera extrinsic parameters to construct the change region.

[0068] Extract all Gaussians from the varying regions of the stratified farmland model, and use them as a local Gaussian set;

[0069] Based on the local image of the changing region, a local sparse point cloud is generated, and the coverage label is assigned, hierarchical Gaussian initialization is performed, and regional parallel optimization is performed to generate a local update set adapted to the changing region.

[0070] The local Gaussian set is replaced with the local update set, and the transformation of the changed region is smoothed.

[0071] A 3DGS modeling system for farmland scenarios includes: a data acquisition module, a point cloud processing module, and a Gaussian optimization module;

[0072] The data acquisition module uses a drone or a ground mobile platform to cruise in different locations and orientations of the target farmland to acquire RGB image sets. In response to the identification of the monitoring area, it performs local supplementary acquisition to obtain supplementary images of the monitoring area.

[0073] The point cloud processing module uses the RGB image set to construct an initial sparse point cloud. At the same time, it performs semantic segmentation on the images in the RGB image set to construct a target mask image, and maps the semantic information in the target mask image to the initial sparse point cloud in the form of labels to generate a sparse point cloud set.

[0074] The Gaussian optimization module is used to perform Gaussian merging on the sparse point cloud to generate a 3D Gaussian set. By configuring the control threshold and optimization weight, it performs regional parallel optimization to construct a layered farmland model, configures the monitoring area of ​​the target farmland, and optimizes the layout of the Gaussian set based on the pixel-level comparison between the supplementary image and the layered farmland model.

[0075] The beneficial effects of this invention are:

[0076] By generating initial point clouds optimized for farmland scenes and using high-precision semantic annotation, a mapping from pixels to 3D points to semantic categories was established, providing a semantic foundation for subsequent differentiated control throughout the entire process. A hierarchical Gaussian initialization strategy based on coverage labels was adopted, configuring differentiated point-Gaussian transformation rules for the reconstruction needs of different farmland elements. This precisely controls the quantity and spatial distribution of Gaussians from the initialization stage, avoiding the generation of a large number of invalid redundant Gaussians in large areas of uniform soil, while reserving sufficient Gaussian expression capacity for the crop canopy that requires fine representation. Through semantically aware density control and a regional parallel optimization mechanism, computing resources are prioritized for crop areas that are most important for agricultural production decisions, improving global optimization efficiency and significantly reducing memory usage and training time. Furthermore, through semantically differentiated model lightweighting, the model size is effectively compressed without sacrificing the accuracy of crop area reconstruction. A multi-source data-driven incremental update mechanism is constructed, and through intelligent identification and monitoring of monitoring areas and local change detection, Gaussian optimization and replacement are performed only on areas where changes have actually occurred. This can promptly reflect the dynamic changes in farmland caused by crop growth, agricultural operations, and meteorological disasters, significantly improving the overall quality and practical application value of the farmland 3D model. Attached Figure Description

[0077] Figure 1 A schematic diagram of a 3DGS modeling method for farmland scenes;

[0078] Figure 2 This is a flowchart of the multi-perspective voting mechanism in this invention;

[0079] Figure 3 This is a flowchart of the process for generating a stratified farmland model in this invention;

[0080] Figure 4 This is a structural diagram of a 3DGS modeling system for farmland scenarios. Detailed Implementation

[0081] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0082] Example 1

[0083] refer to Figures 1 to 3 As shown, this embodiment introduces a 3DGS modeling method for farmland scenes, including the following steps:

[0084] Step S1: Based on the continuous acquisition of RGB image sequences at different locations and orientations of the target farmland by UAV or ground mobile platform, the absolute three-dimensional coordinates (longitude, latitude, altitude) of the acquisition device are obtained simultaneously through GNSS receiver as positioning data, and the three-axis attitude angles (roll angle, pitch angle, yaw angle) of the acquisition device are obtained through IMU inertial measurement unit as attitude data. Through visual-inertial fusion algorithm, the positioning data, attitude data and visual feature points in the image are fused to calculate the camera extrinsic parameter matrix corresponding to each image. The calculated camera extrinsic parameter matrix is ​​then attached to the corresponding RGB image as a pose label, thereby constructing an RGB image set with pose label for the target farmland. Feature extraction and matching are then performed on the RGB image set with pose label to generate an initial sparse point cloud.

[0085] Step S2: Perform semantic segmentation on the images in the RGB image set to identify the coverage type, including crop plants, soil background, irrigation ditches, field ridges, and farmland facilities, thereby generating a target mask map. Through a multi-view voting mechanism, the coverage information in the target mask map of each view is mapped to the initial sparse point cloud according to the corresponding pose label, thereby assigning a coverage label to each 3D point in the initial sparse point cloud, thus generating a sparse point cloud set with coverage labels.

[0086] Step S3: Based on different coverage labels in the sparse point cloud, Gaussian merging is performed on some 3D points using the configured hierarchical Gaussian initialization strategy to generate a hierarchically initialized 3D Gaussian set. This avoids generating a large number of redundant Gaussians in large soil areas. For each semantic subset in the 3D Gaussian set, control thresholds and optimization weights are configured. Regional parallel optimization is used to bind different semantic subsets to independent computational flows of the GPU for parameter optimization. The pixel difference between the rendered image and the RGB image set is minimized through the gradient descent algorithm to avoid resource competition during global optimization and generate a hierarchical farmland model.

[0087] Step S4: Based on the area in the target farmland covered with crop plants, configure the monitoring area in the target farmland. Through local supplementary acquisition of the monitoring area, configure the incremental update mechanism to compare the supplementary image with the corresponding view image of the layered farmland model at the pixel level, identify the changed area and optimize the local Gaussian set of the changed area. Replace all Gaussian parameters in the changed area of ​​the original layered farmland model with the new local Gaussian set, while keeping the Gaussian parameters of the remaining areas unchanged, thus completing the local update of the layered farmland model.

[0088] Furthermore, the steps for generating the initial sparse point cloud include:

[0089] Based on a single image in the RGB image set, the ORB algorithm is used to extract corner feature points and corresponding binary feature descriptors from the image, such as setting the FAST corner detection threshold to 20, the number of pyramid layers to 8, and the maximum number of feature points per layer to 1000. This generates a set of feature point coordinates and a set of feature descriptors for each image.

[0090] The FLANN nearest neighbor search algorithm is used, such as constructing a KD tree index, setting the number of trees to 4 and the number of search checks to 50, to perform initial feature matching on the feature descriptors between images from different viewpoints, generating a matching set, where each matching point pair contains corresponding feature points from two images from different viewpoints;

[0091] For any pair of cross-view matching points in the matching set, the triangulation algorithm is used to calculate the three-dimensional spatial coordinates of the matching points by combining the camera extrinsic matrix and the pre-calibrated camera intrinsic matrix corresponding to the two images, thereby generating a three-dimensional point set of the target farmland.

[0092] Using the random sample consensus algorithm, the farmland ground plane is fitted from a three-dimensional point set to obtain the ground plane equation, as follows: ,in, It is a plane normal vector. The distance from the plane to the origin. , , These are the x-coordinate, y-coordinate, and vertical coordinate of any three-dimensional point in the world coordinate system;

[0093] The 3D point set and all camera extrinsic matrices are used as variables to be optimized. The sum of the squared differences between the pixel coordinates projected onto the corresponding image and the actual feature point coordinates of each 3D point is used as the projection error term. The sum of the squared distances from all 3D points less than the ground distance threshold to the ground plane is used as the distance error term. The objective function is constructed by summing the projection error term and the distance error term. The ground distance threshold is determined based on the typical topographic relief features of farmland and the measurement accuracy of the initial point cloud. For example, it is set to 0.2m to filter out 3D points belonging to the ground area, while avoiding misclassifying crop plant points as ground.

[0094] Using the Levenberg-Marquardt algorithm, with an initial damping factor of 0.001, a damping factor adjustment coefficient of 10, and a convergence threshold of 1e-6, the objective function is minimized, and the coordinates of all 3D points and the extrinsic matrix of all cameras are optimized simultaneously to obtain the optimized point cloud.

[0095] Based on a statistical outlier removal method, outlier filtering is performed on the optimized point cloud, and the distance from each point to the target point is calculated. The average distance between the nearest neighbors is used to remove isolated points whose average distance exceeds twice the standard deviation of the global average distance, resulting in an initial sparse point cloud; among which, As a preset value, this embodiment takes... .

[0096] Furthermore, semantic segmentation of images in the RGB image set includes:

[0097] Based on the encoder-decoder architecture and the MobileNetV2 backbone network, a semantic segmentation network is constructed, including an input layer, an encoding layer, a bottleneck layer, a decoding layer, and an output layer.

[0098] The input layer receives RGB images and maps pixel values ​​to normalized values. interval;

[0099] The encoding layer consists of 5 downsamplers. Each downsampler contains a convolutional layer, a batch normalization layer, and an activation layer. The convolutional layer is used to extract multi-scale features from the normalized image and generate a feature map. The batch normalization layer is used to normalize the feature map, accelerate network convergence, and prevent overfitting. The activation layer is used to perform a non-linear transformation on the normalized feature map. Each downsampler reduces the resolution of the feature map to half of its original value while doubling the number of feature channels. Through 5 downsampling operations, features are gradually extracted from low-level edge and texture features to high-level semantic features, generating 5 sets of feature maps at different scales.

[0100] The bottleneck layer consists of three convolutional layers with dilation rates of 2, 4, and 8, respectively. Each convolutional layer captures global contextual information at different scales by expanding the receptive field of the convolutional kernel without downsampling or reducing the resolution of the feature map. The three convolutional layers with different dilation rates are connected in series to fuse global contextual features at multiple scales and generate a high-level feature map that fuses global semantic information.

[0101] The decoding layer consists of four upsamplers. Each upsampler includes an interpolation layer, a convolutional layer, a normalization layer, and an activation layer. The interpolation layer is used to double the resolution of the high-level feature map. The convolutional layer is used to perform feature fusion and dimensionality reduction on the upsampled high-level feature map. The normalization layer and the activation layer have the same function as the encoding layer. Through four upsampling operations, the feature map resolution is gradually restored, and finally a feature map with the same size as the input image is generated.

[0102] The output layer uses a 1×1 convolutional layer to map the number of channels of the feature map output by the decoding layer to the number of 5 preset coverage categories. The probability distribution of each pixel belonging to each coverage category is calculated by the Softmax activation function, and the coverage category with the highest probability is selected as the final predicted category by the argmax operation.

[0103] Based on RGB images of farmland from different regions, crop types, and growth stages, corresponding semantic label maps were generated through manual annotation to construct a segmentation dataset. The segmentation dataset was divided into a training set and a validation set. The Adam optimizer was used to train the network, setting the initial learning rate, batch size, and number of training epochs. For example, the initial learning rate was 0.001, the batch size was 16, and the number of training epochs was 50. A learning rate decay strategy was adopted, reducing the learning rate to 0.5 times the original value every 10 epochs. During training, the average intersection-union ratio (IUU) and pixel accuracy of the network were calculated on the validation set after each training epoch. Training was stopped early when the average IUU on the validation set no longer improved for 5 consecutive epochs.

[0104] Each image in the RGB image set is sequentially input into the trained semantic segmentation network to obtain the predicted coverage category of each pixel;

[0105] Configure a digital identifier for each coverage category. For example, 0 represents soil background, 1 represents crop plants, 2 represents soil background, 3 represents irrigation canal, 4 represents field ridge, and 5 represents farmland facilities. Based on the prediction result of each pixel, fill the corresponding digital identifier into a single-channel image with the same size as the original image to generate an initial mask image. Generate the corresponding initial mask image for each image in the RGB image set in turn.

[0106] The crop plants in the initial mask image are binarized to generate a crop binary image. Hough transform is used to detect straight line segments in the crop binary image, and the angle distribution of all straight line segments is statistically analyzed. The angle with the highest frequency is used as the main direction of the crop row, thus generating crop row lines. Crop row constraint correction is then performed, including: based on the detected crop row lines, expanding outwards by a preset half-width distance from each line to form continuous crop row regions. The regions between adjacent crop row regions are the crop row gap regions. Crop row region masks and crop row gap region masks are generated separately. For areas located within the crop row region mask and classified as soil background... If a pixel is classified as a crop plant if the percentage of its surrounding 3×3 neighborhood pixels of the crop plant category exceeds a neighborhood percentage threshold, it is corrected to the crop plant category. For pixels located within the crop row gap area mask that are classified as crop plants, if the percentage of their surrounding 3×3 neighborhood pixels of the soil background category exceeds a neighborhood percentage threshold, they are corrected to the soil background category. The neighborhood percentage threshold is determined by those skilled in the art based on the crop planting density and image resolution in order to correct misjudgments while avoiding over-correction. In this embodiment, it is set to 0.6. The crop row half-width is determined based on the typical planting row spacing in farmland. In this embodiment, it is taken as the image pixel width corresponding to 0.3m.

[0107] Connectivity analysis was performed on the soil background in the initial mask image. The 8-neighborhood connectivity criterion was used to identify all independent soil connected components, and the pixel area of ​​each soil connected component was calculated.

[0108] Soil connected regions with pixel areas smaller than the minimum connectivity threshold are selected as isolated soil regions. The minimum connectivity threshold is determined by those skilled in the art based on the image resolution and the minimum actual size of farmland soil patches to filter out isolated misjudgments while retaining real small soil areas. In this embodiment, it is set to 50 pixels.

[0109] The highest percentage of the coverage category in the 8 neighboring regions around an isolated soil domain is determined, and all pixels in the corresponding isolated soil domain are corrected to the highest percentage of the coverage category.

[0110] Morphological opening and closing operations are used to process the mask for each coverage category, i.e., 3×3 rectangular structural elements. First, morphological opening operations are performed on the mask for each coverage category, and then morphological closing operations are performed to eliminate small bright spots and burrs, and fill small holes and gaps. The processed masks of each category are merged to generate the target mask map of the corresponding image. The rectangular structural elements are determined based on the noise characteristics and boundary accuracy requirements of the farmland semantic mask.

[0111] Furthermore, the steps of a multi-perspective voting mechanism include:

[0112] For any three-dimensional point in the initial sparse point cloud Iterate through each image in the RGB image set, read the corresponding camera extrinsic matrix, and use a pinhole camera projection to project the 3D points onto the image. From the plane of the image, obtain the corresponding pixel coordinates. ;in, , This represents the number of images in the RGB image set.

[0113] Get the The effective range of the image, and the pixel coordinates To perform valid verification, i.e., to determine the pixel coordinates Whether it is within the valid range; where the valid range includes image width and image height;

[0114] In response to valid coordinates, i.e., pixel coordinates Within the effective range, acquire 3D points. Coordinate vectors in the world coordinate system, the first The coordinate vector of the camera optical center in the world coordinate system corresponding to each image is used to calculate the 3D point using the Euclidean norm. To the Camera optical center distance of the image and along the camera's optical center to the three-dimensional point Given the ray direction, iterate through all other 3D points (3D points) along the current ray direction. Except for), filter out projections to the first Construct a set of related 3D points by considering all 3D points with the same pixel coordinate in an image. Calculate each relevant 3D point in the relevant 3D point set Distance to the optical center of the camera And perform viewpoint occlusion detection; among which, For three-dimensional points The number of relevant three-dimensional points, ;

[0115] If any relevant three-dimensional point exists Distance to the optical center of the camera Less than and and In the If the coordinate difference on the images is less than the preset effective observation threshold, it indicates that the three-dimensional point... pixel coordinates Unobstructed, judged as number one The viewpoint corresponding to each image is a three-dimensional point. The effective observation angle; wherein, the effective observation threshold is set based on the image resolution and the projection accuracy of the initial point cloud, such as in this embodiment where the effective observation threshold is set to 1 pixel;

[0116] Integrating 3D points Construct a list of all valid observation perspectives. Simultaneously, a corresponding target mask image is associated with each viewpoint; where, For three-dimensional points The number of corresponding effective observation angles;

[0117] For a three-dimensional point Each observation perspective Obtain the optical axis direction vector and 3D point of the corresponding camera. The surface normal vector is used to calculate the dot product of the optical axis direction vector and the surface normal vector, as well as the product of the Euclidean norm of the optical axis direction vector and the Euclidean norm of the surface normal vector. The viewpoint weight is then calculated using the ratio of the dot product to the product, thus obtaining the 3D point. The number of successful feature point matches in the corresponding image, i.e., the number of 3D points in the matching set. The frequency of occurrences in the corresponding image is used to calculate the feature matching weight by performing a logarithmic operation on the number of successful matches. The observation viewpoint is then calculated by multiplying the viewpoint weight by the feature matching weight. The observation weights;

[0118] For example, three-dimensional points Let be a feature point on the corn canopy, whose surface normal vector is a unit vector. At this point, the image is vertically upward, aligning with the crop's growth direction. Viewpoint 1 is taken from directly below the drone at a 30° angle, with the camera's optical axis as a unit vector. This point participated in 5 successful feature matches from this perspective. Viewpoint 2 was taken from the side of the drone at a 60° angle, with the camera's optical axis direction as a unit vector. This point participated in three successful feature matches from this perspective, and the logarithmic operation used the natural logarithm. ;

[0119] Then the view weight of viewpoint 1 at this time Feature matching weights Then the observation weight ;

[0120] Similarly, the viewpoint weights of viewpoint 2 Feature matching weights Then the observation weight ;

[0121] For a three-dimensional point All observation weights are normalized so that the sum of all weights is 1, and the normalized observation weights are filled into the observation view list.

[0122] For a three-dimensional point Each observation perspective Based on the pixel coordinates obtained from the projection, the coverage category is read from the corresponding target mask image. ;

[0123] Based on the coverage category, for 3D points The observation perspectives are classified. Within the same coverage category, the observation weights are summed to form the weight of the coverage category. The coverage type corresponding to the maximum weight is used as the 3D point. The covered labels have a maximum weight equal to the confidence level.

[0124] For example, three-dimensional points There are 5 valid observation perspectives. The normalized weights and labeled coverage categories for each perspective are as follows: the observation weight for perspective 1 is 0.25 and the coverage category is 1 (crop plant); the observation weight for perspective 2 is 0.30 and the coverage category is 1 (crop plant); the observation weight for perspective 3 is 0.20 and the coverage category is 2 (soil background); the observation weight for perspective 4 is 0.15 and the coverage category is 1 (crop plant); and the observation weight for perspective 5 is 0.10 and the coverage category is 1 (crop plant).

[0125] The total weights obtained for each cover category are calculated. The total weight for crop plants is 0.25 + 0.30 + 0.15 + 0.1 = 0.8, the total weight for soil background is 0.2, and the total weight for other cover categories is 0. At this point, the 3D points... The coverage label is 1 (crop plant) with a confidence level of 0.80.

[0126] Furthermore, the steps of the hierarchical Gaussian initialization strategy include:

[0127] Based on the coverage label of each 3D point in the sparse point cloud, the sparse point cloud is divided into 5 mutually independent semantic subsets, including crop plant subset, soil background subset, irrigation canal subset, field ridge subset, and farmland facility subset.

[0128] Since the soil background is a large, uniform area, no fine geometric representation is required. The K-means spatial clustering algorithm is used to perform cluster analysis on all three-dimensional points in the soil background subset. After the cluster analysis is completed, multiple spatially continuous clusters are generated. Each cluster contains several three-dimensional points with the same label whose spatial distance is less than the clustering distance threshold. The clustering distance threshold is set according to the typical size of farmland soil patches and the reconstruction accuracy requirements, and is determined through statistical analysis of the distribution characteristics of farmland soil in different regions of the country.

[0129] All 3D points in each cluster are merged into an initial 3D Gaussian. The cluster center is used as the position parameter of the Gaussian, and the average RGB color value of all 3D points in the cluster is used as the color parameter of the Gaussian. Based on the visual characteristics of the soil area and the 3DGS rendering principle, the initial opacity of the Gaussian is set, such as to 0.3, to ensure a natural transition effect in large areas of soil. No adaptive initialization of shape parameters is performed. The shape parameters (such as a 3×3 covariance matrix) are fixed as isotropic matrices, with all diagonal elements set to preset fixed values ​​and all off-diagonal elements set to 0.

[0130] For the crop plant subset, in order to provide sufficient Gaussian representation, each 3D point is independently initialized as an initial 3D Gaussian, which initializes the position, shape, color, and opacity. This includes using the spatial coordinates of the 3D point as the position parameter, the RGB color corresponding to the 3D point as the color parameter, setting the initial opacity parameter, and initializing the shape parameter as an anisotropic covariance matrix stretched along the vertical direction, where the vertical direction is the Z-axis direction of the world coordinate system. The element value corresponding to the Z-axis in the diagonal elements of the matrix is ​​twice the value of the element value corresponding to the X-axis and Y-axis, and the off-diagonal elements are all 0.

[0131] For irrigation ditches, field ridges, and farmland features, continuity and integrity must be maintained. For the 3D points in the subsets of irrigation ditches, field ridges, and farmland facilities, cluster analysis is performed. At the same time, based on the typical width of the corresponding features and the reconstruction accuracy requirements, the cluster distance thresholds corresponding to the subsets of irrigation ditches, field ridges, and farmland facilities are set, which are less than the cluster distance threshold of the soil background subset. Each cluster is merged into an initial 3D Gaussian, and its position, shape, color, and opacity are initialized. This includes: using the geometric center of the cluster as the position parameter of the Gaussian, using the average RGB color value of all 3D points in the cluster as the color parameter of the Gaussian, setting the initial opacity of the Gaussian to 0.2, extracting the principal direction of the cluster point cloud through principal component analysis as the linear structure direction, initializing the covariance matrix as an anisotropic matrix stretched along the linear structure direction as the initial shape parameter, and setting the diagonal element value corresponding to the principal direction to be 3 times the element value corresponding to the other two directions, while the off-diagonal elements are all 0.

[0132] Integrate the initial Gaussians of each semantic subset, and construct semantic Gaussian subsets for each semantic subset to generate a hierarchically initialized 3D Gaussian set.

[0133] Furthermore, the steps for generating a stratified farmland model include:

[0134] Independent Gaussian density control parameters are configured for each semantic subset. For the soil background subset, Gaussian cloning and splitting operations are disabled, and only Gaussian deletion operations are retained. A maximum upper limit for the number of Gaussians is set, and the density adjustment of the subset is stopped immediately after the upper limit is reached. For the crop plant subset, full Gaussian cloning, splitting, and deletion operations are enabled, and the density threshold is set to 5 times that of the soil subset to improve the sensitivity of Gaussian splitting. When the opacity gradient of the Gaussians exceeds the splitting trigger threshold, splitting is triggered immediately. For the irrigation canal, field ridge, and farmland facility subsets, only Gaussian splitting and deletion operations are enabled, cloning operations are disabled, and a medium density threshold is set, which is between that of the soil and crop subsets. The splitting threshold is set based on the structural characteristics of the crop canopy and the requirements for fine reconstruction. For example, in this embodiment, it is set to 0.02, which can accurately detect areas with insufficient detail expression in the crop leaf edges, veins, and overlapping areas of leaves, and promptly trigger splitting to generate more Gaussians to restore the morphological and textural features of the leaves, while avoiding the generation of unnecessary redundant Gaussians.

[0135] The reconstruction importance of semantic subsets is assigned values, and the values ​​of each semantic subset are normalized to calculate the optimization weight of the corresponding semantic subset. For example, since the reconstruction accuracy of the crop canopy needs to be guaranteed, the crop plant subset is assigned the highest value, set to 1. In order to ensure the reconstruction accuracy of key infrastructure such as irrigation equipment and power facilities, the farmland facility subset is assigned the second highest value, set to 0.6. In order to ensure the continuity of the irrigation system and the observability of water level changes, the irrigation canal subset is assigned a medium value, set to 0.5. Since it is only necessary to ensure that the basic boundaries and direction of the field ridges are clear, the field ridge subset is assigned a low value, set to 0.4. At the same time, since it is only necessary to ensure that the overall color and texture of a large area of ​​soil are consistent, the soil background subset is assigned the lowest value, set to 0.2.

[0136] Using the 3DGS differentiable renderer, the Gaussian set of the current semantic subset is rendered into a rendered image with the same viewpoint as the original RGB image. The pixel mean square error between the rendered image and the corresponding original RGB image is calculated as the reprojection loss of the corresponding semantic subset.

[0137] The loss term for the corresponding semantic subset is calculated based on the product of the optimized weights and the reprojection loss.

[0138] Based on the GPU's hardware specifications, create computational flows that are the same number as the semantic subset, with each computational flow having its own instruction queue and execution context;

[0139] The Gaussian data, density control parameters, and corresponding reprojection loss of each semantic subset are bound to a computational stream. At the same time, independent GPU memory space is allocated to each computational stream to store the Gaussian data, rendering intermediate results, and gradient data of the computational stream, so as to avoid memory access conflicts between different subsets.

[0140] All computational streams are triggered in parallel for optimization loops. Using the 3DGS differentiable renderer, the Gaussian data of the corresponding semantic subset in the computational stream is rendered into a rendered image with the same viewpoint as the original RGB image, and the corresponding loss value is calculated. This includes: obtaining the rendered image of the semantic subset and the original RGB image of the corresponding viewpoint, calculating the absolute difference of the color values ​​at corresponding positions of the two images pixel by pixel, taking the average of the absolute differences of all pixels to obtain the L1 loss value, and multiplying the L1 loss value by the pre-calculated normalized optimization weight of the semantic subset to obtain the semantic weighted loss value of the semantic subset in this iteration.

[0141] The gradient of all Gaussian parameters relative to the loss value is calculated using automatic differentiation. The position, shape, color and opacity parameters of the Gaussian are updated using the Adam optimizer. The cloning, splitting and deletion operations of the Gaussian are performed according to the density control parameters bound to the corresponding semantic Gaussian subset. After each iteration of the computation flow is completed, an intermediate Gaussian set of the corresponding semantic subset is generated.

[0142] For any semantic subset, the loss reduction rate of the last 5 iterations is calculated. At the same time, based on the statistical analysis of historical loss data, a convergence threshold is set. The convergence threshold is dynamically set as the average loss reduction rate of the previous 10 iterations for the corresponding semantic subset. The convergence threshold is set as an adjustment ratio of the average reduction rate. The crop plant subset uses the minimum adjustment ratio, such as 0.01-0.03, to meet the requirements of fine reconstruction of leaf morphology and texture. The farmland facilities and irrigation canal subset uses the medium adjustment ratio, such as 0.04-0.06, to balance structural integrity and optimization efficiency. The field ridges and soil background subset uses the maximum adjustment ratio, such as 0.07-0.1, to ensure the overall consistency of large-area uniform regions.

[0143] If the rate of decrease in loss is less than the convergence threshold, the semantic subset is marked as converged; if the rate of decrease in loss is not less than the convergence threshold, the semantic subset is marked as non-converged.

[0144] Configure a resource scheduling mechanism, traverse the convergence status of each semantic subset, and for converged subsets, retain only the minimum computing resources to maintain parameter updates, integrate all non-converged subsets, and distribute the remaining computing resources evenly to the non-converged subsets.

[0145] The optimization loop and resource scheduling are iteratively calculated until all semantic subsets reach a convergent state or reach the preset maximum number of iterations, thereby generating the optimized Gaussian set of each semantic subset.

[0146] The optimized Gaussian set of all semantic subsets is merged to generate a hierarchical farmland model.

[0147] Furthermore, the incremental update mechanism includes the following steps:

[0148] Based on local images of the monitored area, a local image set is constructed, and a baseline image consistent with the local images is rendered based on a hierarchical farmland model.

[0149] Calculate the pixel-level grayscale difference between the local image and the reference image to generate a difference heatmap;

[0150] Historical change data is acquired, and areas without change are filtered out. The gray standard deviation of the areas without change is calculated. A change threshold is set at three times the gray standard deviation. Pixels with gray differences greater than the change threshold are marked as changed pixels. A continuous two-dimensional change region is constructed through a region growing algorithm. The two-dimensional change region is projected into three-dimensional space through camera extrinsic parameters to form the change region that needs to be updated.

[0151] Extract all Gaussians from the varying regions of the stratified farmland model, and use them as a local Gaussian set;

[0152] Based on the local image of the changing region, a local sparse point cloud is generated, and the coverage label is assigned, hierarchical Gaussian initialization is performed, and regional parallel optimization is performed to generate a local update set adapted to the changing region.

[0153] The original local Gaussian set is replaced with a locally updated set, and the transformation of the changed region is smoothed to eliminate the splicing traces between the old and new Gaussian sets, so as to achieve local update of the hierarchical farmland set.

[0154] Example 2

[0155] Please see Figure 4 Another embodiment of the present invention provides a 3DGS modeling system for farmland scenes, comprising: a data acquisition module, a point cloud processing module, and a Gaussian optimization module;

[0156] The data acquisition module uses drones or ground mobile platforms to cruise in different locations and orientations of the target farmland, continuously acquiring RGB image sequences, positioning data, and attitude data. It uses visual-inertial fusion to calculate the camera extrinsic matrix of each image and uses the camera extrinsic matrix as pose label to construct an RGB image set for the target farmland. At the same time, it responds to local supplementary acquisition of the monitoring area to obtain supplementary images of the monitoring area.

[0157] The point cloud processing module uses the RGB image set to construct an initial sparse point cloud. At the same time, it performs semantic segmentation on the images in the RGB image set to construct a target mask image. The semantic information in the target mask image is then mapped to the initial sparse point cloud in the form of labels to generate a sparse point cloud set.

[0158] The Gaussian optimization module is used to perform Gaussian merging on sparse point clouds to generate a 3D Gaussian set. It then performs parallel optimization in different regions by configuring control thresholds and optimization weights to build a layered farmland model. At the same time, it configures the monitoring area of ​​the target farmland and optimizes the layout of the Gaussian set based on the pixel-level comparison between the supplementary image and the layered farmland model, thus completing the local update of the layered farmland model.

[0159] Working principle and effects:

[0160] By optimizing semantic segmentation and multi-view voting in farmland scenes, each 3D point is assigned an accurate coverage label, establishing a precise mapping between pixels, 3D points, and semantics. A hierarchical Gaussian initialization strategy is adopted, configuring differentiated point-Gaussian transformation rules for different farmland elements. This avoids generating a large amount of redundant Gaussian in the soil area while reserving sufficient expressive capacity for the crop canopy. Through semantically aware density control and regional parallel optimization, computing resources are prioritized for crop areas, improving optimization efficiency and reducing memory usage. Through semantically differentiated lightweighting, the model size is compressed without sacrificing crop accuracy, enabling real-time rendering. A multi-source driven incremental update mechanism is constructed, performing local optimization and replacement only on changed areas, solving the update problem of full retraining of traditional models and significantly improving the practicality of farmland 3D models.

[0161] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A 3DGS modeling method for farmland scenes, characterized in that, include: Obtain an RGB image set of the target farmland equipped with camera pose information, perform feature extraction and semantic segmentation on the RGB image set, and generate an initial sparse point cloud and target mask map; A multi-view voting mechanism is constructed to map the semantic information of the target mask image to the initial sparse point cloud, assign a cover label to each 3D point, generate a sparse point cloud set, and divide the sparse point cloud set into multiple semantic subsets based on the cover label; Using the configured hierarchical Gaussian initialization strategy, Gaussian merging is performed on different semantic subsets to generate a hierarchically initialized 3D Gaussian set. Control thresholds and optimization weights are configured to perform regional parallel optimization and generate a hierarchical farmland model. The target farmland is divided into monitoring areas, and supplementary data is collected locally within these areas. Using an incremental update mechanism, the changes in the farmland are identified and locally optimized by comparing the supplementary images with the corresponding viewpoint images of the layered farmland model at the pixel level, thereby locally updating the layered farmland model.

2. The 3DGS modeling method for farmland scenes according to claim 1, characterized in that, The steps to generate an initial sparse point cloud include: Based on a single image in the RGB image set, the ORB algorithm is used to extract corner feature points and their corresponding binary feature descriptors from the image. The FLANN nearest neighbor search algorithm is used to perform initial feature matching on feature descriptors between images from different viewpoints to generate a matching set. For any pair of matching points in the matching set, the triangulation algorithm is used to calculate the coordinates of the three-dimensional spatial points corresponding to the matching points, and a three-dimensional point set of the target farmland is generated. Using the random sampling consensus algorithm, the ground plane equation is fitted based on the three-dimensional point set; The three-dimensional point set and all camera extrinsic matrices are used as variables to be optimized. Projection error term and distance error term are set to construct the optimization objective function. The optimization objective function is minimized using the Levenberg-Marquardt algorithm to obtain the set of optimization points; Based on the statistical outlier removal method, outlier filtering is performed on the optimized point cloud to obtain the initial sparse point cloud.

3. The 3DGS modeling method for farmland scenes according to claim 2, characterized in that, The steps for semantic segmentation include: Based on the encoder-decoder architecture and the MobileNetV2 backbone network, a semantic segmentation network is constructed, including an input layer, an encoding layer, a bottleneck layer, a decoding layer, and an output layer. The semantic segmentation network is trained by constructing a segmentation dataset through manual annotation. Images from the RGB image set are sequentially input into the semantic segmentation network to obtain the predicted coverage category of each pixel. An initial mask image is generated by combining the numerical identifier configured for each coverage category. For the initial mask image with the coverage category of crop plants, a crop binary image is constructed. The Hough transform is used to detect the straight line segments in the crop binary image, the crop row straight lines are filtered, and the crop row constraint is corrected. Connectivity analysis was performed on the soil background in the initial mask image to filter out isolated soil domains. The cover category with the highest proportion in the eight neighboring areas surrounding the isolated soil domain is used to replace the cover category in the isolated soil domain. Morphological opening and closing operations are used to process the mask of each coverage category. By merging the category mask rows, the target mask map of the corresponding image is generated.

4. The 3DGS modeling method for farmland scenes according to claim 3, characterized in that, The multi-perspective voting mechanism includes: For any three-dimensional point in the initial sparse point cloud Using a pinhole camera projection, three-dimensional points are projected onto the first... From the plane of the image, obtain the corresponding pixel coordinates. ;in, , This represents the number of images in the RGB image set. Get the The effective range of the image, and the pixel coordinates Conduct effective verification; Calculate 3D points in response to valid coordinates. To the The distance from the camera's optical center to the image, and the distance from the camera's optical center to the 3D point. Given the ray direction, traverse all relevant 3D points on the ray and calculate the distance from each relevant 3D point to the camera optical center; 3D points are filtered based on viewpoint occlusion. Effective observation perspectives are identified, and a list of observation perspectives is constructed.

5. The 3DGS modeling method for farmland scenes according to claim 4, characterized in that, The multi-perspective voting mechanism also includes: For a three-dimensional point For each observation viewpoint, obtain the corresponding camera's optical axis direction vector and 3D point. The surface normal vector is used to calculate the viewpoint weight by performing a cosine operation on the angle between the vectors. Obtain 3D points The number of successful feature point matchings in the corresponding image is used to calculate the feature matching weight through logarithmic operations. The observation weight of the observation viewpoint is calculated by multiplying the viewpoint weight by the feature matching weight. For a three-dimensional point All observation weights are normalized, and the normalized observation weights are then filled into the observation view list. For a three-dimensional point For each observation viewpoint, the coverage category is read from the corresponding target mask image based on the pixel coordinates obtained from the projection; Based on the coverage category, for 3D points The observation perspectives are classified, and within the same coverage category, the observation weights are summed to serve as the weights of the coverage category. The coverage type corresponding to the maximum weight is used as the 3D point. The overlay tag.

6. The 3DGS modeling method for farmland scenes according to claim 5, characterized in that, The hierarchical Gaussian initialization strategy includes: Based on the coverage labels of three-dimensional points, the sparse point cloud is divided into a crop plant subset, a soil background subset, an irrigation canal subset, a field ridge subset, and a farmland facility subset. The K-means spatial clustering algorithm is used to perform cluster analysis on the soil background subset, generating multiple spatially continuous clusters; Merge the 3D points in each cluster into an initial 3D Gaussian and initialize the position and color parameters; For the crop plant subset, each 3D point is initialized as an initial 3D Gaussian, and its position, shape, color, and opacity are also initialized. Cluster analysis was performed on the subsets of irrigation canals, field ridges, and farmland facilities. Each cluster was merged into an initial 3D Gaussian, and its position, shape, color, and opacity were initialized. Integrate the initial Gaussians of all semantic subsets to generate a hierarchically initialized 3D Gaussian set.

7. The 3DGS modeling method for farmland scenes according to claim 6, characterized in that, The steps to generate a stratified farmland model include: Configure independent Gaussian density control parameters for each semantic subset; The importance level of the semantic subsets is reconstructed and assigned a value. The values ​​of each semantic subset are then normalized, and the optimized weight of the corresponding semantic subset is calculated. The 3DGS differentiable renderer renders based on the Gaussian set of the current semantic subset to generate a rendered image. Based on the pixel mean square error between the rendered image and the corresponding original RGB image, the reprojection loss of the semantic subset is obtained; The loss term for the corresponding semantic subset is calculated based on the product of the optimized weights and the reprojection loss. Based on the GPU's hardware specifications, a computational flow with the same number of semantic subsets is created, and the Gaussian data, density control parameters, and corresponding reprojection loss of the semantic subsets are bound together.

8. The 3DGS modeling method for farmland scenes according to claim 7, characterized in that, The steps for generating a stratified farmland model also include: All computational flows are triggered in parallel for optimization loops, including: differentiable rendering, loss calculation, gradient backpropagation, parameter update, and density control. After each computational flow completes one iteration, an intermediate Gaussian set of the corresponding semantic subset is generated; For any semantic subset, calculate the loss reduction rate of the last 5 iterations; If the loss reduction rate is less than a preset convergence threshold, the semantic subset is marked as converged; otherwise, it is marked as non-converged. Configure a resource scheduling mechanism, traverse the convergence state of each semantic subset, allocate the minimum computing resources to the converged subset, and distribute the remaining computing resources evenly to the non-converged subset. The optimization loop and resource scheduling are iteratively calculated until all semantic subsets reach a convergent state or reach the preset maximum number of iterations, thereby generating the optimized Gaussian set of each semantic subset. By merging the optimized Gaussian set of all semantic subsets, a hierarchical farmland model is constructed.

9. The 3DGS modeling method for farmland scenes according to claim 8, characterized in that, The incremental update mechanism includes: Based on local images of the monitored area, a local image set is constructed, and a baseline image consistent with the local images is rendered based on a hierarchical farmland model. Calculate the pixel-level grayscale difference between the local image and the reference image to generate a difference heatmap; A change threshold is set based on the gray standard deviation of unchanged areas in historical change data. Pixels with a grayscale difference greater than the change threshold are marked as changed pixels. A two-dimensional change region is constructed using a region growing algorithm. The two-dimensional change region is then projected onto a three-dimensional space using camera extrinsic parameters to construct the change region. Extract all Gaussians from the varying regions of the stratified farmland model, and use them as a local Gaussian set; Based on the local image of the changing region, a local sparse point cloud is generated, and the coverage label is assigned, hierarchical Gaussian initialization is performed, and regional parallel optimization is performed to generate a local update set adapted to the changing region. The local Gaussian set is replaced with the local update set, and the transformation of the changed region is smoothed.

10. A 3DGS modeling system for farmland scenes, used to implement the 3DGS modeling method for farmland scenes as described in any one of claims 1-9, characterized in that, include: Data acquisition module, point cloud processing module, and Gaussian optimization module; The data acquisition module uses a drone or a ground mobile platform to cruise in different locations and orientations of the target farmland to acquire RGB image sets. In response to the identification of the monitoring area, it performs local supplementary acquisition to obtain supplementary images of the monitoring area. The point cloud processing module uses the RGB image set to construct an initial sparse point cloud. At the same time, it performs semantic segmentation on the images in the RGB image set to construct a target mask image, and maps the semantic information in the target mask image to the initial sparse point cloud in the form of labels to generate a sparse point cloud set. The Gaussian optimization module is used to perform Gaussian merging on the sparse point cloud to generate a 3D Gaussian set. By configuring the control threshold and optimization weight, it performs regional parallel optimization to construct a layered farmland model, configures the monitoring area of ​​the target farmland, and optimizes the layout of the Gaussian set based on the pixel-level comparison between the supplementary image and the layered farmland model.

Images