Cultivated land boundary identification method and system

Abstract

The invention provides a cultivated land boundary identification method and system, and the method comprises the steps: collecting a remote sensing image of a target cultivated land, carrying out the cultivated land semantic segmentation of the remote sensing image, obtaining an initial cultivated land probability graph, and determining a color-texture gradient magnitude graph of the remote sensing image; the category probability gradient of each pixel in multiple directions in the cultivated land probability graph and the color-texture gradient magnitude graph of the remote sensing image are initialized, and a boundary significance graph is constructed; according to the semantic probability gradient magnitude and the color-texture gradient magnitude of each pixel point in the boundary saliency map, determining the semantic-geometric feature mismatch degree of each pixel point, and performing adaptive filtering on the boundary saliency map through the semantic-geometric feature mismatch degrees of all the pixel points; and generating a cultivated land boundary vector of the target cultivated land based on the filtered boundary saliency map. By adopting the scheme of the invention, the dislocation of the semantic boundary and the visual edge in the cultivated land boundary image recognition process can be quantified.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of boundary recognition, in particular, the present application relates to a cultivated land boundary recognition method and system. BACKGROUND

[0002] Boundary recognition is a basic task in the field of computer vision and image processing, and its core goal is to automatically detect the boundary line between different regions or objects in a digital image. The essence is to find a set of pixel points where the image properties change dramatically. These changes can be reflected in color / brightness, texture, depth, and semantic category. The result of boundary recognition is usually a set of edge pixels or a contour line composed of these pixels.

[0003] Cultivated land boundary recognition refers to automatically and accurately extracting the external contour line of each independent cultivated land unit from remote sensing images. Traditional cultivated land boundary recognition mainly includes methods based on traditional image processing and methods based on deep learning semantic segmentation. In actual situations, the semantic boundary predicted by the deep learning model (such as the position where the cultivated land probability changes abruptly) may not completely coincide with the real visual edge in the image (such as the color / texture changes caused by dikes and roads). This misalignment can cause the vector boundary to deviate from the true position or be undecided between the two, resulting in noise. Therefore, how to quantify the misalignment between semantic boundaries and visual edges in the cultivated land boundary image recognition process to improve the geometric accuracy of cultivated land boundary vectors has become a difficult problem in the industry. SUMMARY

[0004] Based on this, the present application provides a cultivated land boundary recognition method and system for quantifying the misalignment between semantic boundaries and visual edges in the cultivated land boundary image recognition process.

[0005] In a first aspect, the present application provides a cultivated land boundary recognition method, comprising the following steps: Collecting a remote sensing image of a target cultivated land; Performing cultivated land semantic segmentation on the remote sensing image to obtain an initial cultivated land probability map; Determining a color-texture gradient amplitude map of the remote sensing image, and then constructing a boundary saliency map according to the class probability gradient of each pixel in the initial cultivated land probability map in multiple directions and the color-texture gradient amplitude map of the remote sensing image; Determining the semantic-geometric feature mismatch degree of each pixel point in the boundary saliency map according to the semantic probability gradient amplitude and the color-texture gradient amplitude of each pixel point, wherein the semantic-geometric feature mismatch degree is the absolute difference value between the class probability gradient direction and the color-texture gradient direction of each pixel point. Adaptively filter the boundary saliency map through semantic-geometric feature mismatch degrees of all pixel points, and then generate the cultivated land boundary vector of the target cultivated land based on the filtered boundary saliency map.

[0006] In some embodiments, the remote sensing image is subjected to cultivated land semantic segmentation to obtain an initial cultivated land probability map, specifically comprising: An initial cultivated land probability map is obtained by performing semantic segmentation on the remote sensing image. The remote sensing image is input into the cultivated land semantic segmentation model for forward inference to obtain an initial segmentation result. The initial segmentation result is subjected to probabilistic post-processing to obtain the initial cultivated land probability map.

[0007] In some embodiments, determining the color-texture gradient amplitude map of the remote sensing image specifically comprises: The color gradient amplitude and the texture gradient amplitude of each pixel in the remote sensing image are determined. The color-texture gradient amplitude map of the remote sensing image is determined according to the color gradient amplitude and the texture gradient amplitude of each pixel.

[0008] In some embodiments, the boundary saliency map is constructed according to the class probability gradient in multiple directions of each pixel in the initial cultivated land probability map and the color-texture gradient amplitude map of the remote sensing image, specifically comprising: The class probability gradient amplitude and the class probability gradient direction of each pixel in the initial cultivated land probability map are determined. The color-texture gradient direction of each pixel in the remote sensing image is determined. The gradient consistency weight is determined according to the class probability gradient direction of each pixel in the initial cultivated land probability map and the color-texture gradient direction of each pixel in the remote sensing image. The class probability gradient amplitude of each pixel in the initial cultivated land probability map and the corresponding color-texture gradient amplitude of each pixel in the color-texture gradient amplitude map are weighted and fused based on the gradient consistency weight to generate the boundary saliency map.

[0009] In some embodiments, the semantic-geometric feature mismatch degree of each pixel point in the boundary saliency map is determined according to the semantic probability gradient amplitude and the color-texture gradient amplitude of each pixel point, specifically comprising: The class probability gradient direction and the color-texture gradient direction of each pixel point are obtained. The absolute direction difference between the class probability gradient direction and the color-texture gradient direction of each pixel point is calculated. The absolute direction difference corresponding to each pixel point is normalized to obtain a semantic-geometric feature mismatch degree of each pixel point.

[0010] In some embodiments, adaptively filtering the boundary saliency map by the semantic-geometric feature mismatch degrees of all pixel points specifically comprises: determining an adaptive filtering parameter of a local region in the boundary saliency map based on the semantic-geometric feature mismatch degrees of all pixel points; performing filtering processing on the boundary saliency map according to the adaptive filtering parameter to obtain an enhanced boundary saliency map.

[0011] In some embodiments, generating the cultivated land boundary vector of the target cultivated land based on the filtered boundary saliency map specifically comprises: performing non-maximum suppression and edge connection on the filtered boundary saliency map to extract all candidate boundary pixels; converting all candidate boundary pixels into a vector boundary line, taking the semantic-geometric feature mismatch degree of the corresponding pixel point as a constraint, fusing broken line segments at a vector level, and removing isolated false boundaries to generate a final cultivated land boundary vector of the cultivated land target.

[0012] In a second aspect, the present application provides a cultivated land boundary recognition system, comprising: a collection module configured to collect a remote sensing image of a target cultivated land; a processing module configured to perform cultivated land semantic segmentation on the remote sensing image to obtain an initial cultivated land probability map; the processing module is further configured to determine a color-texture gradient amplitude map of the remote sensing image, and then construct a boundary saliency map according to the class probability gradient of each pixel in multiple directions in the initial cultivated land probability map and the color-texture gradient amplitude map of the remote sensing image; the processing module is further configured to determine a semantic-geometric feature mismatch degree of each pixel point in the boundary saliency map according to the semantic probability gradient amplitude and the color-texture gradient amplitude of each pixel point, wherein the semantic-geometric feature mismatch degree is an absolute difference value of the class probability gradient direction and the color-texture gradient direction of each pixel point; an execution module configured to adaptively filter the boundary saliency map by the semantic-geometric feature mismatch degrees of all pixel points, and then generate a cultivated land boundary vector of the target cultivated land based on the filtered boundary saliency map.

[0013] In a third aspect, the present application provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the cultivated land boundary recognition method described above when executing the computer program.

[0014] In a fourth aspect, the present application provides a computer readable storage medium, which stores a computer program. When the computer program is executed by a processor, the steps of the cultivated land boundary recognition method described above are implemented.

[0015] The technical scheme provided by the embodiments disclosed in the present application has the following beneficial effects: In the cultivated land boundary recognition method and system provided by the present application, firstly, the collected remote sensing image is subjected to cultivated land semantic segmentation to obtain an initial cultivated land probability map. The initial cultivated land probability map quantifies the possibility of each position in the remote sensing image belonging to the cultivated land category with a pixel-level probability value. This step can realize intelligent identification and quantification of the possibility of cultivated land existing at the pixel level from a complex remote sensing image, thereby avoiding the dependence of traditional methods on manual interpretation or simple threshold segmentation. Then, the color-texture gradient amplitude map of the remote sensing image is determined, and a boundary saliency map is constructed according to the class probability gradient of each pixel in the initial cultivated land probability map in multiple directions and the color-texture gradient amplitude map of the remote sensing image. This step can adaptively fuse the class probability gradient representing high-level semantics and the color-texture gradient representing low-level visual features, so as to simultaneously strengthen the real semantic boundary and suppress irrelevant texture responses in the cultivated land interior in a single map, thereby improving the comprehensive representation capability and integrity of the real cultivated land boundary. Secondly, the semantic-geometric feature mismatch degree of each pixel point is determined according to the semantic probability gradient amplitude and the color-texture gradient amplitude of each pixel point in the boundary saliency map. The semantic-geometric feature mismatch degree is the absolute difference value between the class probability gradient direction and the color-texture gradient direction of each pixel point. This step measures the consistency degree of the semantic judgment and the visual evidence in the boundary direction at each pixel point, so as to convert the abstract "feature conflict" into a specific and calculable mismatch degree map. Finally, the boundary saliency map is adaptively filtered through the semantic-geometric feature mismatch degree of all pixel points, and a cultivated land boundary vector of the target cultivated land is generated based on the filtered boundary saliency map. This step uses the semantic-geometric feature mismatch degree to guide the local adaptive intelligent filtering of the boundary map, so as to suppress high mismatch degree noise while retaining low mismatch degree real edges, thereby finally generating a cultivated land boundary vector with high geometric precision and correct topological relationship. In summary, the scheme of the present application can quantify the misalignment between the semantic boundary and the visual edge in the cultivated land boundary image recognition process, thereby improving the geometric precision of the cultivated land boundary vector. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is an exemplary flowchart of a cultivated land boundary recognition method according to some embodiments of the present application; Figure 2 is an application scenario schematic diagram of a cultivated land boundary recognition data processing system according to some embodiments of the present application; Figure 3 is a flowchart of determining a boundary saliency map according to some embodiments of the present application; Figure 4 is a structural schematic diagram of a cultivated land boundary identification system according to some embodiments of the present application; Figure 5 is a structural schematic diagram of a computer device for implementing a cultivated land boundary identification method according to some embodiments of the present application. DETAILED DESCRIPTION

[0017] In order to better understand the above technical solutions, the above technical solutions will be described in detail below in combination with the drawings in the specification and specific embodiments.

[0018] Reference Figure 1 , which is an exemplary flowchart of a cultivated land boundary identification method according to some embodiments of the present application, mainly includes the following steps: In step 101, a remote sensing image of a target cultivated land is collected.

[0019] In a specific implementation, the remote sensing image of the target cultivated land can be collected in the following manner, that is, first, a suitable data source can be selected according to the monitoring requirements, for example, multi-spectral satellite data can be used for large-scale survey, and unmanned aerial vehicle carrying multi-spectral sensor can be used for fine identification to collect centimeter-level images; then, a data collection process is started, and the original data packet of the specified area is downloaded through the satellite ground receiving station or the unmanned aerial vehicle is controlled to fly according to the preset flight route to collect, and in the collection process, it is ensured that the cloud coverage rate is less than 10% and the solar elevation angle is greater than 30 degrees to obtain the best imaging quality; then, the original data is standardized and preprocessed, the DN value is converted into apparent reflectivity by radiation calibration, atmospheric correction is performed by using 6S model, geometric precision correction is completed based on regional network adjustment method, finally, ortho-corrected image with accurate geographic coordinates is generated, and it is cut to the target cultivated land range, so that the remote sensing image of the target cultivated land is obtained, and in other embodiments, other methods can also be used for implementation, which are not limited in the present application.

[0020] In some embodiments, referring to Figure 2 , which is an application scenario schematic diagram of a cultivated land boundary identification data processing system according to some embodiments of the present application, the diagram includes three main components: a collection device, a server and a data storage device, the collection device is responsible for collecting the remote sensing image of the target cultivated land, and sending the collected remote sensing image to the server through the communication network, the cultivated land boundary identification data processing system is running in the server, the server stores the processing result in the data storage device, and performs visualization.

[0021] In step 102, the remote sensing image is subjected to cultivated land semantic segmentation to obtain an initial cultivated land probability map.

[0022] In some embodiments, the farmland semantic segmentation on the remote sensing image to obtain the initial farmland probability map can be implemented by the following steps, that is: obtaining a pre-trained farmland semantic segmentation model; inputting the remote sensing image into the farmland semantic segmentation model for forward inference to obtain an initial segmentation result; performing probabilistic post-processing on the initial segmentation result to obtain the initial farmland probability map.

[0023] It should be noted that the pre-training of the farmland semantic segmentation model in the present application can be implemented in the following manner, for example: first, a semantic segmentation model architecture suitable for remote sensing image processing can be selected, such as U-Net or DeepLabV3+, and a ResNet backbone network pre-trained on ImageNet can be selected for the encoder part to accelerate convergence; then, a public remote sensing farmland segmentation dataset is used to train the selected model, and the dataset is enhanced before training, including random rotation, flipping and color jittering, to improve the generalization ability of the model; during the training process, a cross-entropy loss function and an Adam optimizer are used, and a decreasing learning rate strategy is set, and the model parameters are converged through multiple iterations, and finally the trained model parameters are saved as a pre-trained model file for subsequent inference, and the pre-trained model has the ability to distinguish farmland and non-farmland at the pixel level from multi-band remote sensing images, and the non-farmland is, for example, buildings, forests, water bodies, etc. In other embodiments, other methods can also be used, which are not limited in the present application.

[0024] When specifically implemented, the inputting of the remote sensing image into the farmland semantic segmentation model for forward inference to obtain an initial segmentation result can be implemented in the following manner, for example: first, the pre-processed remote sensing image can be adjusted to the input size specified by the model, such as 512x512 pixels, and channel normalization processing is performed to make the pixel values of each band conform to the distribution set during model training; then, the processed image tensor is input into the pre-trained farmland semantic segmentation model to perform a complete forward propagation calculation, and the model finally outputs the original probability score of each pixel point belonging to the two categories of "farmland" and "non-farmland" through a Softmax layer; the output is a two-channel probability tensor with the same spatial size as the input image, and the two-channel probability tensor is the initial segmentation result, which completely retains the original confidence information of each position belonging to the farmland category, and other methods can also be used in other embodiments, which are not limited in the present application.

[0025] In a specific implementation, the probabilistic post-processing of the initial segmentation result to obtain the initial cultivated land probability map can be implemented in the following manner, for example: the probability value of the corresponding channel representing the "cultivated land" category is extracted from the double-channel probability tensor obtained by the forward reasoning; subsequently, Gaussian filtering is applied to smooth the corresponding single-channel probability map representing the "cultivated land" category to suppress isolated noise points and fine patches caused by model prediction uncertainty, while maintaining the continuity of the probability value of the main cultivated land region, wherein the Gaussian filtering uses a 3x3 convolution kernel and the standard deviation is set to 1; the probability map after smoothing is the final initial cultivated land probability map, wherein the value range of each pixel is [0, 1], and the higher the value, the greater the confidence that the pixel belongs to cultivated land. In other embodiments, other methods can also be used, which are not limited in the present application.

[0026] It should be noted that the initial cultivated land probability map in the present application quantifies the possibility of each position in the remote sensing image belonging to the cultivated land category with pixel-level probability values, which is a two-dimensional data layer. It serves as the semantic basis for the subsequent boundary identification process and converts abstract image information into a continuous cultivated land existence confidence field.

[0027] In step 103, the color-texture gradient amplitude map of the remote sensing image is determined, and then a boundary saliency map is constructed according to the class probability gradient of each pixel in multiple directions in the initial cultivated land probability map and the color-texture gradient amplitude map of the remote sensing image.

[0028] In some embodiments, the color-texture gradient amplitude map of the remote sensing image can be determined by the following steps, that is: determining the color gradient amplitude and the texture gradient amplitude of each pixel in the remote sensing image; determining the color-texture gradient amplitude map of the remote sensing image according to the color gradient amplitude and the texture gradient amplitude of each pixel.

[0029] In a specific implementation, the color gradient amplitude of each pixel in the remote sensing image can be determined in the following manner, for example: for the calculation of the color gradient amplitude of each pixel in the remote sensing image, the preprocessed RGB remote sensing image can be converted to the CIELAB color space to better perceive color differences, and then the Sobel operator is used to calculate the gradient components of the luminance L channel, the green-red component A channel and the blue-yellow component B channel in the horizontal and vertical directions, respectively. The Sobel operator uses a 3x3 convolution kernel, then the gradient components of the three channels are squared and the square roots are calculated for each pixel to obtain the gradient amplitude of each channel, and finally the Euclidean norm of the gradient amplitudes of the three channels is taken as the final color gradient amplitude of the corresponding pixel. In other embodiments, other methods can also be used, which are not limited in the present application.

[0030] In some embodiments, determining the texture gradient magnitude of each pixel in the remote sensing image specifically comprises: converting the remote sensing image into a gray scale image; filtering the gray scale image using a multi-directional Gabor filter bank to obtain texture response maps of the gray scale image in different directions; For each pixel, determining the texture gradient magnitude of each pixel according to the texture response values in different directions.

[0031] In specific implementation, the determination of the texture gradient magnitude of each pixel in the remote sensing image can be implemented in the following manner: first, a set of Gabor filters in different directions is constructed, with the direction parameters Typically covering typical angles of 0°, 45°, 90°, 135°, etc. Then, the gray scale image is convolved with each Gabor filter to obtain the texture response map in the corresponding direction ; then, for each pixel , the texture gradient magnitude thereof is calculated by vector synthesis, specifically: first, the comprehensive response vector is calculated, and then the texture gradient magnitude of the pixel is In other embodiments, other methods can also be used, which are not limited in the present application.

[0032] In specific implementation, the determination of the color-texture gradient magnitude map of the remote sensing image according to the color gradient magnitude and the texture gradient magnitude of each pixel can be implemented in the following manner, for example: the color gradient magnitude and the texture gradient magnitude of each pixel obtained in the foregoing steps can be linearly weighted and fused, wherein the color gradient magnitude is given a weight of 0.6 and the texture gradient magnitude is given a weight of 0.4, the color-texture gradient magnitude of the corresponding pixel is calculated by weighted summation of the color gradient magnitude and the texture gradient magnitude of each pixel; after the weighted fusion calculation of all pixels in the image is completed, the fusion magnitudes of all pixels are arranged according to the original image space position, i.e. the final color-texture gradient magnitude map is generated, and in other embodiments, other methods can also be used, which are not limited in the present application.

[0033] It should be noted that in this process, the horizontal and vertical gradient components generated in the calculation process need to be synchronously reserved for subsequent steps of calculating the gradient direction.

[0034] In some embodiments, with reference Figure 3As shown, the figure is a flowchart of determining the boundary saliency map in some embodiments of the present application, and the boundary saliency map can be constructed by using the following steps, i.e.: In step 1031, the gradient amplitude and the gradient direction of the class probability in multiple directions of each pixel in the initial cultivated land probability map are determined. In step 1032, the color-texture gradient direction of each pixel in the remote sensing image is determined. In step 1033, the gradient consistency weight is determined according to the gradient direction of the class probability in multiple directions of each pixel in the initial cultivated land probability map and the color-texture gradient direction of each pixel in the remote sensing image. In step 1034, the gradient amplitude of the class probability in multiple directions of each pixel in the initial cultivated land probability map and the color-texture gradient amplitude of each pixel in the color-texture gradient amplitude map are weighted and fused based on the gradient consistency weight to generate the boundary saliency map.

[0035] In specific implementation, the gradient amplitude and the gradient direction of the class probability in multiple directions of each pixel in the initial cultivated land probability map can be determined by using the following method, for example, the Sobel gradient operator can be used to perform convolution calculation on the initial cultivated land probability map in the horizontal direction and the vertical direction to obtain the gradient component in the horizontal direction and the gradient component in the vertical direction of each pixel; then, the gradient amplitude of the class probability of each pixel is calculated by squaring and square root operation of the gradient components in the horizontal direction and the vertical direction; at the same time, the gradient direction of the class probability of each pixel is calculated by using the arctan2 function, and the result is normalized to the radian range of 0 to 2π; for example, the horizontal gradient component of a certain pixel is 3 and the vertical gradient component is 4, then the gradient amplitude of the class probability of the pixel is √(3²+4²)=5, and the gradient direction of the class probability of the pixel is calculated by arctan2(4, 3), the result is about 0.93 radian (about 53 degrees), and other methods can also be used in other embodiments, which are not limited in the present application.

[0036] In a specific implementation, the color-texture gradient direction of each pixel in the remote sensing image can be determined in the following manner. First, the gradient components of the color gradient and the texture gradient of each pixel in the remote sensing image in the horizontal direction and the vertical direction are obtained. Second, the gradient components of the color gradient and the texture gradient of each pixel in the horizontal direction are linearly weighted according to the same weight when the color gradient amplitude and the texture gradient amplitude are linearly weighted and fused, and the gradient components of the color gradient and the texture gradient of each pixel in the vertical direction are linearly weighted according to the same weight, so as to obtain the gradient components of the color-texture gradient of each pixel in the horizontal direction and the vertical direction. Third, the color-texture gradient direction of each pixel is calculated using an arctan2 function, and the result is normalized to the radian range of 0 to 2π. In other embodiments, other methods can also be used, which are not limited in the present application.

[0037] In a specific implementation, the gradient consistency weight of each pixel in the remote sensing image can be determined in the following manner. For each pixel, the absolute difference between the class probability gradient direction and the color-texture gradient direction of the pixel is calculated. Because the periodicity of the direction information is considered, the direction inconsistency can be measured by calculating the absolute value of the sine value of the difference, and then the inconsistency value is substituted into an exponential decay function to calculate the gradient consistency weight of each pixel. The decay coefficient β is set to 2.0, so that the weight is close to 1 when the two gradient directions are consistent, and the weight is the smallest when the directions are perpendicular. For example, the class probability gradient direction of a certain pixel is 0.93 radians, and the color-texture gradient direction is 2.356 radians. The absolute direction difference between the two is |0.93-2.356|=1.426 radians. Then, the gradient consistency weight is exp(-2.0×|sin(1.426)|), where sin(1.426)≈0.99, so the gradient consistency weight≈exp(-2.0×0.99)≈exp(-1.98)≈0.14. The low gradient consistency weight indicates that the class probability gradient and the color-texture gradient have a large difference in direction at the pixel, and the semantic boundary and the image edge are inconsistent at this point. In other embodiments, other methods can also be used, which are not limited in the present application.

[0038] It should be noted that the attenuation coefficient β in the gradient consistency weight calculation in the present application is used to control the influence strength of the direction consistency on the weight, and the value range is usually [0.5, 5.0], and the preferred value range is [1.0, 3.0]. For medium-high resolution remote sensing images, the recommended initial value is 2.0, and other methods can also be used in other embodiments, which are not limited in the present application.

[0039] In a specific implementation, the gradient consistency weight is used to weight and fuse the category probability gradient amplitude of each pixel in the initial cultivated land probability graph in multiple directions and the color-texture gradient amplitude corresponding to each pixel in the color-texture gradient amplitude graph to generate a boundary saliency graph. The following method can be used, for example: for each pixel, the gradient consistency weight, the category probability gradient amplitude and the color-texture gradient amplitude corresponding to the pixel are calculated according to the weighted fusion formula. Specifically, the gradient consistency weight is used as the weight of the category probability gradient amplitude, and the complement of the gradient consistency weight is used as the weight of the color-texture gradient amplitude. The boundary saliency value of each pixel is calculated by weighted summation. Finally, the boundary saliency values of all pixels are combined into a boundary saliency graph with the same size as the original image. It should be noted that this fusion mechanism can strengthen the semantic gradient response in the area where the semantic boundary is consistent with the image edge, and rely on the image feature gradient in the inconsistent area. In other embodiments, other methods can also be used, which are not limited in the present application.

[0040] It should be noted that the boundary saliency graph in the present application comprehensively represents the overall saliency degree of each pixel point in the image as a cultivated land boundary. It is an enhanced boundary response graph that fuses semantic information and image features. It highlights the real cultivated land boundary while effectively suppressing irrelevant texture or color changes in the cultivated land by fusing high-level semantic boundary strength and low-level image edge features.

[0041] In step 104, the semantic-geometric feature mismatch degree of each pixel point in the boundary saliency graph is determined according to the semantic probability gradient amplitude and the color-texture gradient amplitude of each pixel point. The semantic-geometric feature mismatch degree is the absolute difference value of the category probability gradient direction and the color-texture gradient direction of each pixel point.

[0042] In some embodiments, the semantic-geometric feature mismatch degree of each pixel point in the boundary saliency graph can be determined according to the semantic probability gradient amplitude and the color-texture gradient amplitude of each pixel point by the following steps, that is: The category probability gradient direction and the color-texture gradient direction of each pixel point are obtained. The absolute direction difference between the category probability gradient direction and the color-texture gradient direction of each pixel point is calculated. The absolute direction difference corresponding to each pixel point is normalized to obtain the semantic-geometric feature mismatch degree of each pixel point.

[0043] In a specific implementation, the absolute direction difference of each pixel point between the category probability gradient direction and the color-texture gradient direction can be calculated in the following manner: for each pixel point, the category probability gradient direction and the color-texture gradient direction of the pixel point are directly read from the process of constructing the boundary saliency map, then a simple arithmetic difference between the category probability gradient direction and the color-texture gradient direction of the pixel point is calculated, and an absolute value is taken to obtain an initial direction difference. Since the direction information is periodic, the initial direction difference needs to be periodically corrected, that is, when the difference is greater than π, 2π is subtracted from the difference, to ensure that the final absolute direction difference always falls within the range of 0 to π radians. For example, it is assumed that the category probability gradient direction of the pixel is 0.93 radians and the color-texture gradient direction is 2.356 radians, which are read from the cache. The initial direction difference is calculated as |0.93-2.356|=1.426 radians. Since 1.426<π, no adjustment is needed, and the final absolute direction difference is 1.426 radians. In other embodiments, other methods can also be used, which are not limited in the present application.

[0044] In a specific implementation, the absolute direction difference corresponding to each pixel point is normalized to obtain the semantic-geometric feature mismatch degree of each pixel point in the following manner: the absolute direction difference of each pixel point is divided by its theoretical maximum value π to linearly scale the direction difference to a standardized range of 0 to 1. The calculation formula is semantic-geometric feature mismatch degree=absolute direction difference / π. Through this simple scaling, the semantic-geometric feature mismatch degree of each pixel point is uniformly normalized, where 0 indicates that the category probability gradient direction and the color-texture gradient direction of the pixel point are completely consistent, and 1 indicates that the category probability gradient direction and the color-texture gradient direction of the pixel point are completely opposite. The obtained semantic-geometric feature mismatch degree constitutes a semantic-geometric feature mismatch degree map of the same size as the original image. In other embodiments, other methods can also be used, which are not limited in the present application.

[0045] It should be noted that the semantic-geometric feature mismatch degree in the present application is the absolute difference between the category probability gradient direction and the color-texture gradient direction of each pixel point, which measures the consistency degree of the semantic segmentation boundary (i.e., the category probability gradient direction) and the visual edge (i.e., the color-texture gradient direction) derived based on the image itself in the spatial direction. It can be used as a basis for reliability in subsequent processing, for example, a high mismatch degree identifies an unreliable area where the semantic judgment and visual evidence are contradictory, and a low mismatch degree identifies a high confidence area where the two mutually confirm each other.

[0046] In step 105, the boundary saliency map is adaptively filtered by the semantic-geometric feature mismatch degree of all pixel points, and then a cultivated land boundary vector of the target cultivated land is generated based on the filtered boundary saliency map.

[0047] In some embodiments, the adaptive filtering of the boundary saliency map by the semantic-geometric feature mismatch degree of all pixel points can be implemented by the following steps, that is: determining an adaptive filtering parameter of a local region in the boundary saliency map based on the semantic-geometric feature mismatch degree of all pixel points; performing filtering processing on the boundary saliency map according to the adaptive filtering parameter to obtain an enhanced boundary saliency map.

[0048] In specific implementation, the adaptive filtering parameter of a local region in the boundary saliency map based on the semantic-geometric feature mismatch degree of all pixel points can be determined by the following manner, for example: first, a local window of a specified size is created with each pixel point as the center, and the average value of the semantic-geometric feature mismatch degrees of all pixels in the local window is calculated as the local average mismatch degree of the corresponding pixel point; then, the local average mismatch degree value is converted into an adaptive range parameter σ_r of the bilateral filter through a linear mapping function, and the specific mapping relationship is adaptive range parameter σ_r = σ_{r0} + γ·local average mismatch degree, wherein σ_{r0} is a basic range parameter, and the recommended value range is [0.05, 0.15], for example, 0.1, γ is a proportional coefficient, and the recommended value range is [0.3, 0.7], for example, 0.5, and the spatial parameter σ_s is set to a fixed value of [1.0, 2.0] pixels according to the image resolution, for example, 1.5, wherein the adaptive filtering parameter includes an adaptive range parameter and a fixed spatial parameter; in this way, the σ_r value is automatically increased in the region with high local average mismatch degree to enhance the smoothing effect, and the σ_r value is reduced in the region with low local average mismatch degree to retain edge details, so as to realize adaptive adjustment of the filtering parameter, and in other embodiments, other methods can also be used for implementation, which is not limited in the present application.

[0049] In a specific implementation, filtering processing is performed on the boundary saliency map according to the adaptive filtering parameter to obtain an enhanced boundary saliency map. The filtering processing can be performed by using a bilateral filter on the boundary saliency map. For each pixel in the map, the spatial weight and the gray value weight of all the pixels in the neighborhood of the pixel are calculated by using the adaptive range parameter and the fixed spatial parameter. The spatial weight and the gray value weight of each neighborhood pixel are multiplied to obtain the maximum weight of the corresponding neighborhood pixel. The maximum weight is multiplied by the boundary saliency value of the neighborhood pixel and accumulated. Finally, the boundary saliency value of the corresponding center pixel after filtering is obtained by dividing the accumulated value by the sum of all the weights. After the operation is performed on all the pixels in the boundary saliency map, the enhanced boundary saliency map is obtained. In other embodiments, other methods can also be used to implement the filtering processing, which are not limited in the present application.

[0050] In some embodiments, the boundary vector of the target cultivated land is generated based on the filtered boundary saliency map by using the following steps. Non-maximum suppression and edge connection are performed on the filtered boundary saliency map to extract all the candidate boundary pixels. The candidate boundary pixels are converted into vector boundary lines, and the semantic-geometric feature mismatch degree of the corresponding pixels is used as a constraint to fuse the broken line segments and remove the isolated false boundaries in the vector layer to generate the final cultivated land boundary vector of the cultivated land target.

[0051] In a specific implementation, non-maximum suppression and edge connection are performed on the filtered boundary saliency map, and then all candidate boundary pixels are extracted by using the following method. First, non-maximum suppression is performed on the enhanced boundary saliency map. Specifically, comparison is performed in a 3x3 neighborhood in the gradient direction of each pixel point, and only the pixel with the largest boundary saliency value is retained while other non-maximum pixels are suppressed, thereby obtaining a refined boundary response. Specifically, assuming that there is a 3x3 local region, the boundary saliency value of the center pixel is 0.8, and the gradient direction is the horizontal direction with an angle of 0°. Therefore, in the horizontal direction, the boundary saliency values of the center pixel and its left and right two neighboring pixels are compared. Assuming that the boundary saliency value of the left pixel is 0.6 and the boundary saliency value of the right pixel is 0.7, since the value 0.8 of the center pixel is greater than the boundary saliency values of the left and right two neighboring pixels, the center pixel is retained, and the boundary saliency values of the left and right two neighboring pixels are suppressed and set to 0. Then, double-threshold method is used for edge connection. A high threshold and a low threshold are set, and the high threshold and the low threshold can be adaptively determined according to the global statistical information of the boundary saliency map. For example, the high threshold and the low threshold are respectively set to 0.7 times and 0.3 times of the maximum boundary saliency value in the boundary saliency map. The pixels with a boundary saliency value higher than the high threshold are marked as strong boundary pixels, the pixels with a boundary saliency value lower than the low threshold are directly suppressed, and the pixels with a boundary saliency value between the two thresholds are marked as weak boundary pixels. Finally, a boundary tracking algorithm is used to connect the weak boundary pixels in the 8-neighborhood of the strong boundary pixels, thereby forming a complete boundary contour. All retained and connected pixels are candidate boundary pixels. In other embodiments, other methods can also be used for implementation, which are not limited in the present application.

[0052] In a specific implementation, all candidate boundary pixels are converted into vector boundary lines, and the semantic-geometric feature mismatch degree of the corresponding pixel points is taken as a constraint to fuse broken line segments and remove isolated false boundaries at the vector level, and finally generate the cultivated land boundary vector of the cultivated land target. The following methods can be used, for example: first, the boundary tracking algorithm based on eight-neighbor connection is used to organize the candidate boundary pixels into continuous boundary chains, and convert the boundary chains into initial vector boundary lines; then, the semantic-geometric feature mismatch degree of each boundary line segment is calculated, and when the average value of the semantic-geometric feature mismatch degree of all pixels covered by each boundary line segment exceeds a set threshold, it is considered that the corresponding line segment has a low credibility and is removed; at the same time, for broken line segments that are close in space and similar in direction, if the average mismatch degree of the pixels at their connection is lower than a set threshold, the line segment fusion operation is performed; finally, the Douglas-Peucker algorithm is applied to the optimized vector boundary to simplify and reduce the number of vertices, and the result is converted into a vector file format with geographic coordinates, to complete the generation of the cultivated land boundary vector. In other embodiments, other methods can also be used, which are not limited in the present application.

[0053] In addition, another aspect of the present application provides a cultivated land boundary identification system in some embodiments, which is described with reference to Figure 4 The figure is a structural schematic diagram of a cultivated land boundary identification system according to some embodiments of the present application, which includes a collection module 401, a processing module 402 and an execution module 403, which are described as follows: The collection module 401 is mainly used for collecting remote sensing images of the target cultivated land in the present application; The processing module 402 is mainly used for performing cultivated land semantic segmentation on the remote sensing images to obtain an initial cultivated land probability map in the present application; The processing module 402 is also used for determining the color-texture gradient amplitude map of the remote sensing images, and then constructing a boundary saliency map according to the class probability gradient in multiple directions of each pixel in the initial cultivated land probability map and the color-texture gradient amplitude map of the remote sensing images; The processing module 402 is also used for determining the semantic-geometric feature mismatch degree of each pixel point according to the semantic probability gradient amplitude and the color-texture gradient amplitude of each pixel point in the boundary saliency map, wherein the semantic-geometric feature mismatch degree is the absolute difference value of the class probability gradient direction and the color-texture gradient direction of each pixel point; The execution module 403 is mainly used for adaptively filtering the boundary saliency map through the semantic-geometric feature mismatch degree of all the pixel points, and generating the cultivated land boundary vector of the target cultivated land based on the filtered boundary saliency map.

[0054] The modules in the cultivated land boundary identification system can be realized by software, hardware, or a combination thereof. The modules can be embedded in or independent of the processor in the computer device in hardware form, or stored in the memory in the computer device in software form, so as to be called and executed by the processor.

[0055] In addition, in an embodiment, the present application provides a computer device, which can be a server, and the internal structure diagram thereof can be as shown in Figure 5 The computer device includes a processor, a memory, and a network interface connected through a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store cultivated land boundary identification data. The network interface of the computer device is used to communicate with an external terminal through a network connection. The computer program is executed by the processor to implement a cultivated land boundary identification method.

[0056] Those skilled in the art can understand that Figure 5 The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different component arrangement.

[0057] In an embodiment, a computer device is also provided, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-mentioned cultivated land boundary identification method embodiments.

[0058] In an embodiment, a computer readable storage medium is provided, which stores a computer program. The computer program is executed by a processor to implement the steps in the above-mentioned cultivated land boundary identification method embodiments.

[0059] In one embodiment, a computer program product or computer program is provided, which includes computer instructions stored in a computer readable storage medium. A processor of a computer device reads the computer instructions from the computer readable storage medium, and the processor executes the computer instructions, so that the computer device performs the steps in the above-mentioned embodiment of the cultivated land boundary identification method.

[0060] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, the computer program can include the processes of the above-mentioned embodiments. Any reference to memory, storage, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. The non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory. The volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, the RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0061] The technical features of the above embodiments can be combined in any manner. To make the description concise, all possible combinations of the technical features in the above embodiments are not described, but as long as the combinations of the technical features do not exist, they should be considered as the scope of the present application.

[0062] The above embodiments only express several implementation manners of the present application, and the description is specific and detailed, but it should not be understood as a limitation on the scope of the patent. It should be pointed out that for those of ordinary skill in the art, without departing from the concept of the present application, some modifications and improvements can be made, which are all within the scope of the present application. Therefore, the scope of the patent of the present application should be subject to the appended claims.

Claims

1. A method of identifying a field boundary, characterized by, The method comprises the following steps: Collecting a remote sensing image of a target farmland; Performing farmland semantic segmentation on the remote sensing image to obtain an initial farmland probability map; Determining a color-texture gradient amplitude map of the remote sensing image, and then constructing a boundary saliency map based on the class probability gradient of each pixel in the initial farmland probability map in multiple directions and the color-texture gradient amplitude map of the remote sensing image; Determining a semantic-geometric feature mismatch degree of each pixel point in the boundary saliency map according to the semantic probability gradient amplitude and the color-texture gradient amplitude of each pixel point, wherein the semantic-geometric feature mismatch degree is an absolute difference value of the class probability gradient direction and the color-texture gradient direction of each pixel point; Adaptively filtering the boundary saliency map through the semantic-geometric feature mismatch degrees of all pixel points, and then generating a farmland boundary vector of the target farmland based on the filtered boundary saliency map.

2. The method of claim 1, wherein, The farmland semantic segmentation on the remote sensing image to obtain an initial farmland probability map specifically comprises: Obtaining a pre-trained farmland semantic segmentation model; Inputting the remote sensing image into the farmland semantic segmentation model for forward reasoning to obtain an initial segmentation result; Performing probabilistic post-processing on the initial segmentation result to obtain the initial farmland probability map.

3. The method of claim 1, wherein, The determination of the color-texture gradient amplitude map of the remote sensing image specifically comprises: Determining the color gradient amplitude and the texture gradient amplitude of each pixel in the remote sensing image; Determining the color-texture gradient amplitude map of the remote sensing image according to the color gradient amplitude and the texture gradient amplitude of each pixel.

4. The method of claim 1, wherein, The construction of the boundary saliency map based on the class probability gradient of each pixel in the initial farmland probability map in multiple directions and the color-texture gradient amplitude map of the remote sensing image specifically comprises: Determining the class probability gradient amplitude and the class probability gradient direction of each pixel in the initial farmland probability map in multiple directions; Determining the color-texture gradient direction of each pixel in the remote sensing image; Determining the gradient consistency weight according to the class probability gradient direction of each pixel in the initial farmland probability map in multiple directions and the color-texture gradient direction of each pixel in the remote sensing image; Generating the boundary saliency map based on the weighted fusion of the class probability gradient amplitude of each pixel in the initial farmland probability map in multiple directions and the corresponding color-texture gradient amplitude of each pixel in the color-texture gradient amplitude map.

5. The method of claim 1, wherein, The determination of the semantic-geometric feature mismatch degree of each pixel point in the boundary saliency map according to the semantic probability gradient amplitude and the color-texture gradient amplitude of each pixel point specifically comprises: Obtaining the class probability gradient direction and the color-texture gradient direction of each pixel point; Calculating the absolute direction difference corresponding to each pixel point between the class probability gradient direction and the color-texture gradient direction; Normalizing the absolute direction difference corresponding to each pixel point to obtain the semantic-geometric feature mismatch degree of each pixel point.

6. The method of claim 1, wherein, The adaptive filtering of the boundary saliency map through the semantic-geometric feature mismatch degrees of all pixel points specifically comprises: Determine adaptive filtering parameters of local regions in the boundary saliency map based on semantic-geometric feature mismatch degrees of all pixel points; Perform filtering processing on the boundary saliency map according to the adaptive filtering parameters to obtain an enhanced boundary saliency map.

7. The method of claim 1, wherein, Generating the cultivated land boundary vector of the target cultivated land based on the filtered boundary saliency map specifically includes: Perform non-maximum suppression and edge connection on the filtered boundary saliency map, and then extract all candidate boundary pixels; Convert all candidate boundary pixels into vector boundary lines, take semantic-geometric feature mismatch degrees of corresponding pixel points as constraints, fuse broken line segments at a vector level, and remove isolated false boundaries to generate the final cultivated land boundary vector of the cultivated land target.

8. A cultivated boundary identification system, characterized by, It includes: The acquisition module is configured to acquire a remote sensing image of a target cultivated land. The processing module is configured to perform cultivated land semantic segmentation on the remote sensing image to obtain an initial cultivated land probability map. The processing module is further configured to determine a color-texture gradient amplitude map of the remote sensing image, and then construct a boundary saliency map based on class probability gradients of each pixel in the initial cultivated land probability map in multiple directions and the color-texture gradient amplitude map of the remote sensing image. The processing module is further configured to determine a semantic-geometric feature mismatch degree of each pixel point in the boundary saliency map based on a semantic probability gradient amplitude and a color-texture gradient amplitude of the pixel point, wherein the semantic-geometric feature mismatch degree is an absolute difference value between a class probability gradient direction and a color-texture gradient direction of each pixel point. The execution module is configured to perform adaptive filtering on the boundary saliency map based on semantic-geometric feature mismatch degrees of all pixel points, and then generate a cultivated land boundary vector of the target cultivated land based on the filtered boundary saliency map. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The processor executes the computer program to implement the steps of the cultivated land boundary identification method in any one of claims 1 to 7.

10. A computer readable storage medium storing a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the cultivated land boundary identification method in any one of claims 1 to 7.

Images