Hilly mountainous area farmland monitoring and protection method fusing low-altitude AI sensing technology

Abstract

The present application relates to remote sensing mapping and computer vision technical field, especially to the hilly and mountainous area cultivated land monitoring protection method fusing low altitude AI perception technology, including the following steps: S1, the original aerial image sequence with preset overlap rate and attitude position information of the monitoring area are acquired;S2, based on the original aerial image sequence and attitude position information, a three-dimensional dense point cloud is constructed by using a multi-view stereo matching algorithm, and the three-dimensional dense point cloud is subjected to visibility analysis, the texture mapping of the projection occlusion area is removed, and a true projection image is generated, S3, the three-dimensional dense point cloud is subjected to cloth simulation filtering, in the present application, the ground point extraction is carried out on the three-dimensional point cloud, and a digital terrain model is generated, the real ground slope value is calculated, the slope value is combined with the crop type of the land plot to realize the objective quantitative monitoring of the hilly and mountainous area steep slope cultivation behavior.

Description

Technical Field

[0001] This invention relates to the fields of remote sensing mapping and computer vision technology, and in particular to a method for monitoring and protecting cultivated land in hilly and mountainous areas that integrates low-altitude AI sensing technology. Background Technology

[0002] my country's hilly and mountainous areas have a wide distribution of arable land and complex terrain, making them key areas for arable land protection and supervision. Currently, the main methods used are satellite remote sensing or drone aerial photography to obtain two-dimensional orthophotos of the monitoring area. The land cover category is then extracted through manual visual interpretation or automatic computer identification to determine whether there is any illegal occupation of arable land.

[0003] However, existing monitoring methods mainly rely on planar image features and lack effective perception of three-dimensional terrain information. It is difficult to accurately obtain the real slope data of the plot and to quantify and judge illegal behaviors such as cultivation on steep slopes. Moreover, when the surface vegetation is dense, traditional methods are difficult to penetrate the vegetation layer to obtain the real bare soil elevation, which can easily lead to missed detection or misjudgment of illegally cultivated plots hidden in the mountains and forests. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides a method for monitoring and protecting cultivated land in hilly and mountainous areas that integrates low-altitude AI perception technology. It aims to improve the existing monitoring methods, which mainly rely on planar image features, lack effective perception of three-dimensional terrain information, and are difficult to accurately obtain the real slope data of the plots.

[0005] This invention provides the following technical solution: a method for monitoring and protecting cultivated land in hilly and mountainous areas that integrates low-altitude AI sensing technology, comprising the following steps: S1. Acquire the original aerial image sequence and attitude position information of the monitoring area with a preset overlap rate; S2. Based on the original aerial image sequence and attitude position information, a three-dimensional dense point cloud is constructed using a multi-view stereo matching algorithm, and the visibility of the three-dimensional dense point cloud is analyzed to remove the texture mapping of the projected occlusion area and generate a true image. S3. Perform cloth simulation filtering on the three-dimensional dense point cloud, use the iterative settlement process of the virtual cloth model under gravity to separate ground points and non-ground points, use the separated ground points to generate a digital terrain model by interpolation, and after downsampling and smoothing the digital terrain model, use the difference algorithm to calculate the global terrain slope map. S4. Convert the true image to the Lab color space, calculate the shadow threshold using the luminance channel histogram features to obtain a binarized shadow mask, and apply the limited contrast adaptive histogram equalization algorithm to adjust the pixel grayscale values ​​only for the area covered by the binarized shadow mask to generate an enhanced image. S5. Input the enhanced image into a pre-trained semantic segmentation neural network model, output pixel-level land feature classification results, and perform morphological operations and connected component extraction on the pixel-level land feature classification results to generate a land parcel vector plot containing land feature attributes. S6. Using the land parcel vector plot as a spatial statistical unit, perform zonal statistical analysis on the global topographic slope map, calculate the average slope value within the range of each land parcel, and combine the land feature attributes with the average slope value to make logical judgments using a preset rule tree, and output the monitoring results.

[0006] Preferably, in step S2, the construction of a 3D dense point cloud using a multi-view stereo matching algorithm specifically includes the following steps: Feature points are extracted and matched from the original aerial image sequence, and the camera intrinsic parameters, extrinsic parameters, and sparse point cloud are solved using bundle adjustment. Based on the solved camera intrinsic and extrinsic parameters and sparse point cloud, a semi-global matching algorithm is used to calculate the depth map of each image. The depth map is subjected to multi-view consistency verification and fusion to remove redundant points and outlier noise points, generating a three-dimensional dense point cloud.

[0007] Preferably, in step S2, generating the true image specifically includes the following steps: A digital surface model is constructed based on the aforementioned three-dimensional dense point cloud; Establish an orthographic projection grid, and for each pixel unit in the grid, construct a line-of-sight vector from the ground point to the camera center; Detect whether there is an intersection relationship between the gaze vector and the digital surface model. If they intersect, mark it as an occluded area; if they do not intersect, mark it as a visible area. Traverse the original aerial image sequence, select the image texture corresponding to the visible area with the smallest projection angle to fill the pixel unit, and generate a true image.

[0008] Preferably, in step S3, performing cloth simulation filtering on the three-dimensional dense point cloud specifically includes the following steps: The elevation coordinates of the three-dimensional dense point cloud are inverted to construct a flipped three-dimensional scene; A virtual cloth grid with a preset grid resolution and stiffness coefficient is initialized above the flipped 3D scene; Force simulation and iterative calculation are performed on the nodes in the virtual fabric grid. In each iteration, the downward displacement of the nodes is calculated based on the effect of gravity, and the node positions are adjusted according to the internal force constraints between the nodes until the fabric grid shape converges and stabilizes. Calculate the vertical distance from each point in the three-dimensional dense point cloud to the converged cloth grid, and mark the points whose vertical distance is less than a preset threshold as ground points, and the rest as non-ground points.

[0009] Preferably, in step S3, the calculation of the global terrain slope map using the difference algorithm specifically includes the following steps: The digital terrain model is resampled to a preset spatial resolution using a bilinear interpolation algorithm, and the resampled digital terrain model is then smoothed using a Gaussian kernel function. Using a sliding window to traverse the smoothed digital terrain model, the elevation gradient values ​​in the horizontal and vertical directions are calculated for the center pixel of the window using the third-order inverse distance squared weighted difference formula. The composite vector magnitude is calculated based on the elevation gradient values ​​in the horizontal and vertical directions, and the composite vector magnitude is converted into a slope angle value using the arctangent function to generate a global terrain slope map.

[0010] Preferably, in step S4, the step of calculating the shadow threshold using the luminance channel histogram features to obtain the binarized shadow mask specifically includes the following steps: Extract the luminance channel components in the Lab color space and calculate the grayscale histogram of the luminance channel components; The gray levels of the gray histogram are traversed using the maximum inter-class variance method to calculate the inter-class variance value when the image is segmented into foreground and background classes. The gray level corresponding to the maximum inter-class variance is selected as the segmentation threshold; The brightness channel components are compared pixel by pixel with the segmentation threshold, and regions with pixel values ​​lower than the segmentation threshold are marked as shadow regions to generate a binarized shadow mask.

[0011] Preferably, in step S5, the step of inputting the enhanced image into a pre-trained semantic segmentation neural network model and outputting pixel-level land cover classification results specifically includes the following steps: The enhanced image is standardized, and the image size is adjusted to fit the input dimension of the semantic segmentation neural network model; The encoder of the semantic segmentation neural network model extracts deep semantic features and shallow spatial features of the image, and the decoder performs multi-scale fusion and upsampling on the deep semantic features and shallow spatial features. Output a multi-channel probability map with the same resolution as the input image, where each channel corresponds to a land cover category; For each pixel location, the category index corresponding to the channel with the highest probability value is selected as the classification label for that pixel, generating pixel-level land feature classification results.

[0012] Preferably, in step S6, the step of combining the land feature attributes with the average slope value using a preset rule tree for logical determination and outputting the monitoring results specifically includes the following steps: Determine whether the land feature attribute is a food crop. If so, further compare the average slope value with a preset slope threshold. When the average slope value is greater than the slope threshold, it is determined to be a steep slope cultivation state. Determine whether the land feature is any one of forest, orchard, or pond / water surface. If so, it is directly determined to be occupied by non-grain planting. Determine whether the land feature is any of the following: building, road, or paved surface. If so, it is directly determined to be occupied by non-agricultural construction.

[0013] The present invention has the following beneficial effects: 1. In this invention, ground points are extracted from three-dimensional point clouds to generate a digital terrain model, the actual ground slope value is calculated, and the slope value is jointly determined with the crop type of the plot, thereby realizing objective quantitative monitoring of cultivation behavior on steep slopes in hilly and mountainous areas.

[0014] 2. In this invention, geometric projection analysis is used to remove occluded areas in the image, and targeted texture restoration is performed on the shadow area based on brightness characteristics, which solves the problem of difficulty in identifying ground features due to insufficient lighting or obstructed view in complex terrain.

[0015] 3. In this invention, by integrating the semantic information of images with the spatial geometric information of terrain, the system automatically performs multi-dimensional compliance logic judgments on the patches within the monitoring area, thereby enabling rapid investigation of the non-grain and non-agricultural use of arable land. Attached Figure Description

[0016] Figure 1 This is a flowchart of the hilly and mountainous area farmland monitoring and protection method that integrates low-altitude AI sensing technology proposed in this invention. Detailed Implementation

[0017] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] This invention provides a method for monitoring and protecting cultivated land in hilly and mountainous areas that integrates low-altitude AI sensing technology, such as... Figure 1 As shown, it includes the following steps: S1. Acquire the original aerial image sequence and attitude position information of the monitoring area with a preset overlap rate; Specifically, a multi-rotor UAV equipped with a high-resolution visible light camera and a real-time dynamic differential positioning module was selected as the data acquisition platform. Operations were conducted during a time window with sufficient sunlight and a large solar altitude angle, such as between 11:00 AM and 2:00 PM local time, to reduce the obscuring of surface textures by the shadows cast by hilly and mountainous terrain. Before the flight operation, a rough digital elevation model of the survey area was imported into the ground station control software for planning the terrain-following flight path. To overcome the resolution inconsistency caused by the large elevation differences in hilly terrain, the design flight altitude of the UAV relative to the ground was calculated by setting the expected ground sampling distance and camera parameters. The calculation formula is: In the formula, For camera focal length, The target ground sampling distance, This refers to the pixel size of the camera sensor.

[0019] During flight, the drone dynamically adjusts its absolute flight altitude based on the real-time ground elevation. The calculation formula is: In the formula, Given the ground elevation value corresponding to the current UAV's planar coordinates in the digital elevation model, and to meet the feature matching requirements of subsequent high-precision 3D reconstruction, it is necessary to calculate the flight path interval and the image baseline based on the set overlap rate parameter, as well as the distance between two adjacent flight paths. The calculation formula is The baseline distance between two adjacent photos taken along the flight path The calculation formula is In the formula, The width of the ground projection area of ​​a single image. The length of a single image projected onto the ground. This refers to the lateral overlap rate. For the heading overlap rate, in this embodiment, it will be Set to 80%, Set to 70% to ensure sufficient multi-angle texture information can be obtained even in areas with large terrain undulations.

[0020] When the drone flies along the planned route, the flight control system triggers the camera to expose and capture images at predetermined positions, obtaining a sequence of raw aerial images. Simultaneously, the real-time dynamic differential positioning module and the inertial measurement unit synchronously record the precise spatial position and attitude information at the moment of each image exposure. The spatial position information includes longitude, latitude, and altitude, while the attitude information includes roll angle. Pitch angle and yaw angle These position and attitude data are written into the image's metadata or stored as independent positioning and orientation system data files, ultimately completing the data acquisition work for the monitored area.

[0021] By controlling flight altitude and overlap rate through a rigorous mathematical model, equal-resolution imaging and full-coverage acquisition were achieved in complex terrain, effectively avoiding the problems of image blurring and data gaps that are prone to occur in conventional aerial surveys in mountainous areas.

[0022] S2. Based on the original aerial image sequence and attitude position information, a three-dimensional dense point cloud is constructed using a multi-view stereo matching algorithm. Visibility analysis is performed on the three-dimensional dense point cloud to remove texture mapping of the projected occluded area and generate a true image. Furthermore, in S2, the construction of a 3D dense point cloud using a multi-view stereo matching algorithm specifically includes the following steps: Feature points are extracted and matched from the original aerial image sequence, and the camera intrinsic parameters, extrinsic parameters and sparse point cloud are solved by bundle adjustment. Based on the solved camera intrinsic and extrinsic parameters and sparse point cloud, a semi-global matching algorithm is used to calculate the depth map of each image. Multi-view consistency verification and fusion are performed on the depth map to remove redundant points and outliers, generating a 3D dense point cloud.

[0023] Furthermore, in S2, generating a true image specifically includes the following steps: Constructing a digital surface model based on 3D dense point cloud; Establish an orthographic projection grid, and for each pixel unit in the grid, construct a line-of-sight vector from the ground point to the camera center; Detect whether there is an intersection between the gaze vector and the digital surface model. If they intersect, mark them as occluded areas; if they do not intersect, mark them as visible areas. Traverse the original aerial image sequence, select the image texture corresponding to the visible area and with the smallest projection angle to fill the pixel unit, and generate a true image.

[0024] Specifically, after acquiring the original aerial image sequence and attitude position information of the monitoring area, multi-view stereo matching technology is first used for 3D reconstruction. The data acquired in step S1 is then imported into the photogrammetric processing system for feature extraction of the image sequence. A scale-invariant feature transform algorithm is used to extract feature points from each image, and the k-nearest neighbor algorithm is used for feature matching between overlapping images. Based on the matching results, a bundle adjustment model is constructed. The camera intrinsic parameters, extrinsic parameters, and 3D point coordinates are jointly optimized by minimizing the reprojection error. The objective function of the adjustment optimization is... Defined as In the formula, This represents the objective function for the total reprojection error. This represents the total number of images involved in the adjustment. This represents the total number of three-dimensional feature points. As a visibility indicator variable, when the first Zhang's image observations showed the first The value is 1 when there are 3D points, and 0 otherwise. Indicates the first The three-dimensional point at the th t The observed pixel coordinates on the image; Includes the first The camera intrinsic parameter matrix, rotation matrix, and translation vector corresponding to each image; Indicates the first The spatial coordinates of a three-dimensional point in the world coordinate system; This represents the projection function that maps three-dimensional coordinates to a two-dimensional image plane. By solving this function through nonlinear iteration, accurate camera parameters and sparse point clouds can be obtained.

[0025] Based on the calculated camera parameters, a semi-global matching algorithm is used for dense matching. For each image, the cross-correlation or Census transform cost between pixels is calculated, and the path cost is aggregated in multiple directions to find the optimal solution for disparity, generating a depth map. Multi-view geometric consistency is checked on the depth maps generated from all images. The depth maps are then back-projected into 3D space, and the reprojection distance deviation of the same point in different views is calculated. A distance deviation threshold is set. The sampling distance is doubled, redundant points and outliers with deviations greater than the threshold are removed, and the remaining valid points are fused to generate a high-density 3D dense point cloud containing color information.

[0026] Using the constructed 3D dense point cloud, an irregular triangular mesh construction algorithm is employed to generate a detailed digital surface model. Subsequently, an orthographic projection mesh consistent with the survey area is established, with the mesh spacing set to the target output resolution. For each pixel unit in the mesh... Obtain the corresponding ground space point coordinates based on the digital surface model. To eliminate obstruction caused by tall vegetation or mountains, visibility analysis was performed on the ground point, and a visibility model was constructed from the ground point. Pointing to the center of each candidate image camera line-of-sight vector Using a ray casting algorithm to detect the gaze vector Does the viewpoint intersect the digital surface model geometrically? If the view vector intersects with other triangular faces of the digital surface model before reaching the camera center, then the ground point is determined to be at the [missing information - likely a specific point or location]. An image that is obscured is marked as invisible; otherwise, it is marked as visible.

[0027] After determining the set of visible images for each ground pixel unit, all images in the set are traversed, and the projection angle of each image onto the ground point is calculated. Projection angle Defined as gaze vector Normal vector at ground point The angle between them is calculated using the following formula: In the formula, To point from the ground point to the first The gaze vector at the center of each camera. Let be the normal vector of the triangular facet containing the ground point. and Let each represent the magnitude of the corresponding vector, and select... Minimum value (i.e.) The image with the largest value is used as the best texture source. The pixel color value of the corresponding position in the image is read and filled into the current pixel unit of the orthophoto grid. After all pixels in the grid are filled, color uniform processing is applied to eliminate the stitching gaps, and finally the true image is generated.

[0028] Through strict geometric constraints and occlusion detection mechanisms, the phenomena of fallen trees, streaked buildings, and shadows commonly found in traditional orthophotos are effectively avoided, ensuring the true geometric positional relationship between farmland boundaries and vegetation cover, and providing distortion-free base map data for subsequent high-precision plot segmentation.

[0029] S3. Perform cloth simulation filtering on the 3D dense point cloud, use the iterative settlement process of the virtual cloth model under gravity to separate ground points and non-ground points, use the separated ground points to generate a digital terrain model by interpolation, and after downsampling and smoothing the digital terrain model, use the difference algorithm to calculate the global terrain slope map. Furthermore, in S3, performing cloth simulation filtering on a 3D dense point cloud specifically includes the following steps: Invert the elevation coordinates of the 3D dense point cloud to construct a flipped 3D scene; Initialize a virtual cloth grid with preset mesh resolution and stiffness coefficient above the flipped 3D scene; Force simulation and iterative calculation are performed on the nodes in the virtual cloth grid. In each iteration, the downward displacement of the nodes is calculated based on the effect of gravity, and the node positions are adjusted according to the internal force constraints between the nodes until the cloth grid shape converges and stabilizes. Calculate the vertical distance from each point in the 3D dense point cloud to the converged cloth grid, mark points with a vertical distance less than a preset threshold as ground points, and mark the rest as non-ground points.

[0030] Furthermore, in S3, the calculation of the global terrain slope map using the difference algorithm specifically includes the following steps: The digital terrain model is resampled to a preset spatial resolution using a bilinear interpolation algorithm, and the resampled digital terrain model is then smoothed using a Gaussian kernel function. Using a sliding window to traverse the smoothed digital terrain model, the elevation gradient values ​​in the horizontal and vertical directions are calculated for the center pixel of the window using the third-order inverse distance squared weighted difference formula. The composite vector magnitude is calculated based on the elevation gradient values ​​in the horizontal and vertical directions, and the composite vector magnitude is converted into slope angle values ​​using the arctangent function to generate a global terrain slope map.

[0031] Specifically, after obtaining the 3D dense point cloud generated in step S2, in order to accurately obtain the bare soil elevation information reflecting the real terrain undulations, the point cloud is first classified using a cloth simulation filtering algorithm, classifying all points in the 3D dense point cloud. The elevation coordinates of the axes are set to negative values, which means flipping the terrain, turning undulating peaks into valleys, and setting a virtual cloth grid with a preset grid resolution above the flipped point cloud.

[0032] Force analysis and position update iterations are performed on each node in the virtual cloth grid. In each iteration, the node is mainly subjected to gravity, resulting in downward displacement. The unconstrained position update formula is as follows: In the formula, This represents the updated node position in the current iteration step. This represents the node position from the previous iteration. The vector of gravitational acceleration. Using the time step, after calculating the unconstrained displacement, an internal force constraint check is immediately performed. If the distance between adjacent nodes exceeds the threshold defined by the cloth stiffness coefficient, the node position is moved to maintain the rigidity of the cloth. At the same time, collision detection is performed. If a node passes through the flipped terrain point cloud surface, its position is forcibly fixed on the terrain surface and marked as immovable. When the maximum movement distance of all nodes is less than the convergence threshold, the iteration stops.

[0033] At this point, the vertical distance from each point in the original point cloud to the converged cloth grid is calculated. ,like If the ground thickness is less than a preset threshold, the point is marked as a ground point, and the remaining points are marked as non-ground points. Using the separated ground points, an inverse distance weighted interpolation algorithm is used to generate an initial digital terrain model, which is then resampled to the standard resolution required for monitoring. To eliminate the influence of micro-terrain noise on slope calculation, a Gaussian kernel function is used to perform convolution smoothing on the resampled digital terrain model. Defined as In the formula, and It is the distance from the center of the convolution. This represents the standard deviation, used to control the degree of smoothness.

[0034] Based on the smoothed digital terrain model, the global slope is calculated using a third-order inverse distance squared weighted difference algorithm for any center pixel in the model. its surroundings The pixel elevations within the neighborhood are denoted as follows: ,in The top left corner The bottom right corner shows the horizontal elevation gradient. The calculation formula is Vertical elevation gradient The calculation formula is In the formula, Given the grid cell size, the gradient in both directions is used to calculate the composite vector magnitude, which in turn yields the slope angle value. The calculation formula is: After traversing all pixels to complete the calculation, a global terrain slope map is generated.

[0035] By using fabric simulation filtering, the interference from surface vegetation and artificial facilities was effectively removed, accurately restoring the true bare soil morphology of hilly and mountainous areas. Furthermore, through multi-scale smoothing and weighted difference algorithms, the problem of inflated slope values ​​caused by micro-topographic noise was avoided, providing accurate quantitative basis for subsequent determination of whether there is steep slope reclamation behavior in cultivated land.

[0036] S4. Convert the true image to the Lab color space, calculate the shadow threshold using the luminance channel histogram features to obtain a binarized shadow mask, and apply the limited contrast adaptive histogram equalization algorithm to adjust the pixel grayscale values ​​only for the area covered by the binarized shadow mask to generate an enhanced image. Furthermore, in S4, calculating the shadow threshold using the luminance channel histogram features to obtain the binarized shadow mask specifically includes the following steps: Extract the luminance channel components in the Lab color space and calculate the grayscale histogram of the luminance channel components; Based on the method of maximum inter-class variance, the gray levels of the gray histogram are traversed, and the inter-class variance value is calculated when the image is divided into foreground and background classes. The gray level corresponding to the maximum inter-class variance is selected as the segmentation threshold. The brightness channel components are compared with the segmentation threshold pixel by pixel. Regions with pixel values ​​lower than the segmentation threshold are marked as shadow regions, and a binarized shadow mask is generated.

[0037] Specifically, in order to improve the texture details of the true image generated in step S2 in low-light areas such as shaded slopes, the true image is first converted from the RGB color space to the CIELAB color space, and the luminance channel component is separated. Perform grayscale statistics on the luminance channel components and calculate the normalized histogram. The calculation formula is: In the formula, The brightness value is The number of pixels, This represents the total number of pixels in the image. The value range is 0 to 255. Based on the statistical grayscale histogram, the maximum inter-class variance method is used to automatically calculate the shadow segmentation threshold.

[0038] Assuming the histogram is set at the brightness level The data is divided into two classes: background and foreground. The variance between classes is calculated. The calculation formula is: In the formula, The probability of shadow appearance is calculated using the following formula: ; The average brightness of the shadow class is calculated using the following formula: ; and Let the occurrence probability and average brightness of non-shadow classes be represented, and then iterate through them. Choose the value that makes Reaching the maximum value As the final segmentation threshold The luminance channel is binarized using this threshold to generate a binarized shadow mask. The brightness value is less than The position is marked as 1, and the rest are marked as 0.

[0039] For masks The shaded areas marked as 1 are enhanced using a contrast-limited adaptive histogram equalization algorithm. This involves dividing the image into continuous rectangular patches, calculating the local histogram for each patch, and setting a truncation threshold. Limiting the peak value of the histogram will exceed... The number of pixels is evenly distributed across the other gray levels of the histogram to obtain the corrected histogram. Based on the corrected histogram, a cumulative distribution function is constructed as a grayscale mapping curve to transform the pixels within the image patch. The transformation formula is as follows: In the formula, This is the original brightness value. The transformed brightness value. Maximum gray level This represents the total number of pixels within the current tile. For the corrected histogram in grayscale To eliminate block artifacts, bilinear interpolation was used to smoothly blend the pixel values ​​at the tile boundaries. Finally, the enhanced luminance channel was merged with the original... , The chroma channels are merged and inversely transformed back to the RGB color space to generate visually enhanced images.

[0040] This step utilizes statistical features to automatically locate black hole regions in images and specifically restores texture information within the shadows without disrupting the color balance of the bright areas. This effectively solves the problem of heterogeneous spectra of similar objects caused by uneven lighting in hilly and mountainous areas, and improves the accuracy of subsequent AI models in recognizing farmland on shady slopes.

[0041] S5. The pre-trained semantic segmentation neural network model of the enhanced image input is used to output pixel-level land feature classification results. Morphological operations and connected component extraction are performed on the pixel-level land feature classification results to generate land parcel vector plots containing land feature attributes. Furthermore, in S5, the semantic segmentation neural network model pre-trained from the image input is enhanced, and the pixel-level land cover classification results are output, specifically including the following steps: The enhanced images are standardized and their dimensions are adjusted to fit the input dimension of the semantic segmentation neural network model. The encoder of the semantic segmentation neural network model extracts deep semantic features and shallow spatial features of the image, and the decoder performs multi-scale fusion and upsampling of the deep semantic features and shallow spatial features. Output a multi-channel probability map with the same resolution as the input image, where each channel corresponds to a land cover category; For each pixel location, the category index corresponding to the channel with the highest probability value is selected as the classification label for that pixel, generating pixel-level land feature classification results.

[0042] Specifically, after obtaining the plot vector map containing land feature attributes generated in step S5 and the global topographic slope map generated in step S3, the suitability for cultivation of each plot is quantitatively evaluated using a zonal statistical analysis method. First, the plot vector map is overlaid on the global topographic slope map, and then a polygon rasterization algorithm is used to determine the value of each vector map. The set of pixels of the slope map covered For sets Extract the slope value of each pixel in the array. And calculate the average slope of the plot. The formula for calculating the average slope is: In the formula, This represents the total number of pixels covered by the plot. coordinates The slope angle value at that location.

[0043] After completing the geometric calculations, extract the land feature attributes of each plot's vector map patch. And compare this attribute with the calculated average slope value. The system inputs a preset farmland protection rule tree for automated logic judgment. The rule tree execution logic is as follows: First, it checks the land feature attributes. ,like If it belongs to the category of grain crops such as rice, corn, and wheat, then the average slope value will be further calculated. With the preset slope threshold In this embodiment, a comparison is performed. Set to 25 degrees, if If the land is cultivated on a steep slope, a soil erosion risk warning is triggered; if If it is, then it is determined to be compliant farmland, if the land features are... If the land belongs to any of the following categories: woodland, orchard, or pond / water surface, and its historical planning attribute is permanent basic farmland, then it is directly determined that the land has been occupied for non-grain planting. If the land feature attribute... If the land parcel is classified as a building, road, or paved surface, it is directly determined that there is non-agricultural construction encroachment.

[0044] For plots of land identified as having a risk of violation, their spatial location, plot number, violation type, and relevant evidence parameters are written into the monitoring results database, and a thematic layer is generated. This step achieves three-dimensional supervision of farmland use by spatially overlaying and logically coupling two-dimensional semantic information with three-dimensional topographic information. It can quickly and accurately screen out hidden violations in deep mountains and dense forests, providing objective data support for land law enforcement and ecological restoration.

[0045] S6. Using the land parcel vector plots as spatial statistical units, perform zonal statistical analysis on the overall topographic slope map, calculate the average slope value within each land parcel, and combine the land feature attributes with the average slope value to make logical judgments based on a preset rule tree, and output the monitoring results.

[0046] Furthermore, in S6, the land feature attributes and average slope values ​​are combined with a preset rule tree for logical judgment, and the output monitoring results specifically include the following steps: Determine whether the land cover is a food crop. If so, further compare the average slope value with the preset slope threshold. When the average slope value is greater than the slope threshold, it is determined to be a steep slope cultivation state. Determine whether the land feature is any of the following: forest, orchard, or pond / water surface. If so, it is directly determined to be occupied by non-grain planting. Determine whether the land feature is any of the following: building, road, or paved surface. If so, it is directly determined to be occupied by non-agricultural construction.

[0047] Specifically, after obtaining the plot vector map containing land feature attributes generated in step S5 and the global topographic slope map generated in step S3, a zonal statistical analysis method is used to quantitatively evaluate the cultivation suitability of each plot. First, the plot vector map is overlaid on the global topographic slope map. A polygon rasterization algorithm is used to determine the set of slope map pixels covered by each vector map. For all pixels in the set, the corresponding slope value is extracted, and the average slope value of the plot is calculated. The formula for calculating the average slope is: In the formula, Indicates the first The average slope of each plot of land Indicates the first The total number of pixels in the slope map covering each plot of land. Indicates the first within the scope of this plot of land The slope angle value of each pixel.

[0048] After completing the geometric calculations, extract the land feature attributes of each plot's vector map patch. And compare this attribute with the calculated average slope value. The preset rule tree is input together for automated logic judgment. The rule tree execution logic is as follows: First, check the land feature attributes. ,like If the crop falls under the category of rice, corn, or wheat, then the average slope value will be further calculated. With the preset slope threshold In this embodiment, a comparison is performed. Set to 25 degrees, when At that time, it was determined that the plot of land was being cultivated on a steep slope, triggering a soil erosion risk warning. If the land features... If the land belongs to any of the following categories: woodland, orchard, or pond / water surface, then the land is directly deemed to be occupied by non-grain planting. If the land feature attribute... If the land parcel is classified as any of the following: buildings, roads, or paved surfaces, it is directly determined that there is non-agricultural construction encroachment on the land.

[0049] For plots of land identified as having a risk of violation, their spatial location, plot number, violation type, and relevant evidence parameters are written into the monitoring results database, and a thematic layer is generated. This step achieves three-dimensional supervision of farmland use by spatially overlaying and logically coupling two-dimensional semantic information with three-dimensional topographic information. It can quickly and accurately screen out hidden violations in deep mountains and dense forests, providing objective data support for land law enforcement and ecological restoration.

[0050] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for monitoring and protecting cultivated land in hilly and mountainous areas by integrating low-altitude AI sensing technology, characterized in that, Includes the following steps: S1. Acquire the original aerial image sequence and attitude position information of the monitoring area with a preset overlap rate; S2. Based on the original aerial image sequence and attitude position information, a three-dimensional dense point cloud is constructed using a multi-view stereo matching algorithm, and the visibility of the three-dimensional dense point cloud is analyzed to remove the texture mapping of the projected occlusion area and generate a true image. S3. Perform cloth simulation filtering on the three-dimensional dense point cloud, use the iterative settlement process of the virtual cloth model under gravity to separate ground points and non-ground points, use the separated ground points to generate a digital terrain model by interpolation, and after downsampling and smoothing the digital terrain model, use the difference algorithm to calculate the global terrain slope map. S4. Convert the true image to the Lab color space, calculate the shadow threshold using the luminance channel histogram features to obtain a binarized shadow mask, and apply the limited contrast adaptive histogram equalization algorithm to adjust the pixel grayscale values ​​only for the area covered by the binarized shadow mask to generate an enhanced image. S5. Input the enhanced image into a pre-trained semantic segmentation neural network model, output pixel-level land feature classification results, and perform morphological operations and connected component extraction on the pixel-level land feature classification results to generate a land parcel vector plot containing land feature attributes. S6. Using the land parcel vector plot as a spatial statistical unit, perform zonal statistical analysis on the global topographic slope map, calculate the average slope value within the range of each land parcel, and combine the land feature attributes with the average slope value to make logical judgments using a preset rule tree, and output the monitoring results.

2. The method for monitoring and protecting cultivated land in hilly and mountainous areas integrating low-altitude AI sensing technology according to claim 1, characterized in that, In S2, the construction of a 3D dense point cloud using a multi-view stereo matching algorithm specifically includes the following steps: Feature points are extracted and matched from the original aerial image sequence, and the camera intrinsic parameters, extrinsic parameters, and sparse point cloud are solved using bundle adjustment. Based on the solved camera intrinsic and extrinsic parameters and sparse point cloud, a semi-global matching algorithm is used to calculate the depth map of each image. The depth map is subjected to multi-view consistency verification and fusion to remove redundant points and outlier noise points, generating a three-dimensional dense point cloud.

3. The method for monitoring and protecting cultivated land in hilly and mountainous areas integrating low-altitude AI sensing technology according to claim 1, characterized in that, In S2, generating the true image specifically includes the following steps: A digital surface model is constructed based on the aforementioned three-dimensional dense point cloud; Establish an orthographic projection grid, and for each pixel unit in the grid, construct a line-of-sight vector from the ground point to the camera center; Detect whether there is an intersection relationship between the gaze vector and the digital surface model. If they intersect, mark it as an occluded area; if they do not intersect, mark it as a visible area. Traverse the original aerial image sequence, select the image texture corresponding to the visible area with the smallest projection angle to fill the pixel unit, and generate a true image.

4. The method for monitoring and protecting cultivated land in hilly and mountainous areas by integrating low-altitude AI sensing technology according to claim 1, characterized in that, In S3, the process of performing cloth simulation filtering on the three-dimensional dense point cloud specifically includes the following steps: The elevation coordinates of the three-dimensional dense point cloud are inverted to construct a flipped three-dimensional scene; A virtual cloth grid with a preset grid resolution and stiffness coefficient is initialized above the flipped 3D scene; Force simulation and iterative calculation are performed on the nodes in the virtual fabric grid. In each iteration, the downward displacement of the nodes is calculated based on the effect of gravity, and the node positions are adjusted according to the internal force constraints between the nodes until the fabric grid shape converges and stabilizes. Calculate the vertical distance from each point in the three-dimensional dense point cloud to the converged cloth grid, and mark the points whose vertical distance is less than a preset threshold as ground points, and the rest as non-ground points.

5. The method for monitoring and protecting cultivated land in hilly and mountainous areas integrating low-altitude AI sensing technology according to claim 1, characterized in that, In S3, the calculation of the global terrain slope map using the difference algorithm specifically includes the following steps: The digital terrain model is resampled to a preset spatial resolution using a bilinear interpolation algorithm, and the resampled digital terrain model is then smoothed using a Gaussian kernel function. Using a sliding window to traverse the smoothed digital terrain model, the elevation gradient values ​​in the horizontal and vertical directions are calculated for the center pixel of the window using the third-order inverse distance squared weighted difference formula. The composite vector magnitude is calculated based on the elevation gradient values ​​in the horizontal and vertical directions, and the composite vector magnitude is converted into a slope angle value using the arctangent function to generate a global terrain slope map.

6. The method for monitoring and protecting cultivated land in hilly and mountainous areas integrating low-altitude AI sensing technology according to claim 1, characterized in that, In S4, the step of calculating the shadow threshold using the luminance channel histogram features to obtain the binarized shadow mask specifically includes the following steps: Extract the luminance channel components in the Lab color space and calculate the grayscale histogram of the luminance channel components; The gray levels of the gray histogram are traversed using the maximum inter-class variance method to calculate the inter-class variance value when the image is segmented into foreground and background classes. The gray level corresponding to the maximum inter-class variance is selected as the segmentation threshold; The brightness channel components are compared pixel by pixel with the segmentation threshold, and regions with pixel values ​​lower than the segmentation threshold are marked as shadow regions to generate a binarized shadow mask.

7. The method for monitoring and protecting cultivated land in hilly and mountainous areas integrating low-altitude AI sensing technology according to claim 1, characterized in that, In S5, the step of inputting the enhanced image into a pre-trained semantic segmentation neural network model and outputting pixel-level land cover classification results specifically includes the following steps: The enhanced image is standardized, and the image size is adjusted to fit the input dimension of the semantic segmentation neural network model; The encoder of the semantic segmentation neural network model extracts deep semantic features and shallow spatial features of the image, and the decoder performs multi-scale fusion and upsampling on the deep semantic features and shallow spatial features. Output a multi-channel probability map with the same resolution as the input image, where each channel corresponds to a land cover category; For each pixel location, the category index corresponding to the channel with the highest probability value is selected as the classification label for that pixel, generating pixel-level land feature classification results.

8. The method for monitoring and protecting cultivated land in hilly and mountainous areas integrating low-altitude AI sensing technology according to claim 1, characterized in that, In S6, the step of combining the land feature attributes with the average slope value using a preset rule tree for logical determination and outputting the monitoring results specifically includes the following steps: Determine whether the land feature attribute is a food crop. If so, further compare the average slope value with a preset slope threshold. When the average slope value is greater than the slope threshold, it is determined to be a steep slope cultivation state. Determine whether the land feature is any one of forest, orchard, or pond / water surface. If so, it is directly determined to be occupied by non-grain planting. Determine whether the land feature is any of the following: building, road, or paved surface. If so, it is directly determined to be occupied by non-agricultural construction.

Images