A method, system, device, and medium for micro-topographic region identification based on segmentation large model and skeleton wavelet features.

Abstract

A method, system, device, and medium for micro-topographic region identification based on a large segmentation model and skeleton wavelet features include: acquiring the original image of the region to be identified; inputting the original image into a pre-constructed terrain semantic segmentation network, performing semantic classification on each pixel in the image to generate a terrain semantic segmentation map, and performing morphological opening operations to obtain classification masks; based on each classification mask, performing skeleton extraction operations, decomposing the skeleton image into connected components to obtain terrain skeleton line segments, and extracting geometric and topological feature vectors; forming skeleton neighborhood wavelet feature vectors based on the terrain skeleton line segments combined with two-dimensional discrete wavelet transform; concatenating the geometric and topological feature vectors and the skeleton neighborhood wavelet feature vectors to form a comprehensive feature vector for each skeleton line segment, and identifying the micro-topographic region category based on the comprehensive feature vector. This invention can effectively improve the automatic identification capability of typical micro-topography and provide terrain identification technology support for railway engineering planning and design.

Description

Technical Field

[0001] This invention relates to the field of computer vision and image processing technology, specifically to a method, system, device, and medium for micro-topographic region recognition based on segmentation large model and skeleton wavelet features. Background Technology

[0002] Micro-topography identification has significant application value in multiple technical fields such as geological disaster early warning, land spatial planning, agricultural production layout, and ecological environment monitoring. Its core requirement is to accurately capture the topographic undulations and morphological features of the earth's surface at a small scale (usually on the order of meters to ten meters), providing data support for subsequent risk assessment and scheme design. In existing technologies, micro-topography identification mainly relies on digital elevation models (DEMs) for basic analysis, supplemented by manual interpretation and verification of remote sensing images and topographic data. However, this technical solution has gradually revealed many technical shortcomings that urgently need to be addressed in practical applications.

[0003] First, the over-reliance on digital elevation models in existing technologies limits the applicability and accuracy of identification. On the one hand, the accuracy of digital elevation models directly determines the effectiveness of micro-topography identification. However, acquiring high-precision digital elevation models is costly and time-consuming, making it difficult to achieve comprehensive coverage and timely updates in complex scenarios such as remote mountainous areas and densely vegetated regions. On the other hand, conventional digital elevation models are insufficient in depicting subtle surface undulations and are easily affected by factors such as data acquisition resolution and terrain filtering algorithms, leading to the omission or misjudgment of some gentle and hidden micro-topographic features.

[0004] Secondly, the manual interpretation process suffers from significant drawbacks, including high workload and strong subjectivity. Due to the small scale, diverse shapes, and scattered distribution of micro-topographic features, large-scale micro-topographic identification tasks require technicians to manually screen, label, and classify topographic data and imagery materials region by region. This not only consumes substantial manpower and time, resulting in low identification efficiency and failing to meet the demands of large-scale, rapid applications, but more importantly, the results of manual interpretation heavily rely on the professional experience and subjective judgment of the operators. Differences in the understanding and classification standards of different personnel regarding the same micro-topographic feature can easily lead to interpretation biases, making it difficult to guarantee the consistency and reliability of the identification results, thereby affecting the scientific validity of subsequent decisions.

[0005] Furthermore, existing technologies are mostly based on coarse-grained geomorphological classification systems for micro-topography identification, which cannot meet the needs of refined applications. Current mainstream identification methods often classify terrain according to macro-topographical types (such as mountains, hills, plains, etc.), failing to construct refined classification standards and identification models for the specific morphology, formation, and functional attributes of micro-topography. This coarse-grained classification method makes it difficult to distinguish between special micro-topography with disaster risks (such as gullies, landslide steps, sinkholes, and weak interlayer outcrops) and ordinary micro-topography, and cannot accurately locate disaster-prone hazard points.

[0006] Finally, existing technologies generally lack the ability to perform detailed modeling of typical disaster-causing micro-topography. The formation and evolution of typical disaster-causing micro-topography are complex and dynamic, and its morphological parameters (such as slope, aspect, undulation, and extension length) are directly related to disaster risk. However, existing technologies can only achieve preliminary contour identification of micro-topography, failing to construct detailed models that encompass its three-dimensional morphology, spatial distribution patterns, and evolutionary trends. This makes it impossible to accurately quantify the hazard level of disaster-causing micro-topography, resulting in a lack of precise technical support for subsequent disaster risk assessment and early warning forecasting, and hindering effective guidance for disaster prevention and mitigation efforts.

[0007] In summary, existing micro-topography recognition technologies suffer from problems such as reliance on digital elevation models, significant deficiencies in manual interpretation, coarse classification granularity, and a lack of fine modeling capabilities for disaster-causing micro-topography. Consequently, their recognition accuracy, efficiency, and reliability cannot meet the refined and precise requirements of practical applications. Summary of the Invention

[0008] To address the problems of existing technologies in micro-topography identification, such as reliance on digital elevation models, heavy manual interpretation and high subjectivity, and reliance on coarse-grained landform classification without fine-grained modeling of typical disaster-causing micro-topography, this invention proposes a micro-topography region identification method based on segmentation large models and skeleton wavelet features, including: Obtain the original image of the region to be identified; The original image is input into a pre-constructed terrain semantic segmentation network. Semantic classification is performed on each pixel in the image to generate a terrain semantic segmentation map. Morphological opening operations are then performed on the terrain semantic segmentation map to obtain the classification masks. Based on each classification mask, a skeleton extraction operation is performed to obtain a skeleton image. The skeleton image is then decomposed into connected components to obtain terrain skeleton line segments, and geometric and topological feature vectors are extracted. Based on the terrain skeleton line segments combined with two-dimensional discrete wavelet transform, a skeleton neighborhood wavelet feature vector is formed to characterize the neighborhood multi-scale texture characteristics of the terrain skeleton line segments. Geometric and topological feature vectors and skeleton neighborhood wavelet feature vectors are concatenated to form a comprehensive feature vector for each skeleton line segment, and the micro-topography region category is identified based on the comprehensive feature vector. The pre-built terrain semantic segmentation network is obtained by fine-tuning the large segmentation model SAM.

[0009] Preferably, after acquiring the original image of the region to be identified, the method further includes: Perform grayscale conversion on the original color image to convert the three-channel image into a single-channel grayscale image. The grayscale image is smoothed and denoised using the discrete Gaussian filtering method to obtain the denoised image. The denoised image is then normalized to obtain the processed image.

[0010] Preferably, the step of inputting the original image into a pre-constructed terrain semantic segmentation network, performing semantic classification on each pixel in the image, and generating a terrain semantic segmentation map includes: The original image is input into a pre-built terrain semantic segmentation network to obtain the response value of each pixel in each terrain category; Perform a Softmax operation on all categories at each pixel location to obtain the probability that the pixel belongs to a certain category; Based on the probability that a pixel belongs to a certain category, the terrain category to which the pixel belongs is determined using the maximum probability criterion; Generate a terrain semantic segmentation map based on the terrain category to which the pixel belongs.

[0011] Preferably, determining the terrain category to which a pixel belongs based on the probability that the pixel belongs to a certain category, using the maximum probability criterion, includes: The semantic category label index of a pixel is calculated based on the maximum probability of the pixel belonging to each category, combined with the semantic category label index calculation formula. The terrain category to which the pixel belongs is obtained based on the mapping relationship between the semantic category label index and each terrain category.

[0012] Preferably, the step of performing morphological opening operations on the terrain semantic segmentation map to obtain classification masks includes: Convert the terrain semantic segmentation map into binary masks for each category; Perform a morphological opening operation on the binary mask to obtain smoothed classification masks.

[0013] Preferably, the step of forming a skeleton neighborhood wavelet feature vector characterizing the neighborhood multi-scale texture characteristics of the terrain skeleton line segment based on the terrain skeleton line segment combined with two-dimensional discrete wavelet transform includes: For each ridge line segment, valley line segment, and water boundary line segment, a strip-shaped neighborhood region is constructed on the normalized image along the skeleton tangential and normal directions. Two-dimensional discrete wavelet transform is performed on the local image patch of the skeleton neighborhood to obtain wavelet sub-band coefficients of multiple scales and multiple directions. Statistical features are calculated on wavelet subbands at various scales and in various directions, including energy and entropy. The statistical features of each local stripe are aggregated along the same skeleton line segment to form a skeleton neighborhood wavelet feature vector that characterizes the multi-scale texture properties of the neighborhood of the skeleton line segment.

[0014] Preferably, the process of performing skeleton extraction based on each classification mask to obtain a skeleton image, performing connected component decomposition on the skeleton image to obtain terrain skeleton line segments, and extracting geometric and topological feature vectors includes: Based on the ridge candidate region mask and the valley candidate region mask, a skeletonization operator is obtained using an iterative refinement-based morphological algorithm. Based on the skeletonization operator combined with the binary image pixel values ​​at the pixel, ridgeline skeleton images and valley skeleton images are obtained. An image of the water body boundary skeleton is obtained by morphological boundary extraction based on a water body region mask; Connected component decomposition is performed on the skeleton image to divide the skeleton into several independent line segments, resulting in terrain skeleton line segments; Each skeleton line segment is represented as a sequence of ordered pixel coordinates, and the total number of pixels, pixel coordinates, skeleton point set, and geometric length of the line segment are recorded.

[0015] Furthermore, this invention also provides a micro-topographic region recognition system based on segmentation large model and skeleton wavelet features, comprising: The image preprocessing module is used to obtain the original image of the region to be identified; The terrain semantic segmentation module is used to input the original image into a pre-constructed terrain semantic segmentation network, perform semantic classification on each pixel in the image, generate a terrain semantic segmentation map, and perform morphological opening operations on the terrain semantic segmentation map to obtain the classification masks. The skeleton extraction module is used to perform skeleton extraction operations based on each classification mask to obtain skeleton images, and to perform connected component decomposition on the skeleton images to obtain terrain skeleton line segments, and extract geometric and topological feature vectors. The skeleton domain wavelet feature extraction module is used to form a skeleton neighborhood wavelet feature vector that characterizes the neighborhood multi-scale texture characteristics of the terrain skeleton line segment based on the terrain skeleton line segment combined with two-dimensional discrete wavelet transform. The micro-topography classification module is used to concatenate geometric and topological feature vectors and skeleton neighborhood wavelet feature vectors to form a comprehensive feature vector for each skeleton line segment, and to identify the micro-topography region category based on the comprehensive feature vector. The pre-built terrain semantic segmentation network is obtained by fine-tuning the large segmentation model SAM.

[0016] Preferably, the terrain semantic segmentation module includes: The semantic segmentation submodule is used to: input the original image into a pre-constructed terrain semantic segmentation network to obtain the response value of each pixel in each terrain category; perform a Softmax operation on all categories at each pixel location to obtain the probability that the pixel belongs to a certain category; determine the terrain category to which the pixel belongs based on the probability that the pixel belongs to a certain category using the maximum probability criterion; and generate a terrain semantic segmentation map based on the terrain category to which the pixel belongs. The smoothing submodule is used to convert the terrain semantic segmentation map into binary masks of various categories; and to perform morphological opening operations on the binary masks to obtain smoothed masks for each category.

[0017] Preferably, the skeleton domain wavelet feature extraction module is specifically used for: For each ridge line segment, valley line segment, and water boundary line segment, a strip-shaped neighborhood region is constructed on the normalized image along the skeleton tangential and normal directions. Two-dimensional discrete wavelet transform is performed on the local image patch of the skeleton neighborhood to obtain wavelet sub-band coefficients of multiple scales and multiple directions. Statistical features are calculated on wavelet subbands at various scales and in various directions, including energy and entropy. The statistical features of each local stripe are aggregated along the same skeleton line segment to form a skeleton neighborhood wavelet feature vector that characterizes the multi-scale texture properties of the neighborhood of the skeleton line segment.

[0018] Preferably, the skeleton extraction module is specifically used for: Based on the ridge candidate region mask and the valley candidate region mask, a skeletonization operator is obtained using an iterative refinement-based morphological algorithm. Based on the skeletonization operator combined with the binary image pixel values ​​at the pixel, ridgeline skeleton images and valley skeleton images are obtained. An image of the water body boundary skeleton is obtained by morphological boundary extraction based on a water body region mask; Connected component decomposition is performed on the skeleton image to divide the skeleton into several independent line segments, resulting in terrain skeleton line segments; Each skeleton line segment is represented as a sequence of ordered pixel coordinates, and the total number of pixels, pixel coordinates, skeleton point set, and geometric length of the line segment are recorded.

[0019] Preferably, the specific steps in the semantic segmentation submodule for determining the terrain category to which a pixel belongs based on the probability of the pixel belonging to a certain category, using the maximum probability criterion, include: The semantic category label index of a pixel is calculated based on the maximum probability of the pixel belonging to each category, combined with the semantic category label index calculation formula. The terrain category to which the pixel belongs is obtained based on the mapping relationship between the semantic category label index and each terrain category.

[0020] Preferably, the specific implementation steps of performing morphological opening operations on the terrain semantic segmentation map in the semantic segmentation submodule to obtain each classification mask include: Convert the terrain semantic segmentation map into binary masks for each category; Perform a morphological opening operation on the binary mask to obtain smoothed classification masks.

[0021] In another aspect, the present invention also provides an electronic device, comprising: at least one processor and a memory; the memory and the processor are connected via a bus; The memory is used to store one or more programs; When the one or more programs are executed by the at least one processor, a micro-topographic region identification method based on segmentation large model and skeleton wavelet features is implemented as described above.

[0022] In another aspect, the present invention also provides a computer-readable storage medium having an executable program stored thereon, which, when executed, implements the micro-topographic region identification method based on segmentation large model and skeleton wavelet features as described above.

[0023] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention provides a micro-topographic region recognition method based on a large segmentation model and skeleton wavelet features, comprising: acquiring the original image of the region to be identified; inputting the original image into a pre-constructed terrain semantic segmentation network, performing semantic classification on each pixel in the image to generate a terrain semantic segmentation map, and performing morphological opening operations on the terrain semantic segmentation map to obtain classification masks; performing skeleton extraction operations based on each classification mask to obtain a skeleton image, and performing connected component decomposition on the skeleton image to obtain terrain skeleton line segments, and extracting geometric and topological feature vectors; forming skeleton neighborhood wavelet feature vectors representing the neighborhood multi-scale texture characteristics of the terrain skeleton line segments based on the terrain skeleton line segments combined with two-dimensional discrete wavelet transform; and concatenating the geometric and topological feature vectors and the skeleton neighborhood wavelet feature vectors to form a comprehensive feature vector for each skeleton line segment. This invention provides a micro-topographic region identification method based on image data, while fully utilizing the topographic skeleton structure and neighborhood multi-scale texture information. It can effectively improve the automatic identification capability of typical micro-topography such as mountain passes, watersheds, topographic uplift, canyons, and water bodies, providing reliable micro-topographic identification technology support for disaster prevention and mitigation and engineering planning and design of power transmission lines. It solves the problems of existing technologies, such as reliance on digital elevation models for micro-topographic identification, large workload and strong subjectivity of manual interpretation, and lack of fine modeling of typical disaster-causing micro-topography based only on coarse-grained landform classification. Attached Figure Description

[0024] Figure 1 This is a flowchart of the micro-topographic region identification method based on segmentation large model and skeleton wavelet features of the present invention; Figure 2 A simplified flowchart of the micro-topographic region identification method based on segmentation large model and skeleton wavelet features of the present invention; Figure 3 This is a flowchart of the five-class micro-terrain classification driven by skeleton geometry and wavelet texture of the present invention. Figure 4 This is a schematic diagram of an electronic device structure according to the present invention. Detailed Implementation

[0025] This invention proposes a micro-topographic region identification method based on segmentation large model and skeleton wavelet features. Relying solely on image data, it utilizes techniques such as segmentation large model, terrain skeleton extraction, and skeleton neighborhood wavelet texture analysis to achieve automatic classification and identification of five typical micro-topographic regions, including mountain passes, watersheds, terrain uplift, canyons, and water bodies. This improves the precision and automation level of micro-topographic identification, providing high-quality micro-topographic basic data for disaster prevention and mitigation of power transmission lines and related engineering applications.

[0026] This method is applicable to the automatic classification and identification of typical micro-topographic areas such as mountain passes, watersheds, terrain uplift, canyons, and water bodies. It is especially suitable for disaster prevention and mitigation along power grid transmission line corridors and related engineering applications.

[0027] To better understand the present invention, the following description, in conjunction with the accompanying drawings and embodiments, will further illustrate the content of the present invention.

[0028] Example 1: A micro-topographic region identification method based on segmentation large model and skeleton wavelet features, such as Figure 1 As shown, it includes: Step 1: Obtain the original image of the region to be identified; Step 2: Input the original image into the pre-constructed terrain semantic segmentation network, perform semantic classification on each pixel in the image to generate a terrain semantic segmentation map, and perform morphological opening operation on the terrain semantic segmentation map to obtain the classification mask; Step 3: Based on each classification mask, perform skeleton extraction to obtain skeleton images, and perform connected component decomposition on the skeleton images to obtain terrain skeleton line segments, and extract geometric and topological feature vectors. Step 4: Based on the terrain skeleton line segments, combine two-dimensional discrete wavelet transform to form a skeleton neighborhood wavelet feature vector that characterizes the neighborhood multi-scale texture characteristics of the terrain skeleton line segments. Step 5: Concatenate the geometric and topological feature vectors and the skeleton neighborhood wavelet feature vectors to form a comprehensive feature vector for each skeleton line segment, and identify the micro-topographic region category based on the comprehensive feature vector; The pre-built terrain semantic segmentation network is obtained by fine-tuning the large segmentation model SAM.

[0029] This invention discloses an image-based method for identifying five types of micro-topographic regions, applicable to complex terrain scenarios such as power transmission line corridors. The method includes the following steps: S1, acquiring the original image; S2, terrain semantic segmentation, inputting the original image into a terrain semantic segmentation network fine-tuned based on a large segmentation model (SAM), classifying each pixel to obtain segmentation results for ridge candidate regions, valley candidate regions, water bodies, and background regions; S3, skeleton extraction, performing skeletonization operations on the ridge and valley candidate regions to obtain one-pixel-wide ridge and valley skeletons, and obtaining water body boundary skeletons through morphological boundary extraction, and performing connected component decomposition on each skeleton to generate a set of skeleton segments; S4, skeleton neighborhood wavelet feature extraction, constructing strip-shaped neighborhoods on both sides of each skeleton segment, and extracting wavelet features from the normalized neighborhoods. A two-dimensional discrete wavelet transform is performed on local strips in the skeleton neighborhood of the image to calculate the energy, entropy, and other statistics of sub-bands at each scale and direction, and these are aggregated along the skeleton lines to form wavelet texture feature vectors for the skeleton neighborhood. In step S5, based on a five-category micro-terrain classification using "geometric + wavelet texture features," geometric and topological feature vectors of the skeleton line segments are constructed and concatenated with the wavelet texture feature vectors to form comprehensive features. These features are input into a trained classification model, and the output is the classification results for five categories of micro-terrain corresponding to the skeleton line segments: mountain passes, watersheds, topographic uplift, canyons, and water bodies. This invention does not rely on a digital elevation model; it can automatically identify five types of micro-terrain based solely on images, offering advantages such as clear structure, strong interpretability, and high recognition accuracy.

[0030] This invention provides a method for identifying five types of micro-topography regions based on images, such as... Figure 2 As shown, the method includes at least the following steps: Step 1: Obtain the original image of the region to be identified, including: Obtain raw images of the area to be identified from drone aerial photography equipment, remote sensing satellites, or ground photography equipment.

[0031] Following step 1, the original image undergoes preprocessing, the details of which are as follows: S1, Image Preprocessing The original image undergoes preprocessing operations such as grayscale conversion, filtering and denoising, brightness and contrast normalization, geometric correction, and cropping to obtain a standardized preprocessed image. Preferably, Gaussian filters or wavelet thresholding methods can be used to suppress noise, and the large image is sliced ​​into several sub-images according to the transmission line corridor range to facilitate subsequent processing.

[0032] Step 2: Input the original image into the pre-constructed terrain semantic segmentation network, perform semantic classification on each pixel in the image to generate a terrain semantic segmentation map, and perform morphological opening operations on the terrain semantic segmentation map to obtain the classification masks, including: The original image is input into a pre-built terrain semantic segmentation network to obtain the response value of each pixel in each terrain category; Perform a Softmax operation on all categories at each pixel location to obtain the probability that the pixel belongs to a certain category; Based on the probability that a pixel belongs to a certain category, the terrain category to which the pixel belongs is determined using the maximum probability criterion; Generate a terrain semantic segmentation map based on the terrain category to which the pixel belongs; Convert the terrain semantic segmentation map into binary masks for each category; Perform a morphological opening operation on the binary mask to obtain smoothed classification masks.

[0033] Furthermore, based on the probability that a pixel belongs to a certain category, the terrain category to which the pixel belongs is determined using the maximum probability criterion, including: The semantic category label index of a pixel is calculated based on the maximum probability of the pixel belonging to each category, combined with the semantic category label index calculation formula. The terrain category to which the pixel belongs is obtained based on the mapping relationship between the semantic category label index and each terrain category.

[0034] Step 2 specifically includes: S2, terrain semantic segmentation. The original or processed image is input into a terrain semantic segmentation network based on the fine-tuned Segment Anything Model (SAM). Semantic classification is performed on each pixel in the image to generate a terrain semantic segmentation map.

[0035] The terrain semantic segmentation map includes at least ridge candidate regions, valley candidate regions, water body regions, and background regions. Preferably, post-processing such as morphological opening operations are performed on the semantic segmentation results to remove isolated noise points and fill small holes, resulting in ridge candidate region masks, valley candidate region masks, and water body region masks with better spatial connectivity.

[0036] Step 3: Based on each classification mask, perform skeleton extraction to obtain skeleton images, and then perform connected component decomposition on the skeleton images to obtain terrain skeleton line segments. Finally, extract geometric and topological feature vectors, including: Based on ridge candidate region masks and valley candidate region masks, a skeletonization operator is obtained using an iterative refinement-based morphological algorithm. The skeletonization operator is then combined with the binary image pixel values ​​at each pixel to obtain ridgeline skeleton images and valleyline skeleton images. Based on water body region masks, morphological boundary extraction is used to obtain water body boundary line skeleton images. Connected component decomposition is performed on the skeleton images to divide the skeleton into several independent line segments, resulting in terrain skeleton line segments. Each skeleton line segment is represented as a sequence of ordered pixel coordinates, recording the total number of pixels, pixel coordinates, skeleton point set, and the geometric length of the line segment.

[0037] Step 3 specifically includes: S3, skeleton extraction.

[0038] Based on various masks obtained from terrain semantic segmentation, skeleton extraction operations are performed: Ridge candidate region masks are refined or skeletonized to obtain a one-pixel-wide ridgeline skeleton; valley candidate region masks are refined or skeletonized to obtain a one-pixel-wide valleyline skeleton; and water body region masks are subjected to boundary extraction operations such as erosion and difference to obtain a water body boundary skeleton. Furthermore, the skeleton images are decomposed into connected components, dividing the skeleton into several independent ridgeline segments, valleyline segments, and water body boundary segments. Geometric and topological attributes such as skeleton point coordinates, segment length, direction, tortuosity, and node degree are extracted to construct a set of terrain skeleton segments.

[0039] Step 4: Based on the terrain skeleton line segments and combined with two-dimensional discrete wavelet transform, form a skeleton neighborhood wavelet feature vector representing the neighborhood multi-scale texture characteristics of the terrain skeleton line segments, including: For each ridge segment, valley segment, and water boundary segment, a strip-shaped neighborhood region is constructed along the tangential and normal directions of the skeleton on the normalized image. A two-dimensional discrete wavelet transform is performed on the local image patch of the skeleton neighborhood to obtain wavelet sub-band coefficients at multiple scales and in multiple directions. Statistical features are calculated on the wavelet sub-bands at each scale and in each direction, including energy and entropy. The statistical features of each local strip along the same skeleton segment are aggregated to form a skeleton neighborhood wavelet feature vector that characterizes the multi-scale texture properties of the neighborhood of the skeleton segment.

[0040] Step 4 specifically includes: S4, skeleton domain wavelet feature extraction. For each ridge segment, valley segment, and water boundary segment, strip-shaped neighborhood regions are constructed along the tangential and normal directions of the skeleton on the normalized image. Two-dimensional discrete wavelet transform is performed on the local image patches within the skeleton neighborhood to obtain wavelet sub-band coefficients at multiple scales and in multiple directions. Statistical features such as energy and entropy are calculated on the wavelet sub-bands at each scale and in each direction. The features of each local strip are then aggregated along the same skeleton segment to form a skeleton neighborhood wavelet feature vector characterizing the multi-scale texture properties of the skeleton segment's neighborhood.

[0041] Step 5: Concatenate the geometric and topological feature vectors and the skeleton neighborhood wavelet feature vectors to form a comprehensive feature vector for each skeleton line segment, and identify the micro-topographic region category based on the comprehensive feature vector, including: Geometric and topological feature vectors and skeleton neighborhood wavelet feature vectors are concatenated to form a comprehensive feature vector for each skeleton line segment. The comprehensive feature vector of each skeleton line segment is then input into a trained micro-topography classification model to obtain the probability of each micro-topography category. The micro-topography category corresponding to the highest probability is taken as the micro-topography region category of the skeleton line segment. The trained micro-topography classification model is obtained by training a supervised learning classification model using samples containing micro-topography category labels; the supervised learning classification model includes: multilayer perceptron, support vector machine, random forest or graph neural network.

[0042] Step 5 specifically includes: S5, five types of micro-terrain classification based on "geometric + wavelet texture features".

[0043] The geometric and topological feature vectors extracted from the skeleton are concatenated with the wavelet texture feature vectors extracted from the skeleton's neighborhood wavelet features to form the comprehensive feature vector of each skeleton line segment. Using training samples labeled with features such as mountain passes, watersheds, topographic uplift, canyons, and water bodies, a pre-defined supervised learning classification model (such as a multilayer perceptron, support vector machine, random forest, or graph neural network) is trained to obtain a micro-topography classification model. This embodiment uses five types of micro-topography as examples. In the actual recognition stage, the comprehensive feature vectors of each ridge line segment, valley line segment, and water body boundary line segment are input into the trained classification model to obtain the probability of each micro-topography category. Based on the category label corresponding to the highest probability, the category to which the line segment belongs—mountain pass, watershed, topographic uplift, canyon, or water body—is output, thus obtaining the five-category micro-topography classification results for all skeleton line segments. The classification results can be directly used as the output of the method of this invention for use by the upper-level disaster prevention and mitigation analysis system without the need for separate additional steps.

[0044] In a preferred embodiment of the above technical solution, the terrain semantic segmentation uses a SAM-based segmentation network as the basis of the encoder-decoder structure. By fine-tuning on the micro-topography annotation dataset, high-precision segmentation of ridge candidate regions, valley candidate regions, and water body regions is achieved.

[0045] In another preferred embodiment, the skeleton neighborhood wavelet feature extraction employs multi-level discrete wavelet decomposition, combined with various statistical features such as energy and entropy, to comprehensively characterize the texture differences of the skeleton neighborhood from different scales and directions, thereby enhancing the separability between micro-topography such as mountain passes, watersheds, topographic uplift, canyons, and water bodies.

[0046] Compared with the prior art, the present invention has the following beneficial effects: 1. Micro-topography recognition can be achieved solely based on images without relying on digital elevation models.

[0047] This invention uses UAV aerial images, remote sensing images, or on-site photographic images as data sources, and can automatically identify five types of micro-topography, namely mountain passes, watersheds, topographic uplift, canyons, and water bodies, without relying on elevation data such as DEM. It is applicable to areas where elevation data is missing or insufficient in accuracy, thus improving the applicability of the method.

[0048] 2. Improve the segmentation accuracy of key terrain regions based on large segmentation models.

[0049] By introducing a large segmentation model (SAM) and fine-tuning it in micro-topographic scenarios, the terrain semantic segmentation module of this invention can achieve high segmentation accuracy under limited labeled sample conditions, reliably extracting ridge candidate regions, valley candidate regions, and water body regions, laying a high-quality foundation for subsequent skeleton extraction and classification.

[0050] 3. By integrating skeleton geometry and topology with neighborhood wavelet texture, multi-dimensional and precise recognition can be achieved.

[0051] This invention further introduces wavelet multi-scale texture features of the skeleton neighborhood on the basis of the skeleton structure, and integrates the geometric and topological properties of line segments with neighborhood texture information, effectively distinguishing micro-terrain types with similar height undulations but different texture features, and improving the separability and recognition accuracy among five types of micro-terrain: mountain passes, watersheds, terrain uplift, canyons, and water bodies.

[0052] 4. The overall process is modular, making it easy to integrate and expand into engineering projects.

[0053] The method of this invention is designed in modules such as image preprocessing, terrain semantic segmentation, skeleton extraction, skeleton neighborhood wavelet feature extraction, and five types of micro-terrain classification. The interfaces between the modules are clear, which facilitates integration and deployment in existing image processing platforms and power grid GIS / disaster prevention assessment systems. At the same time, the specific network structure and classification model can be flexibly replaced or extended according to application requirements, which has good engineering feasibility and scalability.

[0054] In summary, this invention provides a method for identifying five types of micro-topographic regions based solely on image data while fully utilizing topographic skeleton structure and neighborhood multi-scale texture information. This method can effectively improve the automatic identification capability of typical micro-topographic features such as mountain passes, watersheds, topographic uplift, canyons, and water bodies, providing reliable micro-topographic identification technology support for disaster prevention and mitigation and engineering planning and design of power transmission lines.

[0055] Example 2 The following is combined with Figure 3The micro-topographic region identification system based on segmentation large model and skeleton wavelet features of the present invention will be further described.

[0056] S1) Image acquisition module and image preprocessing module The image acquisition module is used to acquire the original image of the region to be identified; The image preprocessing module is used to standardize and improve the quality of the acquired raw micro-topographic images, including steps such as image acquisition and cropping, grayscale conversion, noise reduction and smoothing, brightness and contrast normalization, providing stable input for subsequent terrain semantic segmentation and skeleton extraction.

[0057] 1. Image Acquisition and Coordinate Definition The system first acquires the original image of the area to be identified from drone aerial photography equipment, remote sensing satellites, or ground photography equipment, denoted as... ,in, Indicates pixel coordinates The original image pixel values ​​at that location; Represents the pixel coordinates of the image in the horizontal direction. ; Represents the pixel coordinates of the image in the vertical direction. ; Indicates the width of the image (in pixels); Indicates the height of the image (in pixels).

[0058] When the input is a color image, pixels The image value at that location can be represented as:

[0059] in, Represents pixels The red channel value at that location; Represents pixels The green channel value at that location; Represents pixels The blue channel value at that location.

[0060] To facilitate subsequent processing, the system can crop the large original image to a fixed size according to the spatial range of the transmission line corridor, resulting in several preprocessed sub-images. Each sub-image still uses... express.

[0061] 2. Grayscale processing For the original color image, the system first performs a grayscale conversion operation, transforming the three-channel image into a single-channel grayscale image. The grayscale conversion formula is:

[0062] in, Indicated in pixels The grayscale image pixel value at that location; , , These represent the weighting coefficients of the red, green, and blue channels in luminance composition, respectively; preferably, , , And satisfy .

[0063] The grayscale processing described above can reduce data dimensionality and computational complexity while preserving terrain brightness and texture information.

[0064] 3. Denoising and smoothing (Gaussian filtering) To reduce the impact of sensor noise and environmental interference on subsequent micro-topography recognition, the system processes grayscale images. Perform smoothing and denoising processing to obtain the denoised image. Preferably, a discrete Gaussian filter is used, and the denoising process can be expressed as:

[0065] in, Indicated in pixels The pixel values ​​of the denoised image at the location; Indicates the Gaussian kernel at the offset Weighting coefficients at each location; This represents the radius of the Gaussian kernel in the horizontal and vertical directions, and the kernel size is... ; , Representing relative to pixels Horizontal and vertical offsets, .

[0066] Gaussian kernel It can be defined as:

[0067] in, This represents the standard deviation of the Gaussian filter and is used to control the smoothing intensity. This represents an exponential function.

[0068] Gaussian filtering can effectively suppress high-frequency noise while preserving the overall shape of micro-topographic edges such as ridgelines and valley lines.

[0069] 4. Brightness and contrast normalization To reduce the differences in brightness and contrast caused by different imaging times and lighting conditions, the system processes the denoised image. Normalization is performed to obtain a normalized image. One implementation method is linear normalization based on the global mean and standard deviation:

[0070] in, Indicated in pixels Normalized image pixel values ​​at the location; Representing an image The global grayscale mean is defined as:

[0071] Representing an image The global grayscale standard deviation is defined as:

[0072] This represents the square root operation.

[0073] The above normalization process can unify the grayscale distribution of different images to a relatively stable scale range, which is beneficial for parameter sharing and stable training of subsequent semantic segmentation networks.

[0074] 5. Specific implementation of the image preprocessing module In practice, the image preprocessing module may include the following steps: 1) Image acquisition and cropping: Obtain raw images covering the power transmission line corridor from drones, satellites, or ground-based camera equipment. Based on the center line of the line and the preset buffer width, the original image is cropped into several sub-images of fixed size, and each sub-image is a region to be processed.

[0075] 2) Grayscale conversion: For each colored sub-image, according to Perform grayscale conversion to obtain a grayscale image. .

[0076] 3) Gaussian denoising and smoothing: Select kernel size as Standard deviation is Gaussian kernel For grayscale images Perform convolution operation to obtain the denoised image. It effectively removes isolated noise points and slight texture noise.

[0077] 4) Brightness and contrast normalization: calculate global mean and standard deviation The image is then normalized according to the following formula to obtain the preprocessing result. : .

[0078] 5) Output preprocessing results: Normalized image As input to the subsequent terrain semantic segmentation module (S2), it realizes the transformation from the original image to the standardized preprocessed image.

[0079] Through the aforementioned image preprocessing module, the system can significantly reduce noise interference and imaging condition differences while ensuring micro-topographic morphology and texture information, providing high-quality input data for subsequent semantic segmentation of ridge candidate regions, valley candidate regions, and water body regions.

[0080] S2) Terrain semantic segmentation module The terrain semantic segmentation module is used for the original image or the pre-processed image. Pixel-level classification is performed, dividing each pixel into ridge candidate regions, valley candidate regions, water regions, or background regions, providing basic region segmentation results for the subsequent skeleton extraction module.

[0081] The terrain semantic segmentation module of this invention is built based on the Segment Anything Model (SAM). As a pre-trained large segmentation model, SAM has powerful image feature representation capabilities and can adapt to micro-terrain scenes with a small number of labeled samples.

[0082] (1). Definition of SAM-based semantic segmentation network Define the pre-trained SAM network as ,in, This represents a segmentation network pre-trained based on the Segment AnythingModel; its internal parameter vector is defined as follows: , This represents the initial parameter vector of the SAM pre-trained model.

[0083] In this invention, with Based on this, some parameters were fine-tuned to obtain a network suitable for micro-terrain semantic segmentation tasks: ,in, This represents the fine-tuned terrain semantic segmentation network; This represents the network parameter vector after fine-tuning; and The relationship can be represented as , This represents the parameter increments generated when fine-tuning SAM on micro-topography training data.

[0084] (2). Pixel category response based on SAM Preprocessed image Input the fine-tuned network This allows us to obtain the response value of each pixel across various terrain categories. The total number of categories is defined as follows: ,in, The number of categories representing terrain semantic segmentation, in embodiments of the present invention, These correspond to the candidate areas for mountain ridges, valley bottoms, water bodies, and background areas, respectively.

[0085] Then at pixel coordinates Place, No. The response value for each category is defined as follows:

[0086] in, Indicated in pixels At this point, the SAM network is fine-tuned. The output of the first Response values ​​for each category; Indicates category index, .

[0087] (3). Softmax probability mapping To convert the response value The invention maps pixels to the probability of belonging to each category at each pixel location. Performing the Softmax operation on all categories yields:

[0088] in, Represents pixels Belongs to the The probability of each category; Represents an exponential function; This represents the index used to iterate through all classes during the Softmax normalization process. , To represent in pixels At this point, the SAM network is fine-tuned. The output of the first The response values ​​for each category. The above probabilities satisfy the condition that the sum of the probabilities of each category at each pixel location equals 1, as shown below: .

[0089] (4). Pixel category decision and label graph Using the maximum probability criterion, this invention will use pixels The final category label is defined as:

[0090] in, Represents pixels Semantic category label index; This represents the value of the independent variable that makes the given function reach its maximum value.

[0091] In a typical embodiment of the present invention, the mapping relationship between the tag index and the specific terrain category is as follows: when At that time, pixels It belongs to the ridge candidate area; when At that time, pixels It belongs to the candidate area of ​​elevation valley bottom; when At that time, pixels It belongs to a water body area; when At that time, pixels It belongs to the background area.

[0092] (5). Category binary masking and morphological post-processing To facilitate subsequent skeleton extraction, this invention uses the label image. Convert to various types of binary masks Defined as:

[0093] in, Indicated in pixels First The binary mask value of the class, which takes the value 0 or 1.

[0094] To improve the spatial connectivity of candidate ridge regions, candidate valley regions, and water bodies, this invention performs a morphological opening operation on the binary mask. The structural element is defined as follows: ,in, Structural elements for morphological operations are defined in a finite region. superior; Represents structural element The domain of definition. The erosion operator is defined as: And the dilation operator is: ,in, Representing structural elements Operators that perform erosion operations on binary images; Representing structural elements Operators that perform dilation operations on binary images.

[0095] For each category mask The smoothed mask is obtained by morphological opening operation:

[0096] in, Indicates the morphological post-processing of the first Binary-like mask. After the above processing, the candidate ridge region mask can be obtained. Valley bottom candidate region mask and water area mask This provides a connected and clearly defined region for subsequent skeleton extraction.

[0097] 5. Specific Implementation Method of Terrain Semantic Segmentation Module In one specific implementation, the terrain semantic segmentation module includes the following steps: 1) Preprocessed Image Input: The normalized image output from the image preprocessing module. As input, it is fed into the terrain semantic segmentation module.

[0098] 2) SAM-based network construction and fine-tuning: The publicly available Segment Anything Model is selected as the pre-trained model to obtain the pre-trained segmentation network. and its parameter vector A micro-topography training dataset containing labeled candidate regions for ridges, valleys, water bodies, and background areas is constructed. Training images and their labels are then input into the SAM model. Using the topographic label mask from the training dataset as a supervision signal, only the parameters of the high-level feature layers of the encoder and the decoder are fine-tuned to obtain the parameter vector. Thus, a micro-terrain semantic segmentation network is constructed. .

[0099] 3) Pixel-level category response and probability output: This involves processing the image... Input network This yields the response value of each pixel for each category. ; Use the Softmax formula below to calculate the probability of a pixel belonging to each category. : .

[0100] 4) Label map generation and category mapping: Assign a category label to each pixel according to the following formula to form a label map. ;

[0101] Index the tags according to the preset mapping relationship. These correspond to the candidate areas for mountain ridges, valleys, water bodies, and background areas, respectively.

[0102] Binary mask extraction and morphological post-processing: Generate binary masks of various categories according to the following formula. ;

[0103] Selecting structural elements By using the following formula Perform an opening operation to remove small noise points and smooth the boundaries to obtain a smooth mask. : .

[0104] 5) Output the segmentation results used for skeleton extraction. , , and label images As input to the subsequent S3 skeleton extraction module, it enables automatic segmentation of three key regions from the normalized image: "ridge candidates / valley candidates / water body regions".

[0105] S3) Skeleton Extraction Module The skeleton extraction module is used to extract ridgeline skeletons, valley skeletons, and water body boundaries based on the terrain semantic segmentation results. It transforms regional ridge candidate regions, valley candidate regions, and water body regions into one-pixel-wide linear skeletons, providing structured linear terrain elements for subsequent skeleton neighborhood wavelet feature extraction and five types of micro-terrain classification.

[0106] S3-1 Skeleton Extraction Principle The skeleton extraction module of this invention takes the smooth mask output by the terrain semantic segmentation module as input and includes: a ridge candidate region mask. Valley bottom candidate region mask Water area mask .

[0107] in, Defined in S2 as the first The binary mask is obtained by morphological post-processing.

[0108] 1. Definition of skeletonization operator Define the skeletonization operator as: ,in, This represents an operator that performs thinning / skeletonization operations on a binary image, used to gradually shrink a region to a skeleton line one pixel wide while preserving connectivity and the overall topology.

[0109] Given any binary image: ,in, Indicated in pixels The binary image pixel value at a given location takes the value 0 or 1, and its skeleton image is defined as:

[0110] in, Indicates by The skeleton binary image obtained by the skeletonization operator takes values ​​of 0 or 1; when When, it indicates that the pixel belongs to the skeleton line; when When the value is zero, it indicates that the pixel does not belong to the skeleton line.

[0111] In this invention, the skeletonization operator Φ(•) is implemented using a morphological algorithm based on iterative thinning (Zhang-Suen thinning algorithm).

[0112] 2. Extraction of ridgeline and valleyline skeletons Based on the aforementioned skeletonization operator, this invention masks the ridge candidate region. and valley bottom candidate region mask Convert to the corresponding skeleton image: Ridgeline skeleton image:

[0113] in, Indicated in pixels The binary image pixel values ​​of the ridgeline skeleton are either 0 or 1. Valley skeleton image:

[0114] in, Indicated in pixels The binary image pixel values ​​of the valley skeleton are either 0 or 1. In the skeleton image above, pixels with a value of 1 form a ridgeline or valley line that is one pixel wide.

[0115] 3. Extraction of water body boundary lines For water area masking This invention obtains the water body boundary skeleton through morphological boundary extraction. Water body boundary skeleton Defined as:

[0116] in, Indicated in pixels The pixel values ​​of the binary image of the water body boundary skeleton, which take values ​​of 0 or 1; The erosion operator is already defined in S2; For the structuring elements already defined in S2; "" indicates a pixel-by-pixel subtraction operation. Through the above boundary extraction, a one-pixel wide outline of the outer boundary of the water area is preserved, which is used to depict the boundary between the water body and the surrounding terrain.

[0117] 4. Connected component decomposition of skeleton line segments To perform subsequent geometric and topological feature calculations on the skeleton, this invention performs connected component decomposition on the skeleton image, dividing the skeleton into several independent line segments.

[0118] 1) Connected component decomposition of the ridgeline framework Define the ridgeline skeleton image The total number of connected segments is ,in, This represents the number of connected segments obtained by decomposing the ridgeline skeleton. For each ridgeline segment, its corresponding binary mask is defined as: ,in, Indicates the first The ridge line segment in pixels The mask value at that location can be either 0 or 1. For ridgeline segment indexing, .

[0119] The ridgeline skeleton image can be represented as a pixel-by-pixel summation of all line segment masks:

[0120] 2) Connected component decomposition of the valley line skeleton Similarly, define the valley skeleton image. The total number of connected segments is: ,in, This represents the number of connected segments obtained by decomposing the valley skeleton. Define the first... The binary mask for the valley line segment is: ,in, Indicates the first Valley line segments in pixels The mask value at that location can be either 0 or 1. For valley line segment index, Then we have:

[0121] 3) Connectivity decomposition of the water body boundary skeleton Define the skeleton image of the water body boundary. The total number of connected segments is: ,in, This represents the number of connected line segments obtained by decomposing the water body boundary skeleton. Define the... The binary mask for the water body boundary line segment is: ,in, Indicates the first The water body boundary line segment in pixels The mask value at that location can be either 0 or 1. For water body boundary line segment index, Then we have: .

[0122] 5. Geometric property representation of skeleton line segments To calculate the skeleton's geometric features in subsequent steps, this invention represents each skeleton line segment as a sequence of ordered pixel coordinates. Taking a ridge line segment as an example, the first... The total number of pixels on the ridge line segment is: ,in, Indicates the first The number of pixels contained in a ridgeline segment. Define the first... The first ridge segment The coordinates of the skeleton pixels are: ,in, Indicates the first The first ridge segment The horizontal coordinates of each pixel; Indicates the first The first ridge segment The vertical coordinates of each pixel; Indicates the sequence number along the skeleton line segment. Then the first The framework of a ridge segment can be represented as an ordered set:

[0123] in, Indicates the first The set of skeletal points for a ridgeline segment. The geometric length of the ridgeline segment can be approximated as:

[0124] in, Indicates the first The length of the ridgeline segment; The square root operation is indicated by the symbol "". Valley lines and water body boundaries can be represented and their geometric properties calculated in the same way. The above skeleton set and length formula are called in the subsequent construction of micro-topographic geometric and topological features.

[0125] Detailed implementation of the S3-2 skeleton extraction module: In one specific implementation, the skeleton extraction module includes the following steps: 1. Input semantic segmentation mask: Read the smooth mask output by the terrain semantic segmentation module. , , These correspond to the candidate areas for mountain ridges, valley bottoms, and water bodies, respectively.

[0126] 2. Ridgeline skeleton extraction: Calling the skeletonization operator Masking candidate areas for ridges Perform iterative refinement to obtain the ridgeline skeleton image. ;right Perform connected component labeling to divide the connected skeleton pixels into The ridgeline segments are defined, and a mask is generated for each segment. ,satisfy:

[0127] 3. Extraction of the valley line skeleton: Calling the skeletonization operator Masking candidate regions at the bottom of the valley Perform iterative refinement to obtain the valley skeleton image. ;right Perform connected component labeling to divide the connected skeleton pixels into Create valley line segments and generate a mask for each segment. ,satisfy:

[0128] 4. Extraction of water body boundary skeleton: Masking of water body areas Calling the erosion operator Obtain the eroded image The water body boundary skeleton image is calculated using the following formula. .

[0129]

[0130] right Perform connected component labeling, dividing the connected boundary skeleton pixels into Create water body boundary line segments and generate a mask for each segment. ,satisfy:

[0131] 5. Calculation of geometric properties of skeleton line segments: Masking each ridge segment Extract skeleton pixel coordinates according to the connected path order Construct a skeleton point set And use the following formula to calculate the geometric properties such as the length of the ridge segment:

[0132] Valley line segment mask Water boundary line segment mask Perform the same skeleton point extraction and geometric attribute calculation to obtain the corresponding valley line set and water body boundary line set and their lengths and other geometric information.

[0133] 6. Output skeleton results Ridgeline skeleton image Valley line skeleton image Water body boundary skeleton image and its corresponding set of line segments , , and geometric attribute parameters (such as) (etc.) as input to the subsequent S4 skeleton neighborhood wavelet feature extraction module, providing a basis for extracting multi-scale texture features along the skeleton line and constructing five types of micro-topographic geometric features.

[0134] S4) Skeleton Neighborhood Wavelet Feature Extraction Module The skeleton neighborhood wavelet feature extraction module is used to extract multi-scale wavelet texture features along the skeleton line neighborhood based on the ridge line skeleton, valley line skeleton and water body boundary line skeleton. It encodes the local texture information near the skeleton into feature vectors, which, together with the skeleton geometric and topological features, provide input for the subsequent five types of micro-terrain classification.

[0135] The input to this module includes: a set of ridgeline skeleton segments. and its mask Valley line skeleton line segment set and its mask Set of water body boundary line skeleton segments and its mask The above symbols have all been defined in S3. The preprocessed normalized image is still denoted as... .

[0136] S4-1 Skeleton Neighborhood Wavelet Feature Extraction Principle 1. Definition of skeleton neighborhood strip regions To depict the local texture near the skeleton, the present invention constructs strip-shaped neighborhood regions on both sides of each skeleton line segment.

[0137] 1) Strip width and length parameters.

[0138] Introducing parameters: : Indicates the width (in pixels) of the skeleton neighborhood strip in the normal direction; : Represents the length (in pixels) of each local strip sub-block along the skeleton line direction in the tangential direction. Wherein, and It is a positive integer and is set according to the image spatial resolution and micro-topography scale.

[0139] 2) Local strip coordinate system.

[0140] To describe the pixel coordinates within a strip, this invention introduces local coordinates. ,in: : Represents the pixel index along the tangential direction in the local coordinate system of the strip. ; : Represents the pixel index along the normal direction in the local coordinate system of the strip. .

[0141] 3) Local strip image.

[0142] Taking the ridgeline as an example, for the first The ridge segment and its upper section skeletal pixels (See definition in S3) In the original normalized image In the middle, extract a point centered on this point along the tangential and normal directions of the skeleton. A local image patch, denoted as:

[0143] in, Indicates the first The first ridge segment Local stripe images of the neighborhood of a skeleton point in local coordinates Pixel value at; For ridgeline segment indexing, ; These are the indexes of the skeletal points along the ridgeline segment. ; , As stated above.

[0144] Similarly, local strip images of valley lines and water body boundaries can be defined. Local image of the valley line neighborhood:

[0145] in, , , Indicates the first The number of skeleton points contained in the valley line segment; local image of the water body boundary neighborhood: ,in, , , Indicates the first The number of skeleton points contained in the water body boundary line segment. The above local strip image is achieved by truncating using a sliding window along the tangential direction of the skeleton.

[0146] 2. Wavelet decomposition and subband coefficients To characterize the multi-scale, multi-directional texture features of striped images, this invention performs multi-scale, multi-directional textural analysis on each local striped image. Perform a two-dimensional discrete wavelet transform.

[0147] 1) Decomposition layer number and orientation index The following new indexes are introduced: : Represents the total number of wavelet decomposition layers, which is a positive integer; : Represents the layer index of wavelet decomposition. ; : Indicates the number of directions considered in each decomposition layer. In a typical embodiment, These correspond to the horizontal, vertical, and diagonal directions, respectively. : Indicates direction index, .

[0148] 2) Wavelet subband coefficients of local stripes Local images of the ridgeline neighborhood Performing a two-dimensional discrete wavelet transform yields the wavelet subband coefficients for each layer and direction, defined as follows: ,in, Indicates the first The first ridge segment A local stripe image in the decomposition layer ,direction Wavelet subband coefficients on local coordinates The value at that location; , ; , Similarly, for a local image of the valley line and the water body boundary, we can obtain: wavelet sub-band coefficients of the valley line neighborhood: Wavelet subband coefficients of the water body boundary: The coefficients mentioned above are calculated using a selected wavelet basis (e.g., Daubechies wavelet) and a two-dimensional discrete wavelet transform algorithm.

[0149] 3. Wavelet texture statistical features This invention calculates several statistics on the wavelet subbands of each local strip to describe the multi-scale texture characteristics of the skeleton neighborhood.

[0150] 1) Subband energy characteristics Define the ridgeline neighborhood in the decomposition layer ,direction The energy characteristics of the surface are:

[0151] in, Indicates the first The first ridge segment The local stripe image in the first Layer, First Energy characteristics of wavelet subbands in each direction; " indicates the squaring operation. Similarly, the energy characteristics of the neighborhood of the valley line and the water body boundary line are defined as follows:

[0152] 2) Subband entropy characteristics To characterize the complexity of the coefficient distribution, this invention may introduce an entropy feature in some embodiments. Taking the ridge neighborhood as an example, the first... Layer, First The coefficient probability of the directional sub-band is:

[0153] in, Indicates the first The first ridge segment The first local stripe image Layer, First In the directional sub-band, The probability corresponding to the normalized coefficient magnitude of the position; , These are the pixel indices used for summation and traversal in the local coordinate system. , ; " represents absolute value operation. Based on this, the characteristics of entropy are defined as follows:

[0154] in, Indicates the first The first ridge segment The first local stripe image Layer, First Entropy characteristics of directional wavelet subbands; It represents logarithmic operations.

[0155] Similarly, wavelet entropy features of the neighborhood of valley lines and water body boundaries can be defined. , .

[0156] 3) Feature aggregation along the skeleton line To obtain an overall wavelet texture description for each skeleton line segment, this invention aggregates the local features of all skeleton points on the same line segment. Taking a ridge line segment as an example, the first... The ridge segment in the decomposition layer ,direction The average energy characteristic on is defined as:

[0157] in, Indicates the first The ridge segment in the first Layer, First Average energy characteristics in the direction; For the first The number of skeleton points contained in each ridge segment (see S3). Similarly, the average entropy feature can be defined:

[0158] Valley lines and water boundary segments can be aggregated using the same method to obtain the first... The valley line and the water body boundary segment in the first Layer, First Average energy characteristics and average entropy characteristics in different directions , and , .

[0159] 4) Skeleton neighborhood wavelet feature vector Finally, the average energy features and entropy features of each layer and direction are concatenated in a fixed order to form the multi-scale wavelet texture feature vector for each skeleton line segment. For example, the first... The wavelet eigenvector of a ridge segment can be represented as:

[0160] in, Indicates the first The skeleton neighborhood wavelet texture feature vector of the ridge line segment; " indicates the vector transpose operation.

[0161] Eigenvectors of valley lines and water body boundary segments , They can be constructed using the same rules. These feature vectors, along with the skeleton geometry and topological features, will be used as input to the five-class micro-terrain classification model in the subsequent S5 module.

[0162] Detailed Implementation of the S4-2 Skeleton Neighborhood Wavelet Feature Extraction Module In one specific implementation, the skeleton neighborhood wavelet feature extraction module includes the following steps: 1. Input the skeleton line segment set and the normalized image, then read the ridgeline skeleton line segment set output by the skeleton extraction module. Valley line skeleton line segment set Set of water body boundary line skeleton segments and the corresponding skeleton point coordinates , , And load the normalized image output by the preprocessing module. .

[0163] 2. Set the strip size and wavelet parameters to set the width of the skeleton neighborhood strip. With strip sub-block length ,For example Select 10-30 pixels. Select 16-64 pixels; choose the wavelet decomposition level. and direction number ,For example or , Choose a wavelet basis for two-dimensional discrete wavelet transform, such as the Daubechies wavelet.

[0164] 3. Constructing a local strip image of the skeleton neighborhood for each ridge segment Along its skeleton point sequence Process sequentially: Calculate the tangential and normal directions of the skeleton at that point; with that point as the center, in... Construct a dimension along the tangential and normal directions. Local image patches, to obtain If the skeleton is long, several representative skeleton points can be evenly selected along the skeleton line according to the step size to construct a strip image, thereby controlling the computational load. For valley lines... Boundary line with water body The processing steps are the same, and the results are obtained respectively. , .

[0165] 4. Perform a two-dimensional discrete wavelet transform on each local strip image. Perform two-dimensional discrete wavelet decomposition to obtain wavelet subband coefficients for each layer and direction. ;right , Perform the same wavelet decomposition to obtain , .

[0166] 5. Calculate wavelet energy and entropy features For each local strip image and its sub-band coefficients, the following formula is used for calculation:

[0167] Calculate the energy characteristics; if necessary, calculate the entropy characteristics using the following formula:

[0168] The wavelet characteristics of the valley line and the water body boundary are similar.

[0169] 6. Aggregate wavelet features along the skeleton line. For each ridge segment, average the wavelet features of all local stripes on the segment according to the following formula to obtain multi-scale wavelet features at the segment level. Valley lines and water boundary segments are aggregated in the same way.

[0170]

[0171]

[0172] 7. Construct and output the skeleton neighborhood wavelet feature vectors. The same line segment in all decomposition levels and direction The average energy and entropy features are concatenated in a fixed order to form a line segment-level wavelet feature vector. , , The wavelet feature vectors mentioned above are used as output and fed into the subsequent S5 module, where they are concatenated with the skeleton geometry and topological features for five types of micro-terrain classification.

[0173] S5) Five-category micro-terrain classification module based on "geometric + wavelet texture features" This module is used to classify each ridge line segment, valley line segment, and water body boundary line segment into five categories based on the skeleton geometry and topological features and the skeleton neighborhood wavelet texture features. It outputs micro-topographic category labels such as mountain pass, high mountain watershed, topographic uplift, canyon, and water body as the final classification result of the method of this invention.

[0174] The input to this module includes: a set of ridgeline skeleton segments and their wavelet features. , Valley line skeleton segment set and its wavelet characteristics: , The set of water body boundary line skeleton segments and their wavelet features: , The aforementioned skeleton segment set and wavelet eigenvectors have been defined in S3 and S4.

[0175] S5-1 Five-Type Micro-Topography Classification Principles 1. Skeleton geometry and topological eigenvectors Before classifying the skeleton line segments, this invention first constructs geometric and topological feature vectors for each skeleton line segment to characterize the length, shape, and spatial relationships of the line segment.

[0176] 1) Geometric and topological characteristic scalars Taking a ridge segment as an example, the following geometric and topological scalar features are introduced: : indicates the first The geometric length of a ridge segment can be determined using the definition in S3. ; : indicates the first The average strike angle of a ridge line segment is used to describe the overall direction of the segment; : indicates the first The curvature of a ridge line segment can be expressed, for example, as the ratio of the actual length of the segment to the Euclidean distance between its endpoints; : indicates the first The average relative height position of the ridge line segment can be obtained by statistically analyzing the normalized positions of the skeleton points in the vertical direction of the image; : indicates the first The minimum distance between a ridge line segment and the nearest valley line segment; : Indicates the degree of the endpoint of the ridge line segment on the topographic skeleton map, used to reflect the intersection relationship.

[0177] This invention can also extend other geometric or topological features as needed. For a unified representation, a feature dimension parameter is introduced: : Represents the total dimension of geometric and topological features, and is a positive integer; for the , The first ridge segment Each geometric and topological feature scalar is denoted as ,in .

[0178] 2) Geometric and topological eigenvectors By concatenating the above scalars in a fixed order, we can construct the geometric and topological feature vectors of the ridgeline segment:

[0179] in, Indicates the first Geometric and topological characteristic vectors of a ridgeline segment; The '' operator represents the vector transpose operation. Similarly, the geometric and topological characteristic vectors of valley lines and water body boundary segments are defined as follows:

[0180] in, , The first segment represents the boundary line between the valley line and the water body. A geometric and topological scalar.

[0181] 2. Fusion of geometric features and wavelet texture features To fully utilize the skeleton morphological information and neighborhood texture information, this invention concatenates geometric and topological feature vectors with skeleton neighborhood wavelet feature vectors to form a comprehensive feature vector.

[0182] 1) Comprehensive feature vector of ridge line segments For the The comprehensive feature vector of a ridge segment is defined as follows:

[0183] in, Indicates the first A ridgeline segment is used as a comprehensive feature vector for micro-topography classification; " indicates a vector concatenation operation by dimension; The first one defined in S4 The skeleton neighborhood wavelet texture feature vector of a ridge line segment.

[0184] 2) Combined feature vector of valley segment and water body boundary segment Similarly, the combined eigenvector of the valley line and the water body boundary segment is defined as:

[0185] in, Indicates the first The comprehensive feature vector of the valley line segment; Indicates the first The comprehensive feature vector of the water body boundary line segment; The first Geometric and topological characteristic vectors of valley lines and water body boundary segments. The respective The skeleton neighborhood wavelet texture feature vectors of valley lines and water body boundaries.

[0186] 3. Five-category micro-topography classification model This invention treats each skeleton line segment as a sample to be classified and predicts its micro-terrain category through a supervised learning model.

[0187] 1) Definition of five types of micro-topography Total number of micro-terrain categories introduced: ,in, Indicates the number of micro-topography categories, in this invention These correspond to: mountain pass, high mountain watershed, topographic uplift, canyon, and water body, respectively.

[0188] Define the micro-topography category index as h, where h represents the micro-topography category index. For example, in a typical mapping, it can be specified that: : mountain pass; High mountain watershed; : Terrain uplift; :canyon; : Water body.

[0189] 2) Classification model definition Introducing the classification model function: ,in, This represents a supervised learning model used for five types of micro-topography classification; This represents the parameter vector of the classification model. For a given line segment, the comprehensive feature vector... (represent , or One of the classification models outputs a response value for each micro-terrain category:

[0190] in, Represents the comprehensive feature vector In the Response value for each micro-topography category; subscript This indicates selecting the first element from the model output vector. Each component.

[0191] 3) Category Probability and Decision Rules To convert the response value Converting to probabilities, this invention performs a Softmax operation on the response values ​​for all categories:

[0192] in, This indicates a composite eigenvector. The skeleton line segment belongs to the first The probability of each micro-topography category; This represents the index used to traverse all micro-terrain categories during the Softmax normalization process. Ultimately, this invention uses the maximum probability criterion to determine the micro-topography category label for the skeleton line segment:

[0193] in, The final five micro-terrain category labels represent the skeleton line segments, with values ​​ranging from 1 to 2. to one, This indicates the value of the independent variable that maximizes the objective function. (This tag) This is the micro-topography classification result of the skeleton line segment by the method of the present invention.

[0194] Detailed Implementation of the S5-2 Five-Type Micro-Topography Classification Module In one specific implementation, the five-category micro-terrain classification module based on "geometric + wavelet texture features" includes the following steps: 1. Input the comprehensive feature vector The geometric and topological feature vectors of each ridge segment are read from the skeleton extraction module and the skeleton neighborhood wavelet feature extraction module. and wavelet eigenvectors Construct the comprehensive feature vector as follows:

[0195] Construct a comprehensive feature vector for each valley line segment. Construct a comprehensive feature vector for each water body boundary line segment. .

[0196] 2. Classification model training (offline phase) During the offline training phase, a training sample set containing a large number of labeled micro-topography categories is constructed. Each training sample corresponds to a ridge line segment, valley line segment, or water boundary segment, and has a comprehensive feature vector. and manually labeled micro-topography categories ,in Indicates the training sample index; by For input, with For supervised labeling, cross-entropy loss or other supervised learning loss functions are used, and the parameter vector is iteratively updated through optimization methods such as gradient descent. A well-trained five-class micro-terrain classification model was obtained. In practical implementation, the classification model It can be any of the following: multilayer perceptron (MLP), support vector machine, random forest, or graph neural network. The implementation form does not affect the method flow of this invention.

[0197] 3. Online classification reasoning In the actual identification phase, for each ridge segment to be classified, its comprehensive feature vector is read. Input classification model Response values ​​for each micro-terrain category were obtained. Calculate the probability that the line segment belongs to each micro-topography category using the following formula. and the Micro-topographic category labels for ridgeline segments :

[0198]

[0199] Similarly, the same reasoning process is performed on valley line segments and water body boundary segments to obtain the corresponding category labels:

[0200]

[0201] in, , , These represent the final five categories of micro-topography classification results for ridgeline segments, valley segments, and water boundary segments, respectively. and These represent the probabilities of valley line segments and water body boundary segments belonging to various micro-topographic categories.

[0202] 4. Output classification results This invention provides a set of classification labels for all skeleton line segments: : Results of micro-topography categories corresponding to ridgeline segments; : Micro-topographic category results corresponding to valley line segments; : Micro-topographic category results corresponding to water body boundary segments; The final micro-topography classification output of this method is directly provided to upper-level application systems, such as power transmission line disaster prevention assessment systems and icing risk assessment systems, without the need for separate additional output and mapping steps.

[0203] Example 3 Based on the same inventive concept, this invention also provides a micro-terrain region identification system based on segmentation large model and skeleton wavelet features, including: The image preprocessing module is used to obtain the original image of the region to be identified; The terrain semantic segmentation module is used to input the original image into a pre-constructed terrain semantic segmentation network, perform semantic classification on each pixel in the image, generate a terrain semantic segmentation map, and perform morphological opening operations on the terrain semantic segmentation map to obtain the classification masks. The skeleton extraction module is used to perform skeleton extraction operations based on each classification mask to obtain skeleton images, and to perform connected component decomposition on the skeleton images to obtain terrain skeleton line segments, and extract geometric and topological feature vectors. The skeleton domain wavelet feature extraction module is used to form a skeleton neighborhood wavelet feature vector that characterizes the neighborhood multi-scale texture characteristics of the terrain skeleton line segment based on the terrain skeleton line segment combined with two-dimensional discrete wavelet transform. The micro-topography classification module is used to concatenate geometric and topological feature vectors and skeleton neighborhood wavelet feature vectors to form a comprehensive feature vector for each skeleton line segment, and to identify the micro-topography region category based on the comprehensive feature vector. The pre-built terrain semantic segmentation network is obtained by fine-tuning the large segmentation model SAM.

[0204] Preferably, the terrain semantic segmentation module includes: The semantic segmentation submodule is used to: input the original image into a pre-constructed terrain semantic segmentation network to obtain the response value of each pixel in each terrain category; perform a softmax operation on all categories at each pixel location to obtain the probability that the pixel belongs to a certain category; determine the terrain category to which the pixel belongs based on the probability that the pixel belongs to a certain category using the maximum probability criterion; and generate a terrain semantic segmentation map based on the terrain category to which the pixel belongs. The smoothing submodule is used to convert the terrain semantic segmentation map into binary masks of various categories; and to perform morphological opening operations on the binary masks to obtain smoothed masks for each category.

[0205] Preferably, the skeleton domain wavelet feature extraction module is specifically used for: For each ridge segment, valley segment, and water boundary segment, a strip-shaped neighborhood region is constructed along the tangential and normal directions of the skeleton on the normalized image. A two-dimensional discrete wavelet transform is performed on the local image patch of the skeleton neighborhood to obtain wavelet sub-band coefficients at multiple scales and in multiple directions. Statistical features are calculated on the wavelet sub-bands at each scale and in each direction, including energy and entropy. The statistical features of each local strip along the same skeleton segment are aggregated to form a skeleton neighborhood wavelet feature vector that characterizes the multi-scale texture properties of the neighborhood of the skeleton segment.

[0206] Preferably, the skeleton extraction module is specifically used for: Based on the ridge candidate region mask and the valley candidate region mask, a skeletonization operator is obtained using an iterative refinement-based morphological algorithm; based on the skeletonization operator and the binary image pixel value at the pixel, the ridge line skeleton image and the valley line skeleton image are obtained; based on the water body region mask, the water body boundary line skeleton image is obtained through morphological boundary extraction. The skeleton image is decomposed into connected components, dividing the skeleton into several independent line segments to obtain the terrain skeleton line segments; each skeleton line segment is represented as a sequence of ordered pixel coordinates, and the total number of pixels, pixel coordinates, skeleton point set and geometric length of the line segment are recorded.

[0207] Preferably, the specific steps in the semantic segmentation submodule for determining the terrain category to which a pixel belongs based on the probability of the pixel belonging to a certain category, using the maximum probability criterion, include: The semantic category label index of the pixel is calculated based on the maximum probability of the pixel belonging to each category and the semantic category label index calculation formula; the terrain category to which the pixel belongs is obtained based on the mapping relationship between the semantic category label index and each terrain category.

[0208] Preferably, the specific implementation steps of performing morphological opening operations on the terrain semantic segmentation map in the semantic segmentation submodule to obtain each classification mask include: The terrain semantic segmentation map is converted into binary masks for each category; morphological opening operations are performed on the binary masks to obtain smoothed category masks.

[0209] Example 4 like Figure 4 As shown, the present invention also provides an electronic device, which may be a computer device, a microcontroller device, a smart mobile device, etc. The electronic device in this embodiment may include a processor, a memory, a transceiver component, etc. The memory, processor, and transceiver component are connected via a bus; the memory can be used to store executable programs, and an exemplary executable program may include instructions; the processor is used to execute the instructions stored in the memory. The memory can also be used to store data, which can be accessed and / or modified when instructions are executed.

[0210] The processor may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, and it is suitable for implementing one or more instructions. Specifically, it is suitable for loading and executing one or more instructions in the storage medium to realize the corresponding method flow or corresponding function, so as to realize the steps of the micro-topographic region recognition method based on segmentation large model and skeleton wavelet features in the above embodiment.

[0211] Example 5 Based on the same inventive concept, this invention also provides a readable storage medium, specifically an electronic device readable storage medium (Memory). This readable storage medium is a memory device within an electronic device used to store programs and data. It is understood that the storage medium here can include both built-in storage media within the electronic device and extended storage media supported by the electronic device. The storage medium provides storage space, which stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more executable programs (including program code). It should be noted that the storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. Loading and executing one or more instructions stored in the storage medium by the processor can implement the steps of the micro-topographic region recognition method based on segmentation large model and skeleton wavelet features in the above embodiments.

[0212] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0213] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0214] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The function specified in one or more boxes.

[0215] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0216] The above are merely embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of the claims of the present invention pending approval.

Claims

1. A method for identifying micro-topographic regions based on segmentation large models and skeleton wavelet features, characterized in that, include: Obtain the original image of the region to be identified; The original image is input into a pre-constructed terrain semantic segmentation network. Semantic classification is performed on each pixel in the image to generate a terrain semantic segmentation map. Morphological opening operations are then performed on the terrain semantic segmentation map to obtain the classification masks. Based on each classification mask, a skeleton extraction operation is performed to obtain a skeleton image. The skeleton image is then decomposed into connected components to obtain terrain skeleton line segments, and geometric and topological feature vectors are extracted. Based on the terrain skeleton line segments combined with two-dimensional discrete wavelet transform, a skeleton neighborhood wavelet feature vector is formed to characterize the neighborhood multi-scale texture characteristics of the terrain skeleton line segments. Geometric and topological feature vectors and skeleton neighborhood wavelet feature vectors are concatenated to form a comprehensive feature vector for each skeleton line segment, and the micro-topography region category is identified based on the comprehensive feature vector. The pre-built terrain semantic segmentation network is obtained by fine-tuning the large segmentation model SAM.

2. The method as described in claim 1, characterized in that, After obtaining the original image of the region to be identified, the following steps are also included: Perform grayscale conversion on the original color image to convert the three-channel image into a single-channel grayscale image. The grayscale image is smoothed and denoised using the discrete Gaussian filtering method to obtain the denoised image. The denoised image is then normalized to obtain the processed image.

3. The method as described in claim 1, characterized in that, The step of inputting the original image into a pre-constructed terrain semantic segmentation network, performing semantic classification on each pixel in the image, and generating a terrain semantic segmentation map includes: The original image is input into a pre-built terrain semantic segmentation network to obtain the response value of each pixel in each terrain category; Perform a Softmax operation on all categories at each pixel location to obtain the probability that the pixel belongs to a certain category; Based on the probability that a pixel belongs to a certain category, the terrain category to which the pixel belongs is determined using the maximum probability criterion; Generate a terrain semantic segmentation map based on the terrain category to which the pixel belongs.

4. The method as described in claim 3, characterized in that, The process of determining the terrain category to which a pixel belongs based on the probability that the pixel belongs to a certain category, using the maximum probability criterion, includes: The semantic category label index of a pixel is calculated based on the maximum probability of the pixel belonging to each category, combined with the semantic category label index calculation formula. The terrain category to which the pixel belongs is obtained based on the mapping relationship between the semantic category label index and each terrain category.

5. The method as described in claim 3, characterized in that, The morphological opening operation is then performed on the terrain semantic segmentation map to obtain various classification masks, including: Convert the terrain semantic segmentation map into binary masks for each category; Perform a morphological opening operation on the binary mask to obtain smoothed classification masks.

6. The method as described in claim 1, characterized in that, The skeleton neighborhood wavelet feature vector, which characterizes the neighborhood multi-scale texture properties of the terrain skeleton line segments by combining two-dimensional discrete wavelet transform, includes: For each ridge line segment, valley line segment, and water boundary line segment, a strip-shaped neighborhood region is constructed on the normalized image along the skeleton tangential and normal directions. Two-dimensional discrete wavelet transform is performed on the local image patch of the skeleton neighborhood to obtain wavelet sub-band coefficients of multiple scales and multiple directions. Statistical features are calculated on wavelet subbands at various scales and in various directions, including energy and entropy. The statistical features of each local stripe are aggregated along the same skeleton line segment to form a skeleton neighborhood wavelet feature vector that characterizes the multi-scale texture properties of the neighborhood of the skeleton line segment.

7. The method as described in claim 1, characterized in that, Based on each classification mask, a skeleton extraction operation is performed to obtain a skeleton image. The skeleton image is then decomposed into connected components to obtain terrain skeleton line segments, and geometric and topological feature vectors are extracted, including: Based on the ridge candidate region mask and the valley candidate region mask, a skeletonization operator is obtained using an iterative refinement-based morphological algorithm. Based on the skeletonization operator combined with the binary image pixel values ​​at the pixel, ridgeline skeleton images and valley skeleton images are obtained. An image of the water body boundary skeleton is obtained by morphological boundary extraction based on a water body region mask; Connected component decomposition is performed on the skeleton image to divide the skeleton into several independent line segments, resulting in terrain skeleton line segments; Each skeleton line segment is represented as a sequence of ordered pixel coordinates, and the total number of pixels, pixel coordinates, skeleton point set, and geometric length of the line segment are recorded.

8. A micro-topographic region identification system based on segmentation large model and skeleton wavelet features, characterized in that, include: The image acquisition module is used to acquire the original image of the region to be identified; The terrain semantic segmentation module is used to input the original image into a pre-constructed terrain semantic segmentation network, perform semantic classification on each pixel in the image, generate a terrain semantic segmentation map, and perform morphological opening operations on the terrain semantic segmentation map to obtain the classification masks. The skeleton extraction module is used to perform skeleton extraction operations based on each classification mask to obtain skeleton images, and to perform connected component decomposition on the skeleton images to obtain terrain skeleton line segments, and extract geometric and topological feature vectors. The skeleton domain wavelet feature extraction module is used to form a skeleton neighborhood wavelet feature vector that characterizes the neighborhood multi-scale texture characteristics of the terrain skeleton line segment based on the terrain skeleton line segment combined with two-dimensional discrete wavelet transform. The micro-topography classification module is used to concatenate geometric and topological feature vectors and skeleton neighborhood wavelet feature vectors to form a comprehensive feature vector for each skeleton line segment, and to identify the micro-topography region category based on the comprehensive feature vector. The pre-built terrain semantic segmentation network is obtained by fine-tuning the large segmentation model SAM.

9. The system as described in claim 8, characterized in that, The terrain semantic segmentation module includes: The semantic segmentation submodule is used to: input the original image into a pre-constructed terrain semantic segmentation network to obtain the response value of each pixel in each terrain category; perform a Softmax operation on all categories at each pixel location to obtain the probability that the pixel belongs to a certain category; determine the terrain category to which the pixel belongs based on the probability that the pixel belongs to a certain category using the maximum probability criterion; and generate a terrain semantic segmentation map based on the terrain category to which the pixel belongs. The smoothing submodule is used to convert the terrain semantic segmentation map into binary masks of various categories; and to perform morphological opening operations on the binary masks to obtain smoothed masks for each category.

10. The system as described in claim 8, characterized in that, The skeleton domain wavelet feature extraction module is specifically used for: For each ridge line segment, valley line segment, and water boundary line segment, a strip-shaped neighborhood region is constructed on the normalized image along the skeleton tangential and normal directions. Two-dimensional discrete wavelet transform is performed on the local image patch of the skeleton neighborhood to obtain wavelet sub-band coefficients of multiple scales and multiple directions. Statistical features are calculated on wavelet subbands at various scales and in various directions, including energy and entropy. The statistical features of each local stripe are aggregated along the same skeleton line segment to form a skeleton neighborhood wavelet feature vector that characterizes the multi-scale texture properties of the neighborhood of the skeleton line segment.

11. The system as described in claim 8, characterized in that, The skeleton extraction module is specifically used for: Based on the ridge candidate region mask and the valley candidate region mask, a skeletonization operator is obtained using an iterative refinement-based morphological algorithm. Based on the skeletonization operator combined with the binary image pixel values ​​at the pixel, ridgeline skeleton images and valley skeleton images are obtained. An image of the water body boundary skeleton is obtained by morphological boundary extraction based on a water body region mask; Connected component decomposition is performed on the skeleton image to divide the skeleton into several independent line segments, resulting in terrain skeleton line segments; Each skeleton line segment is represented as a sequence of ordered pixel coordinates, and the total number of pixels, pixel coordinates, skeleton point set, and geometric length of the line segment are recorded.

12. An electronic device, characterized in that, include: At least one processor and memory; The memory and processor are connected via a bus; The memory is used to store one or more programs; When the one or more programs are executed by the at least one processor, the micro-topographic region identification method based on segmentation large model and skeleton wavelet features as described in any one of claims 1 to 7 is implemented.

13. A readable storage medium, characterized in that, It contains an executable program, which, when executed, implements the micro-topographic region identification method based on segmentation large model and skeleton wavelet features as described in any one of claims 1 to 7.

Images