A map feature recognition method and system based on deep learning

Abstract

The application relates to the technical field of map element recognition, in particular to a map element recognition method and system based on deep learning, which extracts features by acquiring multi-scale multi-channel images through remote sensing, forms an edge track sequence by scanning brightness coherent pixels, merges consistent tracks to generate continuous line segments, analyzes line segment connection relationships to form a closed region contour, determines a main response channel by counting brightness changes of each channel, and realizes map element classification recognition by combining the channel and element mapping rules. According to the application, multi-scale spatial information is cooperatively organized to form a region expression with consistent direction and continuous constraint, so that the map element can maintain a stable overall contour in a complex background, the spatial structure integrity is strengthened by combining the boundary trend and the region closure characteristics, the region attribute is depicted by introducing the channel main response difference, the element differentiation is realized based on the overall response characteristics to achieve clear identification, and the spatial consistency and attribute matching reliability are improved.

Description

Technical Field

[0001] This invention relates to the field of map feature recognition technology, and in particular to a map feature recognition method and system based on deep learning. Background Technology

[0002] The field of map feature recognition technology encompasses the automatic extraction and classification of various spatial entity information in maps. Its core content includes identifying geographic features such as roads, water bodies, buildings, and green spaces from remote sensing images, vector data, or other geospatial data sources, and converting them into structured information. Map feature recognition technology covers multiple aspects, including image processing, feature extraction, graphic analysis, and geographic information system data conversion. Its development aims to improve map update efficiency, enhance the accuracy and completeness of geographic data, and provide fundamental support for navigation, urban planning, and environmental monitoring. Traditional methods for map feature recognition often rely on manual annotation or rule-based classification, resulting in low recognition efficiency and poor adaptability. Therefore, how to utilize data-driven approaches to achieve efficient automatic identification of features in multi-source heterogeneous geographic data has become a key research direction in this field.

[0003] One of the methods for map feature recognition based on deep learning refers to a method that uses a deep neural network model to train and classify map features in remote sensing images. This patent primarily targets features such as building outlines, road boundaries, water areas, and vegetation cover contained in map data. It employs a convolutional neural network model for pixel-level classification. The process involves acquiring high-resolution remote sensing image data, performing uniform size processing and normalization on the images, and then inputting the images into a multi-layer convolutional network for feature extraction and image segmentation. During the training phase, supervised learning is performed using labeled geographic feature image data, with parameters optimized using a cross-entropy loss function and stochastic gradient descent used for iterative model updates. In the recognition phase, the image to be recognized is input into the trained network model, which outputs the corresponding feature classification result image, thus completing the automatic recognition of map features.

[0004] Existing technologies focus on classifying pixels independently using a unified model. In practical applications, when faced with complex terrain features, they tend to overlook the need for continuous representation of the internal structure of features. Under conditions of high-density feature distribution or overlapping boundaries, the classification results lack overall constraints and are prone to regional fragmentation or boundary drift. At the same time, the use of multi-channel information tends to favor numerical discrimination and weakens regional response differences, resulting in recognition biases in scenarios with similar spectra but different semantics, which in turn affects the stability and usability of map feature representation. Summary of the Invention

[0005] To address the technical problems existing in the prior art, embodiments of the present invention provide a map feature recognition method and system based on deep learning. The technical solution is as follows: A deep learning-based map feature recognition method includes the following steps: S1: Acquire multi-channel images of the covered map feature area, construct multi-scale layers according to resolution, extract multi-level feature images, scan each layer of feature images row by row and column by column, extract brightness coherent pixel segments, locate pixel fragments, acquire and encode horizontal and vertical edge trajectories, and form an edge-aware channel sequence. S2: Extract adjacent edge trajectory segments from the edge sensing channel sequence, compare the brightness change direction and trend, select consistent edge trajectory pairs and merge them along the extension direction, connect adjacent edge trajectory segments into continuous line segments and record the start and end positions to obtain a continuous line segment representation map. S3: Obtain the start and end positions of each continuous line segment in the continuous line segment representation diagram, analyze the spatial continuity relationship based on the relative distribution relationship of the start and end positions, trace the connection relationship between the beginning and end along the extension direction of the continuous line segments, determine the closed path formed by the connection of multiple line segments, verify the pixel connectivity of the enclosed area, and obtain the contour image of the closed area. S4: For each closed region in the closed region contour image, count the frequency of brightness change in the red, green and blue channels of each closed region, select the main response channel according to the sorting result formed between the channels based on the brightness change frequency, record the corresponding order, and form a list of main response features of the region.

[0006] As a further aspect of the present invention, the edge-aware channel sequence includes edge direction category, edge continuity identifier and edge arrangement index; the continuous line segment representation map specifically includes line segment set representation, line segment spatial position description and line segment direction consistency attribute; the closed region contour image includes closed contour unit, region boundary structure and region integrity identifier; and the region main response feature list specifically includes region identifier number, main response channel category and channel priority order.

[0007] As a further aspect of the present invention, the step of obtaining S1 is as follows: S101: Acquire multi-channel images covering map feature areas through remote sensing image acquisition equipment. Based on the spatial resolution identification information corresponding to each channel in the multi-channel images, divide the images into multiple scale levels consistent with the resolution level setting rules, and generate a unique number and location information mark for each scale level image to generate a scale level identification set. S102: Based on the channel category and corresponding image matrix labeled in each item of the scale level identifier set, the image matrix is ​​sequentially input into the perception structure of the deep learning network front end according to the execution order of the perception process, the structural feature images of the corresponding level are extracted, and an index path is established for each group of feature images to generate a feature map index path set. S103: Based on the image position covered by each path in the feature map index path set, the pixel value of the corresponding position is called according to the scanning direction of the horizontal path and the vertical path, the change trend of brightness value between adjacent pixels is tracked along the path, the pixel segment information with continuously changing brightness value is retained, and a pixel order mapping relationship is established to generate a brightness change pixel sequence. S104: Based on the pixel positions contained in each pixel segment in the brightness change pixel sequence, the gray value difference direction between adjacent pixels is compared relatively, pixel combinations with the same gray value change direction are selected, sequential numbers are established for pixel combinations in the horizontal and vertical scanning directions, and the numbering information is integrated to construct an overall trajectory sequence group to generate an edge-aware channel sequence.

[0008] As a further aspect of the present invention, the step of obtaining S2 is as follows: S201: Extract edge trajectory segments at adjacent positions in the edge sensing channel sequence. Based on the pixel sequence corresponding to each pair of trajectory segments at adjacent positions, call the brightness value set in each pixel sequence, extract the grayscale change trend of adjacent pixels according to the path scanning direction, make a corresponding judgment on the grayscale value change direction and trajectory direction, and generate a trajectory direction consistency comparison result. S202: Based on the trajectory combination marked as having the same direction and the same brightness change trend in the trajectory direction consistency comparison result, locate the start coordinates and end coordinates of each trajectory segment in the combination, connect adjacent trajectory segments in order according to the extension direction, establish a set of position information for the connected trajectory path according to the continuity of pixel points, and generate a sequence of line segment connection paths. S203: Call the start and end coordinates and direction markers of each path in the line segment connection path sequence, combine them with the two-dimensional spatial index layer in the image, map each path to a unified result layer in the processing order, check the consistency between the line segment direction of each path and the original connection direction, record the path number and mapping position, and generate a continuous line segment representation map.

[0009] As a further aspect of the present invention, the step of obtaining S3 is as follows: S301: Obtain the start and end positions of each continuous line segment in the continuous line segment representation diagram in the image, call the vector pair formed by the coordinates of the two ends, generate a direction mark sequence between the start and end positions for each continuous line segment, establish an extension and continuation sequence diagram between continuous line segments according to the relative distribution order between each direction mark, and generate a line segment direction connection relationship group. S302: Based on the line segment index corresponding to each group of marked paths in the line segment direction connection relationship group, along the line segment order arranged in each group of index paths, sequentially call the end point and start point coordinate values ​​of adjacent line segments, perform the start-end connection judgment on all line segments in the continuous connection path segment by segment, extract all line segment combination paths that can be completely closed, and generate a closed path contour sequence. S303: Based on the image region covered by each path in the closed path contour sequence, mark the pixel region enclosed by the corresponding contour in the original image matrix, extract the coordinate group of all pixels in the region, check the connection status between adjacent pixels in the four directions of up, down, left and right, filter out the region range with complete connectivity features, and generate a closed region contour image.

[0010] As a further aspect of the present invention, the step of obtaining S4 is as follows: S401: For each closed region in the closed region contour image, collect the set of pixel positions corresponding to the closed region in the red channel, green channel and blue channel, match the corresponding pixel distribution positions in each channel one by one according to the region boundary coordinates, collect them to form the correspondence between the channel and the closed region, and generate the set of channel pixel positions. S402: Based on the pixel position corresponding to the closed area in each channel in the set of channel pixel positions, extract the number of brightness changes of adjacent pixels in the order of continuous scanning, sort the brightness change frequencies in the red channel, green channel and blue channel, record the combination result of sorting position and channel type, and generate a channel brightness change frequency sequence. S403: Based on the channel type that ranks first in the channel brightness change frequency sequence, take it as the main response channel corresponding to the closed region. Match the channel number for each closed region, and summarize the corresponding channel types according to the region number to establish a list of main response channels for the region.

[0011] As a further aspect of the present invention, the method further includes: S5: Based on the main response channel information in the main response feature list of the region, locate the pixel region corresponding to the main response channel in the original image, obtain the set of pixel coordinates in the pixel region and match the channel element mapping rules, and establish the correspondence between the main response channel region and the preset map element category by combining the region connectivity and enclosure relationship, and obtain the map element recognition image. The map feature recognition image includes feature category labels, feature spatial coverage, and feature regional distribution layers.

[0012] As a further aspect of the present invention, the step of obtaining S5 is as follows: S501: Based on the main response channel information in the main response feature list of the region, locate the pixel region corresponding to each main response channel in the original remote sensing image, call the correspondence between the region number and the channel type to extract the pixel positions contained in all main response channels in the image, classify and collect the coordinate numbers according to the channel, and generate a set of channel positioning coordinates. S502: Based on the pixel position corresponding to each channel in the channel positioning coordinate set, call the channel type identifier and map feature attribute type in the channel and feature mapping rule base, perform mapping on the matching relationship between channel number and feature type, write all mapping relationships into the channel correspondence table, and generate the channel feature correspondence table. S503: Based on the established channel number and map element category relationship in the channel element correspondence table, and combined with the connected contour of each channel pixel coordinate set in the image, the connected region is marked according to the connected region number and assigned the corresponding map element category label, and mapped to the image spatial location of the original remote sensing image to generate a map element recognition image.

[0013] A deep learning-based map feature recognition system, the system comprising: The multi-scale edge perception module acquires multi-channel images covering the map feature area, constructs multi-scale layers according to resolution, extracts multi-level feature images, scans each layer of feature images row by row and column by column, extracts brightness coherent pixel segments, locates pixel fragments, acquires and encodes horizontal and vertical edge trajectories, and forms an edge perception channel sequence. The directional trajectory fusion module extracts adjacent edge trajectory segments from the edge perception channel sequence, compares the brightness change direction and trend, filters consistent edge trajectory pairs and merges them along the extension direction, connects adjacent edge trajectory segments into continuous line segments and records the start and end positions to obtain a continuous line segment representation map. The closed structure parsing module obtains the start and end positions of each continuous line segment in the continuous line segment representation diagram. Based on the relative distribution relationship of the start and end positions, it analyzes the spatial continuity relationship, traces the connection relationship between the beginning and end along the extension direction of the continuous line segments, determines the closed path formed by the connection of multiple line segments, verifies the pixel connectivity of the enclosed area, and obtains the contour image of the closed area. The region channel determination module, for each closed region in the closed region contour image, counts the brightness change frequency in the red, green and blue channels of each closed region, selects the main response channel according to the sorting result formed between the channels based on the brightness change frequency, records the corresponding order, and forms a list of region main response features. The feature type mapping module locates the pixel region corresponding to the main response channel in the original image based on the main response channel information in the region main response feature list, obtains the set of pixel coordinates within the pixel region and matches the channel feature mapping rules, and establishes the correspondence between the main response channel region and the preset map feature category by combining the region connectivity and enclosure relationship, thereby obtaining the map feature recognition image.

[0014] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: In this invention, a regional expression method with directional consistency and continuous feature constraints is formed through the collaborative organization of multi-scale spatial information, enabling map elements to present a stable overall outline shape in complex backgrounds. By combining the description of boundary trends and regional closure characteristics, the complete presentation of the spatial structure of elements is enhanced. At the same time, the channel main response difference is introduced to characterize regional attribute features, so that element differentiation is based on the overall response characteristics of the region rather than isolated feature judgments. This maintains a clear differentiation effect in scenarios with multiple adjacent elements or similar textures, and improves the credibility of the recognition results in terms of spatial consistency and attribute matching. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a flowchart illustrating the acquisition process of S1 in this invention; Figure 3 This is a flowchart illustrating the acquisition process of S2 in this invention; Figure 4 This is a flowchart illustrating the acquisition process of S3 in this invention; Figure 5 This is a flowchart illustrating the acquisition process of S4 in this invention; Figure 6 This is a flowchart of the acquisition process for S5 of the present invention. Detailed Implementation

[0016] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0017] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0018] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0019] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0020] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0021] Please see Figure 1 This invention provides a technical solution: a map feature recognition method based on deep learning, comprising the following steps: S1: Acquire multi-channel images covering map feature areas using remote sensing image acquisition equipment. Divide the multi-channel images into multiple scale levels according to spatial resolution. Input the images of each scale level into the front-end perception process of the deep learning network in sequence to obtain multi-level feature images. Perform row-by-row and column-by-column scanning on each feature image to extract pixel segments with continuous brightness changes in the scanning path. Locate pixel fragments that continuously change along the same scanning direction. Locate edge trends with consistent brightness gradient directions in pixel fragments. Extract horizontal and vertical edge trajectories respectively. Encode them in the order of appearance to form an edge perception channel sequence. S2: Extract edge trajectory segments at adjacent positions in the edge-aware channel sequence, compare and analyze the brightness change direction and trend of each pair of edge trajectory segments, screen out edge trajectory pairs with consistent direction and the same brightness change trend, perform trajectory merging along the trajectory extension direction, connect adjacent edge trajectory segments into line segments with continuous direction, and represent them in the image as an overall path, record the start and end positions of the line segments in the image, keep the initial direction and change characteristics unchanged during the extension of the line segments, and import them into a unified result layer according to the processing order to obtain a continuous line segment representation map; S3: Obtain the start and end positions of each continuous line segment in the continuous line segment representation diagram. Based on the relative distribution relationship between the start and end positions, analyze the spatial continuity relationship of different continuous line segments. Trace the connection relationship between the start and end of each line segment along the extension direction of the continuous line segment itself to determine the closed path formed by the sequential connection of multiple continuous line segments. Summarize the closed path as a set of closed contour line segments. Check the connectivity of pixel distribution within the image range enclosed by the set of closed contour line segments to determine the contour range of the complete closed region and obtain the contour image of the closed region. S4: For each closed region in the closed region contour image, extract the corresponding position of each closed region in the red, green and blue channels, count the brightness change frequency of the red, green and blue channels within the region, select the corresponding channel type according to the sorting result formed by the brightness change frequency among the channels, determine the selected channel as the main response channel of the current closed region, record the order of the main response channels corresponding to each closed region, and form a list of region main response features; S5: Based on the main response channel information in the regional main response feature list, locate the pixel region corresponding to the main response channel in the original remote sensing image, obtain the set of pixel coordinates within the pixel region, call the preset channel and feature mapping rule library, match the map feature attributes represented by different main response channel types, establish the correspondence between the main response channel region and the preset map feature category based on the connectivity relationship and the enclosed area range of the pixel coordinate set in the original remote sensing image, generate the classification result of the map feature region in the image space, and obtain the map feature recognition image.

[0022] The edge-aware channel sequence includes edge direction category, edge continuity identifier, and edge arrangement index. The continuous line segment representation map specifically includes line segment set representation, line segment spatial location description, and line segment direction consistency attribute. The closed region contour image includes closed contour unit, region boundary structure, and region integrity identifier. The region main response feature list specifically includes region identifier number, main response channel category, and channel priority order. The map feature recognition image includes feature category label, feature spatial coverage, and feature region distribution layer.

[0023] Please see Figure 2 The steps to obtain S1 are as follows: S101: Acquire multi-channel images covering map feature areas through remote sensing image acquisition equipment. Based on the spatial resolution identification information corresponding to each channel in the multi-channel images, divide the images into multiple scale levels consistent with the resolution level setting rules, and generate a unique number and location information mark for each scale level image to generate a scale level identification set. Multi-channel images covering map feature areas are acquired using satellite or aerial remote sensing image acquisition equipment equipped with CCD or CMOS sensors. The metadata header file (such as GeoTIFF Tags) contained in the multi-channel images is read, and the ground sampling distance (GSD) parameter recorded therein is extracted as the initial spatial resolution value. For example, the initial resolution of the acquired red channel image is 0.5 meters. If the preset resolution level setting rule is downsampling by powers of 2, the original image is marked as level 0. The rows and columns of the image matrix are traversed to... Using the pixel neighborhood as the unit, the arithmetic mean of the gray values ​​of four pixels within the neighborhood is calculated using the mean pooling technique. The average value is used as the pixel value for the corresponding position in the next level, generating a first-level image with a resolution of 1.0 meter. The above pixel aggregation calculation process is repeated until the preset maximum number of levels is reached. (For example This constructs an image pyramid with four levels: 0.5 meters, 1.0 meters, 2.0 meters, and 4.0 meters. For each generated scale level image, a unique level index key is constructed by combining its corresponding channel spectral type identifier and level depth number. For example, the first level image of the red channel is labeled as "". Meanwhile, the starting and ending storage addresses of the image at this level in the video memory space are obtained through the operating system's memory management interface. The storage address range is used as a location information marker and bound to the level index key value. This operation is repeated for all channels and all levels. The generated key-value pairs are written into a unified list file to generate a scale level identifier set.

[0024] S102: Based on the channel category and corresponding image matrix labeled in each item of the scale-level identifier set, the image matrix is ​​sequentially input into the perception structure of the deep learning network front end according to the execution order of the perception process, the structural feature images of the corresponding level are extracted, and an index path is established for each set of feature images to generate a feature map index path set. Based on the channel category and corresponding image matrix labeled for each item in the scale-level identifier set, the memory address range stored in the list is read. The corresponding image matrix data is loaded into the video memory of a high-performance computing unit (such as a GPU or TPU) via parallel computing interfaces such as CUDA or OpenCL. The convolutional operation layer of the deep learning network front-end is initialized, and the size parameters of the convolutional kernel are set (e.g., using...). The convolution kernel is set to a stride parameter of 1. The loaded image matrix is ​​used as the input tensor, and a sliding window multiplication-add operation is performed between the convolution kernel and the input tensor in the two-dimensional plane. That is, the weight parameters in the convolution kernel are used to perform the operation. Pixel values ​​at positions corresponding to the image matrix Perform point-by-point multiplication and summation ,in The bias parameters are used to process the results through a non-linear activation function, such as the ReLU function. Negative values ​​are set to zero to increase sparsity, and a numerical matrix reflecting the image texture intensity is output. This process is performed in parallel with multiple different convolutional kernel parameters, generating multiple feature maps. For each output feature map, its dimension index in the network output tensor is obtained, for example, the index of the nth feature map. The first layer Each channel feature map records its dimensional coordinates. Simultaneously, the pointer handle of the feature map in the video memory is recorded. The dimension coordinates and pointer handle are combined into an index record. According to the connection order of the network layers, the index records of all feature maps are appended to a dynamic array to generate a set of feature map index paths.

[0025] S103: Based on the image position covered by each path in the feature map index path set, call the corresponding pixel value according to the scanning direction of the horizontal and vertical paths respectively, track the change trend of brightness value between adjacent pixels along the path, retain the pixel segment information with continuously changing brightness value, establish the pixel order mapping relationship, and generate a brightness change pixel sequence. Based on the image location covered by each path in the feature map index path set, the corresponding feature map data in the video memory is accessed using the pointer handle in the index record. A normalized scan threshold parameter is set, for example, a brightness variation threshold of 15 (for a quantization range of 0-255). The horizontal scan process is then initiated, with the row coordinates fixed. Make column coordinates Read the current coordinates sequentially, incrementing from 0 to the feature map width value. and adjacent coordinates The pixel brightness values ​​are used to perform a difference operation to obtain the absolute value of the brightness difference: The absolute value is compared with the brightness change threshold of 15. If... If a significant brightness change is detected at the current location, the current pixel coordinates are immediately recorded. The data is entered into a temporary buffer queue, and the next adjacent pixel is read for difference calculation. If the absolute value of the difference between consecutive subsequent pixels is greater than or equal to 15, the coordinates are continuously added to the same queue until a pixel with an absolute difference value less than 15 is encountered. At this point, the recording of the current segment stops, and all coordinates in the buffer queue are packaged into a single pixel segment unit. If the queue is empty, packaging is not performed, the buffer queue is reset, and the search for the next starting point that meets the conditions continues along the scanning direction. After the horizontal scan is completed, the system switches to the vertical scan mode and fixes the column coordinates. Variable row coordinates Repeat the above differential alignment and fragment extraction actions. For each extracted pixel fragment unit, assign an incremental sequence number to its internal pixels according to the order in which the pixels appear on the scanning path, establish a pixel order mapping relationship, and generate a brightness variation pixel sequence.

[0026] S104: Based on the pixel positions contained in each pixel segment in the brightness change pixel sequence, the gray value difference direction between adjacent pixels is compared relatively, and pixel combinations with the same gray value change direction are selected. Sequential numbers are established for pixel combinations in the horizontal and vertical scanning directions, and the numbering information is integrated to construct an overall trajectory sequence group to generate an edge-aware channel sequence. Based on the pixel positions contained in each pixel segment of the brightness variation pixel sequence, each pixel within the segment is traversed, and the grayscale value of that pixel is read again along the scanning direction and its subsequent pixel value is compared with the grayscale value. A signed subtraction operation is then performed to calculate the grayscale gradient value. Determine the sign of the gradient value. Marked as "Brightening Trend", if Marked as "Darkening Trend", if The segment is then marked as "stable," and a consistency length threshold is set, for example, 5 pixels. The gradient signs of consecutive pixels within the segment are checked to ensure they remain consistent. If a segment of consecutive pixels has all positive gradient signs and the number exceeds 5, it is truncated and confirmed as a valid monotonic edge combination. Similarly, combinations with all consecutive negative signs are filtered out. Scattered pixel segments that do not meet the length threshold or have alternating signs are removed (i.e., not included in subsequent processing). For each selected monotonic edge combination, a unique identifier is assigned from two independent counters based on its horizontal or vertical scan source. For example, horizontal combinations are assigned a unique identifier starting from two independent counters. "Start with an incrementing number, and combine vertically with " "Add incrementing numbers at the beginning, and reorder and stitch the edge combinations with assigned identification codes according to the spatial distribution coordinates of the original image to form a list containing all valid edge information, generating an edge-aware channel sequence."

[0027] Please see Figure 3 The steps to obtain S2 are as follows: S201: Extract edge trajectory segments at adjacent positions in the edge-aware channel sequence. Based on the pixel sequence corresponding to each pair of trajectory segments at adjacent positions, call the brightness value set in each pixel sequence, extract the grayscale change trend of adjacent pixels according to the path scanning direction, and make a corresponding judgment on the grayscale value change direction and trajectory direction to generate trajectory direction consistency comparison results. Extract edge trajectory segments from adjacent positions in the edge-aware channel sequence, read the pixel coordinate list contained in each extracted trajectory segment, and set the neighborhood determination radius, for example, by setting a radius. Pixels, using the Euclidean distance formula Calculate the distance between the endpoints of different trajectory segments. If the distance between the nearest endpoints of two trajectory segments A and B is... If the segments are spatially adjacent, they are marked as candidate pairs. For each candidate trajectory segment, the pixel grayscale value sequence contained therein is read, and a first-order difference operation is performed on the sequence to obtain the grayscale gradient vector. The sign of all elements in the gradient vector is counted. If the proportion of positive gradients in the sequence exceeds 80%, the segment is determined to be "brightness-increasing type"; if the proportion of negative gradients exceeds 80%, it is determined to be "darkness-decreasing type"; otherwise, it is determined to be "mixed type". The geometric slope of the trajectory segment in the image coordinate system is extracted, and the coordinates of each pixel are calculated using the arctangent function. The tangent angle is used to calculate the average angle of the entire segment as the trajectory direction angle. For example, if the direction angle of segment A is 45 degrees and the direction angle of segment B is 48 degrees, the absolute value of the angle difference between the two is calculated, and an angle tolerance threshold (e.g., 10 degrees) is set. The angle is compared with the threshold of 10 degrees. If the angle difference is less than 10 degrees and the grayscale change type (brightness increase or darkness decrease) of the two are exactly the same, it is determined that the two have the connection condition and the comparison result is recorded as the Boolean value "True". If any condition is not met (such as distance exceeding the limit, different type or large angle difference), it is recorded as "False" and the pair is skipped. The above calculation is completed by traversing all candidate pairs to generate the trajectory direction consistency comparison result.

[0028] S202: Based on the trajectory combination marked as having the same direction and the same brightness change trend in the trajectory direction consistency comparison result, locate the start coordinates and end coordinates of each trajectory segment in the combination, connect adjacent trajectory segments in order according to the extension direction, establish a set of position information for the connected trajectory path according to the pixel continuity, and generate a sequence of line segment connection paths. Based on the trajectory direction consistency comparison results, the trajectory combinations marked as having the same direction and the same brightness change trend are read from the previous trajectory segment of the combination. Termination coordinates With the next trajectory segment starting coordinates Using the Bresenham linear interpolation algorithm, calculate and The coordinates of the missing pixels, for example The value is (10,10). If the coordinates are (12, 12), then the interpolation calculation yields the intermediate point (11, 11). The interpolated pixels are then sequentially inserted into the... The end and Between the beginning and end of the list, construct a new list of merged pixels and check the continuity of pixel coordinates in the merged list, that is, confirm that any two adjacent pixel coordinates are connected in their 8-neighborhood (i.e., the coordinate difference is within 8 ohms). (Within the set), if breakpoints exist, interpolation is performed again to fill in the gaps, and a unique path ID is reassigned to the merged new path, for example, by assigning the ID number " "Keep the pre-merger version" The starting point is used as the starting point of the new path, and is retained. The endpoint is used as the new path endpoint. All merged path data is stored in a structured manner. Each record contains a path ID, an ordered set of pixel coordinates, and a timestamp of the merging operation, generating a sequence of line segment connections.

[0029] S203: Call the start and end coordinates and direction markers of each path in the line segment connection path sequence, combine them with the two-dimensional spatial index layer in the image, map each path to the unified result layer in the processing order, check the consistency between the line segment direction of each path and the original connection direction, record the path number and mapping position, and generate a continuous line segment representation. The algorithm calls the start and end coordinates and direction markers of each path in the line segment connection path sequence, creates an all-zero matrix with the same size as the original image as the result layer, sets the background value to 0, and the line value to 255, and iterates through each path in the sequence to read all the pixel coordinates it contains. The values ​​at the corresponding row and column positions in the resulting layer matrix are modified from 0 to 255 to complete the rasterization mapping of the path. During the mapping process, the gradient direction data of the corresponding coordinates in the original image are read again, and the angle between the local tangent direction of the current mapped line segment and the original gradient direction is calculated. A consistency check threshold (e.g., 30 degrees) is set. If more than 20% of the pixels on a certain path have an angle greater than 30 degrees between their tangent direction and the original gradient direction, the path is determined to have a drift error, and the pixel value of the path in the resulting layer is reset to 0 and discarded (considered as noise). For paths that pass the check, their path ID and the bounding box range in the layer are recorded in the layer's metadata table, i.e., the minimum bounding box range is recorded. ,maximum Minimum ,maximum Four boundary values ​​are used to solidify all the remaining line segments in the form of binary lines in the layer, generating a continuous line segment representation diagram.

[0030] Please see Figure 4 The steps to obtain S3 are as follows: S301: Obtain the start and end positions of each continuous line segment in the continuous line segment representation diagram in the image, call the vector pair formed by the coordinates of the two ends, generate a direction mark sequence between the start and end positions for each continuous line segment, establish an extension and continuation sequence diagram between continuous line segments according to the relative distribution order between each direction mark, and generate a line segment direction connection relationship group. To obtain the start and end positions of each continuous line segment in the continuous line segment representation diagram, iterate through each identified line segment object in the image and read the start coordinates recorded in its metadata. and termination coordinates Calculate the displacement vector using vector subtraction. The azimuth angle of the vector relative to the horizontal axis is calculated using the atan2 function. This is then converted into an angle value. For example, if a line segment starts at (100, 100) and ends at (150, 150), its azimuth angle is 45 degrees. This angle value is used as a direction marker for the line segment. An adjacency matrix or directed graph data structure is constructed, where nodes represent line segment IDs, edges represent continuity relationships, and a continuity distance threshold is set. (e.g., 5 pixels) and the angle threshold between the directions (e.g., 30 degrees), for any two line segments and ,calculate The end point and Euclidean distance between the starting points ,as well as End tangential and Angle between the tangents at the beginning ,like and Then determine to If a succession relationship exists, set the corresponding position in the adjacency matrix to 1; And calculate the normalized weights. (Ensure that the dimensions are consistent and the range is between 0 and 1), traverse all line segments to complete the matrix filling, perform depth-first search (DFS) for each node, find subsequent connection nodes according to the principle of minimum weight, arrange the searched line segment ID sequence in order to form one or more potential long path chains, record the flow order of line segments in each chain, and generate line segment direction connection relationship groups.

[0031] S302: Based on the line segment index corresponding to each group of marked paths in the line segment direction connection relationship group, along the line segment order arranged in each group of index paths, sequentially call the end point and start point coordinate values ​​of adjacent line segments, perform the start-end connection judgment on all line segments in the continuous connection path segment by segment, extract all line segment combination paths that can be completely closed, and generate a closed path outline sequence. Based on the segment index corresponding to each marked path in the segment direction connection relationship group, initialize an empty closed path list. For each segment index chain, read the first segment in the chain in sequence. starting coordinates With the last line segment End point coordinates Calculate the spatial Euclidean distance between the two endpoints. Set a closure threshold (For example, set to 3 pixels based on image resolution), and Perform a numerical comparison; if If the chain is geometrically closed at both ends, it is directly marked as a candidate for closure. Then the completion detection procedure will be started. The finish line Search for the radius to see if there are any ungrouped isolated line segments; if so, connect them. End point and The intermediate line segment at the starting point, and the length of the intermediate line segment is less than... If a suitable intermediate segment is not found, the intermediate segment is inserted at the end of the chain to achieve logical closure. If no suitable intermediate segment is found, the chain is marked as "open" and removed from the closure candidate list. For chains that are confirmed to be closed, the number of segments they contain is checked. If the number is less than 3, they are considered invalid noise loops and are removed. Combinations with a number of segments greater than or equal to 3 are retained. All line segment coordinates in the valid closure combinations are reordered in a clockwise or counterclockwise direction to eliminate coordinate jumps, form a continuous set of contour points, and generate a closed path contour sequence.

[0032] S303: Based on the image region covered by each path in the closed path contour sequence, mark the pixel region enclosed by the corresponding contour in the original image matrix, extract the coordinate group of all pixels in the region, check the connection status between adjacent pixels in the four directions of up, down, left and right, screen out the region range with complete connectivity features, and generate a closed region contour image. Based on the image region covered by each path in the closed path contour sequence, a mask matrix of the same size as the original image is created, initially set to all zeros. For each set of closed contour points, a scan-line polygon fill algorithm is used to calculate the coordinates of the intersection points of the contour boundaries on each scan line. The intersection points are then processed according to... The coordinates are sorted, and the pixel intervals between each pair of intersections are marked as 1 in the mask matrix (representing the internal region). The set of pixel coordinates with a value of 1 in the mask matrix is ​​extracted, and the connectivity check window is set to... Iterate through each pixel in the set. Check if the pixels in its 8-neighborhood are also in the set, and count the number of pixels with a value of 1 in the neighborhood. ,like The point is determined to be an isolated noise point and its value is reset to 0. If the total pixel area of ​​a connected region (i.e. the total number of pixels with a value of 1) is less than the set minimum area threshold (e.g., 50 pixels), the entire region is removed (i.e., regarded as a non-target small region). Large connected blocks with a qualified area and no holes are retained. Each retained connected region is assigned a unique region ID label, such as region 1, region 2, etc. The region boundary pixel coordinates and the internal filling pixel coordinates are recorded in two independent arrays to generate a closed region contour image.

[0033] Please see Figure 5 The steps to obtain S4 are as follows: S401: For each closed region in the closed region contour image, collect the set of pixel positions corresponding to the closed region in the red, green and blue channels, match the corresponding pixel distribution positions in each channel one by one according to the region boundary coordinates, collect them to form the correspondence between the channel and the closed region, and generate the set of channel pixel positions. For each closed region in the closed region contour image, read the unique identifier ID of each region and its corresponding internal filling mask. Load the red (R), green (G), and blue (B) spectral channel data matrix of the original remote sensing image acquired by the multispectral imaging sensor. Using NumPy array slicing or similar matrix indexing operations, perform a logical AND operation between the region mask and the data matrix of the three channels respectively, and extract the pixel grayscale value at the corresponding position with a value of True within the mask range. For example, for ID " The mask covers a rectangular area from (50,50) to (100,100) in the image. The system will extract 2500 pixel values ​​from this range in the R channel, 2500 pixel values ​​from this range in the G channel, and 2500 pixel values ​​from this range in the B channel. Each group of extracted grayscale values ​​will be flattened into a one-dimensional array in row-major order and labeled as follows. , , And record the global coordinates of each pixel in the original image. Local index within the region Using the region ID as the primary key and three one-dimensional arrays R, G, and B as values, a dictionary or hash table structure is constructed for storage. This extraction and binding process is repeated for all identified closed regions to generate a set of channel pixel positions.

[0034] S402: Based on the pixel position corresponding to the closed region in each channel in the channel pixel position set, extract the number of brightness changes of adjacent pixels in the order of continuous scanning, sort the brightness change frequency in the red channel, green channel and blue channel, record the combination result of sorting position and channel type, and generate a channel brightness change frequency sequence. Based on the pixel positions corresponding to the closed regions within each channel in the channel pixel position set, differential statistics are performed on the three sets of one-dimensional pixel grayscale arrays (R, G, and B) for each region, and a brightness jump threshold is set. (For example, set to 20 gray levels), iterate through adjacent elements in the array and calculate the absolute value of the difference. If the difference is greater than 20, the counter is incremented by 1, and the total number of transitions in the entire array is counted to obtain the frequency values ​​of the three channels. , , For example, the calculation result for a certain vegetation area is , , Introducing channel information entropy calculation; Using Shannon's entropy formula Calculate the entropy value of the grayscale histogram for each channel. , , (The result is the number of dimensionless bits). Construct a comprehensive scoring index (Multiplying by 100 here is to adjust the magnitude of the entropy value to be approximately the same as the order of counts, so that the weighted sum has numerical significance.) Assume the calculated value is... , , The three scores are sorted from largest to smallest, resulting in a sorted result of (G,R,B). This sorted result is then bound to the channel name, for example, recorded as [('Green',480),('Red',180),('Blue',150)]. This statistical, scoring, and sorting process is performed on each closed region, and the resulting list of ordered tuples is stored to generate a channel brightness change frequency sequence.

[0035] S403: Based on the channel type that ranks first in the channel brightness change frequency sequence, use it as the main response channel corresponding to the closed region. Match the channel number for each closed region and summarize the corresponding channel types according to the region number to establish a list of main response channels for the region. Based on the channel type ranked first in the channel brightness change frequency sequence, read the first element of the sorted list corresponding to each region, such as ('Green', 480) in the example above. Extract its channel label "Green" as the main response feature of that region. Build a mapping lookup table containing a region ID field and a main response channel field. Fill the extracted "Green" label into the ID field. Significance verification was performed on the records of "". Calculate the ratio of the score of the first channel to the score of the second channel. Set a significance threshold If it is 1.2, If so, the first channel is directly confirmed as the primary response channel. If the region is identified as a "mixed response" or "transition region", the first and second channels are combined and marked, for example, "Green-Red_Mix". All region IDs are traversed, and the verified final channel labels (such as "Green", "Red", "Blue" or "Mix") are written to the database or memory list in sequence to ensure that each region has a unique spectral attribute description and generate a list of main response channels for the region.

[0036] Please see Figure 6 The steps to obtain S5 are as follows: S501: Based on the main response channel information in the regional main response feature list, locate the pixel region corresponding to each main response channel in the original remote sensing image, call the correspondence between the region number and the channel type to extract the pixel positions contained in all main response channels in the image, classify and collect the coordinate numbers according to the channel, and generate a set of channel positioning coordinates. Based on the main response channel information in the regional main response feature list, the pixel region corresponding to each main response channel is located in the original remote sensing image. Each region entry recorded in the main response feature list is traversed, and its region ID and main response channel label (e.g., "Green") are read. The region ID is used to index the closed region contour image generated in the previous step, and the corresponding binarized mask matrix is ​​extracted. Using the mask matrix as a Boolean indexer, the global coordinates of all pixels with a mask value of True are selected in the global coordinate system of the original image. Construct a list of coordinate pairs, for example, for the area marked "Green". Extract the coordinates {(10,10),(10,11),…,(20,20)}, create a hash dictionary with channel type as the key, and merge all extracted coordinate lists according to category keys such as "Red", "Green", "Blue", "Mix", etc. If the same channel type corresponds to multiple unconnected region IDs, merge the coordinate lists of these regions and store them in a large list, retaining their respective region IDs as sub-labels. For example, the dictionary structure is as follows: {'Green':[{ :[(10,10)…]},{ After completing the coordinate aggregation of all regions, the number of coordinate points under each channel category is counted. If the total number of pixels in a certain category is lower than the set noise threshold (e.g., 10 pixels), the category is considered invalid noise, and the coordinate data of the category is discarded. Only the coordinate set of the valid categories is retained to generate the channel positioning coordinate set.

[0037] S502: Based on the pixel position corresponding to each channel in the channel positioning coordinate set, call the channel type identifier and map feature attribute type in the channel and feature mapping rule base, perform mapping on the matching relationship between channel number and feature type, write all mapping relationships into the channel correspondence table, and generate the channel feature correspondence table. Based on the pixel positions corresponding to each channel in the channel positioning coordinate set, a pre-built channel-feature mapping rule base file constructed based on spectral feature indices or expert prior knowledge is loaded. This file stores the association logic between spectral channels and land cover types in key-value pairs, such as rule items: Rule_1:{'Channel':'Green','Feature':'Vegetation','Confidence':0.9}; Rule_2:{'Channel':'Red','Feature':'Artificial_Surface','Confidence':0.85}; Rule_3:{'Channel':'Blue','Feature':'Water_Body','Confidence':0.88}; Iterate through each channel key (e.g., "Green") in the channel location coordinate set, search for a matching Rule item in the rule base, extract the corresponding 'Feature' field value (e.g., "Vegetation"), and assign the feature type string to all area IDs under the current channel. If a channel label of type "Mix" is encountered, the secondary discrimination logic is activated, and the combined mapping table in the rule base is called, for example, {'Mix_Green_Red':'Agricultural_Land'}. The corresponding composite feature type is found according to the specific mixed component. If there is no direct match in the rule base, it is marked as "Unclassified". A new relation table is constructed, and each record contains the channel number, the original channel name, the mapped map feature category name, and the confidence score of the mapping rule. For example, the record row is: [Channel_ID:1,Name:'Green',Type:'Vegetation',Score:0.9]. All mapping results are appended to the table line by line to generate the channel feature correspondence table.

[0038] S503: Based on the established channel number and map element category relationship in the channel element correspondence table, and combined with the connected contour of each channel pixel coordinate set in the image, the connected region is marked according to the connected region number and assigned the corresponding map element category label, and mapped to the image spatial location of the original remote sensing image to generate a map element recognition image. Based on the established channel number and map feature category relationship in the channel feature correspondence table, a blank single-channel label image matrix with the same resolution as the original remote sensing image is created, and all pixel values ​​are initialized to 0 (representing the background). Each record in the correspondence table is traversed, and its map feature category name's corresponding integer code (e.g., Vegetation=1, Water=2, Artificial=3) and associated region ID list are read. For each region ID in the list, the corresponding global pixel coordinates are retrieved from the channel positioning coordinate set. The pixel values ​​at these coordinate positions in the label image matrix are modified to the corresponding feature codes. For example, all pixels belonging to vegetation areas are assigned a value of 1. During the assignment process, boundary conflict checks are performed. If a pixel has already been assigned a non-zero value (i.e., there is overlap), the confidence scores of the old and new feature categories are compared, and the label of the category with the higher score is retained. After completing the pixel assignment for all regions, morphological closing operations are performed on the label image. The structural elements are expanded and eroded to fill the small unclassified voids (isolated points with a value of 0) inside the region. Finally, the integer encoding matrix is ​​converted into a visual RGB color mask. Colors are assigned to different encoding values ​​according to a preset legend color table (such as vegetation = green, water = blue). The mask is then overlaid on the original image to generate a map feature recognition image.

[0039] A deep learning-based map feature recognition system, comprising: The multi-scale edge perception module acquires multi-channel images covering the map feature area, constructs multi-scale layers according to resolution, extracts multi-level feature images, scans each layer of feature images row by row and column by column, extracts brightness coherent pixel segments, locates pixel fragments, acquires and encodes horizontal and vertical edge trajectories, and forms an edge perception channel sequence. The directional trajectory fusion module extracts adjacent edge trajectory segments from the edge perception channel sequence, compares the brightness change direction and trend, selects consistent edge trajectory pairs and merges them along the extension direction, connects adjacent edge trajectory segments into continuous line segments and records the start and end positions to obtain a continuous line segment representation map. The closed structure parsing module obtains the start and end positions of each continuous line segment in the continuous line segment representation diagram. Based on the relative distribution relationship of the start and end positions, it analyzes the spatial continuity relationship, traces the connection relationship between the beginning and end along the extension direction of the continuous line segments, determines the closed path formed by the connection of multiple line segments, verifies the pixel connectivity of the enclosed area, and obtains the contour image of the closed area. The region channel determination module calculates the frequency of brightness changes in the red, green, and blue channels of each closed region in the closed region contour image, selects the main response channel based on the sorting results formed between the channels according to the brightness change frequency, records the corresponding order, and forms a list of region main response features. The feature type mapping module locates the pixel region corresponding to the main response channel in the original image based on the main response channel information in the regional main response feature list, obtains the set of pixel coordinates within the pixel region and matches the channel feature mapping rules, and establishes the correspondence between the main response channel region and the preset map feature category by combining the regional connectivity and enclosure relationship, thus obtaining the map feature recognition image.

[0040] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A map feature recognition method based on deep learning, characterized in that, Includes the following steps: S1: Acquire multi-channel images of the covered map feature area, construct multi-scale layers according to resolution, extract multi-level feature images, scan each layer of feature images row by row and column by column, extract brightness coherent pixel segments, locate pixel fragments, acquire and encode horizontal and vertical edge trajectories, and form an edge-aware channel sequence. S2: Extract adjacent edge trajectory segments from the edge sensing channel sequence, compare the brightness change direction and trend, select consistent edge trajectory pairs and merge them along the extension direction, connect adjacent edge trajectory segments into continuous line segments and record the start and end positions to obtain a continuous line segment representation map. S3: Obtain the start and end positions of each continuous line segment in the continuous line segment representation diagram, analyze the spatial continuity relationship based on the relative distribution relationship of the start and end positions, trace the connection relationship between the beginning and end along the extension direction of the continuous line segments, determine the closed path formed by the connection of multiple line segments, verify the pixel connectivity of the enclosed area, and obtain the contour image of the closed area. S4: For each closed region in the closed region contour image, count the frequency of brightness change in the red, green and blue channels of each closed region, select the main response channel according to the sorting result formed between the channels based on the brightness change frequency, record the corresponding order, and form a list of main response features of the region.

2. The map feature recognition method based on deep learning according to claim 1, characterized in that: The edge-aware channel sequence includes edge direction category, edge continuity identifier, and edge arrangement index. The continuous line segment representation map specifically includes line segment set representation, line segment spatial location description, and line segment direction consistency attribute. The closed region contour image includes closed contour unit, region boundary structure, and region integrity identifier. The region main response feature list specifically includes region identifier number, main response channel category, and channel priority order.

3. The map feature recognition method based on deep learning according to claim 1, characterized in that: The steps for obtaining S1 are as follows: S101: Acquire multi-channel images covering map feature areas through remote sensing image acquisition equipment. Based on the spatial resolution identification information corresponding to each channel in the multi-channel images, divide the images into multiple scale levels consistent with the resolution level setting rules, and generate a unique number and location information mark for each scale level image to generate a scale level identification set. S102: Based on the channel category and corresponding image matrix labeled in each item of the scale level identifier set, the image matrix is ​​sequentially input into the perception structure of the deep learning network front end according to the execution order of the perception process, the structural feature images of the corresponding level are extracted, and an index path is established for each group of feature images to generate a feature map index path set. S103: Based on the image position covered by each path in the feature map index path set, the pixel value of the corresponding position is called according to the scanning direction of the horizontal path and the vertical path, the change trend of brightness value between adjacent pixels is tracked along the path, the pixel segment information with continuously changing brightness value is retained, and a pixel order mapping relationship is established to generate a brightness change pixel sequence. S104: Based on the pixel positions contained in each pixel segment in the brightness change pixel sequence, the gray value difference direction between adjacent pixels is compared relatively, pixel combinations with the same gray value change direction are selected, sequential numbers are established for pixel combinations in the horizontal and vertical scanning directions, and the numbering information is integrated to construct an overall trajectory sequence group to generate an edge-aware channel sequence.

4. The map feature recognition method based on deep learning according to claim 1, characterized in that: The steps for obtaining S2 are as follows: S201: Extract edge trajectory segments at adjacent positions in the edge sensing channel sequence. Based on the pixel sequence corresponding to each pair of trajectory segments at adjacent positions, call the brightness value set in each pixel sequence, extract the grayscale change trend of adjacent pixels according to the path scanning direction, make a corresponding judgment on the grayscale value change direction and trajectory direction, and generate a trajectory direction consistency comparison result. S202: Based on the trajectory combination marked as having the same direction and the same brightness change trend in the trajectory direction consistency comparison result, locate the start coordinates and end coordinates of each trajectory segment in the combination, connect adjacent trajectory segments in order according to the extension direction, establish a set of position information for the connected trajectory path according to the continuity of pixel points, and generate a sequence of line segment connection paths. S203: Call the start and end coordinates and direction markers of each path in the line segment connection path sequence, combine them with the two-dimensional spatial index layer in the image, map each path to a unified result layer in the processing order, check the consistency between the line segment direction of each path and the original connection direction, record the path number and mapping position, and generate a continuous line segment representation map.

5. The map feature recognition method based on deep learning according to claim 1, characterized in that: The steps for obtaining S3 are as follows: S301: Obtain the start and end positions of each continuous line segment in the continuous line segment representation diagram in the image, call the vector pair formed by the coordinates of the two ends, generate a direction mark sequence between the start and end positions for each continuous line segment, establish an extension and continuation sequence diagram between continuous line segments according to the relative distribution order between each direction mark, and generate a line segment direction connection relationship group. S302: Based on the line segment index corresponding to each group of marked paths in the line segment direction connection relationship group, along the line segment order arranged in each group of index paths, sequentially call the end point and start point coordinate values ​​of adjacent line segments, perform the start-end connection judgment on all line segments in the continuous connection path segment by segment, extract all line segment combination paths that can be completely closed, and generate a closed path contour sequence. S303: Based on the image region covered by each path in the closed path contour sequence, mark the pixel region enclosed by the corresponding contour in the original image matrix, extract the coordinate group of all pixels in the region, check the connection status between adjacent pixels in the four directions of up, down, left and right, filter out the region range with complete connectivity features, and generate a closed region contour image.

6. The map feature recognition method based on deep learning according to claim 1, characterized in that: The steps for obtaining S4 are as follows: S401: For each closed region in the closed region contour image, collect the set of pixel positions corresponding to the closed region in the red channel, green channel and blue channel, match the corresponding pixel distribution positions in each channel one by one according to the region boundary coordinates, collect them to form the correspondence between the channel and the closed region, and generate the set of channel pixel positions. S402: Based on the pixel position corresponding to the closed area in each channel in the set of channel pixel positions, extract the number of brightness changes of adjacent pixels in the order of continuous scanning, sort the brightness change frequencies in the red channel, green channel and blue channel, record the combination result of sorting position and channel type, and generate a channel brightness change frequency sequence. S403: Based on the channel type that ranks first in the channel brightness change frequency sequence, take it as the main response channel corresponding to the closed region. Match the channel number for each closed region, and summarize the corresponding channel types according to the region number to establish a list of main response channels for the region.

7. The map feature recognition method based on deep learning according to claim 1, characterized in that: The method further includes: S5: Based on the main response channel information in the main response feature list of the region, locate the pixel region corresponding to the main response channel in the original image, obtain the set of pixel coordinates in the pixel region and match the channel element mapping rules, and establish the correspondence between the main response channel region and the preset map element category by combining the region connectivity and enclosure relationship, and obtain the map element recognition image. The map feature recognition image includes feature category labels, feature spatial coverage, and feature regional distribution layers.

8. The map feature recognition method based on deep learning according to claim 7, characterized in that: The steps for obtaining S5 are as follows: S501: Based on the main response channel information in the main response feature list of the region, locate the pixel region corresponding to each main response channel in the original remote sensing image, call the correspondence between the region number and the channel type to extract the pixel positions contained in all main response channels in the image, classify and collect the coordinate numbers according to the channel, and generate a set of channel positioning coordinates. S502: Based on the pixel position corresponding to each channel in the channel positioning coordinate set, call the channel type identifier and map feature attribute type in the channel and feature mapping rule base, perform mapping on the matching relationship between channel number and feature type, write all mapping relationships into the channel correspondence table, and generate the channel feature correspondence table. S503: Based on the established channel number and map element category relationship in the channel element correspondence table, and combined with the connected contour of each channel pixel coordinate set in the image, the connected region is marked according to the connected region number and assigned the corresponding map element category label, and mapped to the image spatial location of the original remote sensing image to generate a map element recognition image.

9. A map feature recognition system based on deep learning, characterized in that, The system is used in the deep learning-based map feature recognition method according to any one of claims 1-8, the system comprising: The multi-scale edge perception module acquires multi-channel images covering the map feature area, constructs multi-scale layers according to resolution, extracts multi-level feature images, scans each layer of feature images row by row and column by column, extracts brightness coherent pixel segments, locates pixel fragments, acquires and encodes horizontal and vertical edge trajectories, and forms an edge perception channel sequence. The directional trajectory fusion module extracts adjacent edge trajectory segments from the edge perception channel sequence, compares the brightness change direction and trend, filters consistent edge trajectory pairs and merges them along the extension direction, connects adjacent edge trajectory segments into continuous line segments and records the start and end positions to obtain a continuous line segment representation map. The closed structure parsing module obtains the start and end positions of each continuous line segment in the continuous line segment representation diagram. Based on the relative distribution relationship of the start and end positions, it analyzes the spatial continuity relationship, traces the connection relationship between the beginning and end along the extension direction of the continuous line segments, determines the closed path formed by the connection of multiple line segments, verifies the pixel connectivity of the enclosed area, and obtains the contour image of the closed area. The region channel determination module, for each closed region in the closed region contour image, counts the brightness change frequency in the red, green and blue channels of each closed region, selects the main response channel according to the sorting result formed between the channels based on the brightness change frequency, records the corresponding order, and forms a list of region main response features. The feature type mapping module locates the pixel region corresponding to the main response channel in the original image based on the main response channel information in the region main response feature list, obtains the set of pixel coordinates within the pixel region and matches the channel feature mapping rules, and establishes the correspondence between the main response channel region and the preset map feature category by combining the region connectivity and enclosure relationship, thereby obtaining the map feature recognition image.

Images