Remote sensing identification method of target tree species based on air-ground collaboration

Abstract

The present invention provides a method for remote sensing identification of target tree species based on air-ground-ground collaboration. The method performs adjacent superpixel merging on the segmented spots after superpixel segmentation, performs feature factor-based model optimization on the current identification model of the target tree species based on the merged spot distribution and the field distribution data, performs target tree species identification based on the updated new identification model and obtains the correct identification spots; performs edge matching detection on the correct identification spots, and adjusts the current merging scale based on the edge matching detection results to obtain the optimal identification model, and then obtains the identification results of the target tree species based on the optimal identification model; the method of the present application effectively improves the accuracy of the spot boundary while ensuring the identification accuracy of the target tree species, so that the matching degree between the boundary of the segmented spot and the actual boundary of the target tree species is effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing monitoring technology, and in particular to a target tree species remote sensing identification method based on air-ground-space collaboration, and more specifically to a target tree species remote sensing identification model acquisition method, a target tree species remote sensing identification method based on air-ground-space collaboration, a terminal and a computer storage medium. Background Art

[0002] Existing tree species identification methods often use manual identification methods, which are not only time-consuming and labor-intensive, but also have low accuracy. In addition, they often require field surveys in deep mountainous areas, which is very dangerous. However, the use of satellite remote sensing technology for tree species identification is not only efficient and cost-effective, but also suitable for large-scale long-term monitoring.

[0003] When using satellite remote sensing images for tree species identification, it is often necessary to segment and classify the images. In the prior art, only the image segmentation method or the image classification method is often optimized, that is, the segmentation process or the classification process is simply optimized without considering the impact of the segmentation method, scale, etc. on the classification results, resulting in the recognition of the boundaries of the image patches being often low, which in turn affects the recognition accuracy of the tree species and reduces it. Especially in the identification of sky canopy tree species, the classification accuracy of the image patches is often reduced due to inaccurate image segmentation.

[0004] Therefore, how to simultaneously improve the segmentation accuracy and recognition accuracy of tree species has become a technical problem that needs to be solved in this field. Summary of the invention

[0005] In view of the shortcomings of the prior art described above, the purpose of the present invention is to provide a target tree species remote sensing identification model optimization method, terminal and computer storage medium, which are used to solve the problems of low accuracy of the boundaries of the identification spots in the existing target tree species identification model, which in turn leads to low accuracy of classification identification.

[0006] To achieve the above-mentioned and other related purposes, the present invention provides, in a first aspect, a method for acquiring a target tree species remote sensing identification model, comprising:

[0007] Construct a sample data set of the target tree species in the work area; the sample data set includes a sample data pair consisting of multidimensional image data and the corresponding field distribution data of the target tree species; perform superpixel segmentation on each of the multidimensional image data to obtain the corresponding initial segmentation patch; perform neighboring superpixel merging based on the current merging scale on each of the initial segmentation patches to obtain the merged patch distribution of the multidimensional image data corresponding to the current merging scale; based on the merged patch distribution corresponding to each of the multidimensional image data and the field distribution data, in combination with the feature factor library of the target tree species, perform feature factor-based model optimization on the current recognition model of the target tree species to update the recognition model; based on the updated new recognition model, perform target tree species recognition on each of the multidimensional image data to obtain the current correct recognition patch; perform edge matching detection on the current correct recognition patch, and adjust the merging scale based on the edge matching detection result to obtain the optimal recognition model and the corresponding optimal merging scale.

[0008] In one embodiment of the present invention, the neighboring superpixels merging based on the current merging scale, when executed, includes:

[0009] Based on the distribution of each of the initial segmentation spots, a regional adjacency graph of the current multidimensional image data is constructed; according to the current merging scale, a clustering method is used to perform neighboring superpixel merging on each of the initial segmentation spots to obtain the distribution of each of the merged spots.

[0010] In one embodiment of the present invention, the implementation method of performing characteristic factor-based model optimization on the current recognition model of the target tree species includes:

[0011] The current target characteristic factor contained in the current recognition model is obtained; based on the current target characteristic factor, a characteristic factor in the characteristic factor library that is different from the current target characteristic factor is used as a current candidate characteristic factor; according to the calculation method of the current candidate characteristic factor, corresponding calculations are performed on each current segmentation patch of the multidimensional image data to obtain a corresponding characteristic distribution map; based on the characteristic distribution map corresponding to each multidimensional image data and the corresponding field distribution data, a characteristic factor associated with target tree species identification is extracted from each current candidate characteristic factor using a characteristic screening method as a new target characteristic factor; based on the new target characteristic factor, the current target characteristic factor is updated.

[0012] In one embodiment of the present invention, the feature screening method includes the Boruta method.

[0013] In one embodiment of the present invention, the performing edge matching detection on the current correctly identified pattern includes:

[0014] Based on the field distribution data of the target tree species, the field distribution spots corresponding to each correctly identified spot are obtained; the spot boundaries of each correctly identified spot are extracted, and the spot boundaries of the corresponding field distribution spots are extracted; based on the boundary distribution of the two, the edge matching detection method is used to extract the edge matching degree between the correctly identified spot and the corresponding field distribution spot; and the edge matching degree is detected to see whether it is greater than a matching degree threshold.

[0015] In one embodiment of the present invention, the step of adjusting the merging scale based on the edge matching detection result includes:

[0016] When the proportion of the edge of the correctly identified spot exceeding the edge of the target spot is greater than the proportion threshold, the current merging scale is reduced; and when the proportion of the edge of the correctly identified spot not exceeding the edge of the target spot is greater than the proportion threshold, the current merging scale is increased.

[0017] In one embodiment of the present invention, the method for acquiring the target tree species remote sensing identification model further includes:

[0018] Construct each characteristic factor of the target tree species, and construct a characteristic factor library of the target tree species based on each characteristic factor; wherein the characteristic factors include habitat factors, texture factors and spectral factors.

[0019] To achieve the above-mentioned and other related purposes, the present invention provides, in a second aspect, a method for remote sensing identification of target tree species based on air-ground collaboration, comprising:

[0020] Obtain an optimal merging scale corresponding to a target tree species in a working area, and an optimal recognition model of the target tree species corresponding to the merging scale; perform superpixel segmentation on the multidimensional image data of the working area to obtain initial segmented patches after segmentation; based on the optimal merging scale, perform neighboring superpixel merging of the corresponding merging scale on each of the initial segmented patches to obtain a corresponding merged patch distribution; use the optimal recognition model to identify each patch in the merged patch distribution to obtain the recognition of the target tree species; wherein the optimal recognition model is obtained by using any of the above-described methods for obtaining a remote sensing recognition model of the target tree species.

[0021] To achieve the above-mentioned purpose and other related purposes, the present invention also provides a storage medium in the third aspect, on which a computer program is stored, which, when executed by a processor, implements a method for acquiring a remote sensing identification model of a target tree species as described above, or implements a method for remote sensing identification of a target tree species based on air-ground-space collaboration as described above.

[0022] In a fourth aspect, the present invention also provides a terminal, comprising a processor and a memory, wherein the memory is communicatively connected to the processor; the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory, so that the terminal executes any method for acquiring a remote sensing identification model of a target tree species as described above, or any method for remote sensing identification of a target tree species based on air-ground-space collaboration as described above.

[0023] As described above, the method for acquiring the target tree species remote sensing identification model, the target tree species remote sensing identification method based on air-ground-space collaboration, the terminal and the computer storage medium provided by the present invention perform segmentation on the multidimensional image by using the superpixel segmentation method, and perform adjacent superpixel clustering and merging based on the merging scale on the spots obtained by the segmentation to obtain initial segmentation spots corresponding to the current merging scale, and screen the characteristic factors based on the distribution of each initial segmentation spot, optimize the recognition model of the target tree species based on the screened target recognition factors, and detect the boundary matching degree of the correct recognition spots of the optimized recognition model, and detect the correct recognition spots and the target tree species. When the boundary of the corresponding actual map patch does not match, the current merging scale is adjusted, and the optimal merging scale matching the target tree species in the working area and the optimal recognition model corresponding to the optimal merging scale are obtained through the above iterative process. While ensuring the recognition accuracy of the target tree species, the accuracy of the patch boundary is effectively improved, so that the matching degree between the boundary of the segmented patch and the actual boundary of the target tree species is effectively improved, so that the target recognition factor corresponding to the target tree species can be obtained quickly, efficiently and accurately, and then the recognition model constructed based on the target recognition factor can accurately and efficiently obtain the recognition patch corresponding to the target tree species, which not only improves the accuracy of tree species recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 Shown is a flow chart of a method for acquiring the target tree species remote sensing identification model provided by the present application in one embodiment;

[0025] Figure 2 Shown is a flow chart of performing superpixel segmentation on the multi-dimensional image data in each of the sample data pairs in the present application in one embodiment;

[0026] Figure 3 Shown is a flow chart of step S300 described in this application in one embodiment;

[0027] Figure 4 Shown is a flow chart of using the Boruta method in the present application to identify all characteristic factors associated with target tree species from various characteristic factors in one embodiment;

[0028] Figure 5Shown is a schematic flow chart of a method for obtaining characteristic factors described in the present application in one embodiment;

[0029] Figure 6 Shown is a flow chart of the target tree species remote sensing identification method provided by the present application in one embodiment;

[0030] Figure 7 Shown is a schematic structural diagram of the electronic terminal provided by the present invention in one embodiment. DETAILED DESCRIPTION

[0031] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

[0032] It should be noted that the illustrations provided in the following embodiments are only for schematically illustrating the basic concept of the present invention.

[0033] In order to solve the technical problems existing in the prior art, the present invention first provides a method for obtaining a remote sensing identification model of a target tree species. In the process of superpixel segmentation of a remote sensing image, the segmented patches are subjected to superpixel merging based on a merging scale, and the identification feature factors of the target tree species are screened based on the merged patches to obtain a target identification factor having a high correlation with the target tree species identification patches. Based on the target identification factor obtained by screening, a pre-constructed identification model corresponding to the target tree species is trained to obtain a trained identification model, which not only effectively improves the accuracy of the identification boundary of the target tree species, but also effectively improves the model recognition efficiency and accuracy.

[0034] Among them, the target tree species are sky canopy tree species, that is, tree species with a high growth height (higher than surrounding trees) and a large tree width; for example, Huangshan pine, Masson pine or Liquidambar formosana.

[0035] See also Figure 1 , showing a flow chart of a method for acquiring a target tree species remote sensing identification model provided in an embodiment of the present application.

[0036] S100, constructing a sample data set of target tree species in the work area;

[0037] Wherein, the sample data set includes multiple groups of sample data pairs of target tree species;

[0038] A single set of sample data pairs includes multi-dimensional image data of target tree species and field distribution data of the target tree species corresponding to the image data in time and space;

[0039] The multi-dimensional image data contains the texture features and spectral features of the target tree species; and the field distribution data of the target tree species contains the real range of the target tree species.

[0040] Specifically, the image acquisition time corresponding to the multidimensional image data is adapted to the acquisition time of the field distribution data of the target tree species to ensure the temporal consistency of the two; and the spatial coordinates corresponding to the multidimensional image data are matched with the spatial coordinates of the field distribution data to ensure the spatial consistency of the two.

[0041] In an optional embodiment, a method for obtaining a single set of sample data pairs includes:

[0042] Acquire a remote sensing image that is located within the working area and covers the target tree species area; the remote sensing image contains multiple spectral bands; extract a band containing spectral characteristics and texture characteristics of the target tree species from each spectral band as a characteristic band; fuse each characteristic band to obtain multidimensional image data;

[0043] Acquire the image acquisition time of the remote sensing image; construct a corresponding target time range based on the image acquisition time; the time intervals between the start time and the end time of the target time range and the image acquisition time are both less than a time interval threshold; acquire the measured distribution data of the target tree species area within the target time range, align the measured distribution data with the remote sensing image, and use the aligned measured distribution data as the field distribution data of the target tree species corresponding to the remote sensing image in time and space.

[0044] The above steps are repeatedly performed on the working area to obtain multiple groups of sample data pairs in the working area.

[0045] S200, performing superpixel segmentation on the multidimensional image data in each of the sample data pairs to obtain an initial segmentation pattern corresponding to the multidimensional image data;

[0046] Specifically, according to a preset superpixel number factor, superpixel segmentation is performed on the multidimensional image data in each of the sample data pairs; wherein the superpixel number factor is a factor that affects the granularity and details of the segmentation result, and when the value of the superpixel number factor is larger, the segmented superpixels are smaller, and a single superpixel contains more details; conversely, when the value of the superpixel number factor is smaller, the segmented superpixels are larger, and a single superpixel contains fewer details.

[0047] In a specific embodiment, for a single multi-dimensional image data, the implementation method of performing superpixel segmentation on the multi-dimensional image data in each sample data pair is as follows: Figure 2 It includes the following sub-steps:

[0048] S201, obtaining the pixel value and the number of superpixels corresponding to each pixel point in the current multi-dimensional image data; based on the pixel value and the number of superpixels, obtaining the pixel mean corresponding to the superpixel, which is:

[0049] Among them, is the pixel mean corresponding to the superpixel; N is the total pixel value corresponding to each pixel point, that is, the accumulated value of the pixel values ​​corresponding to each pixel point; K is the number of superpixels.

[0050] S202, based on the pixel mean, using a cluster analysis method to determine each cluster center point in the current multidimensional image data;

[0051] Specifically, a number of initial seed points are randomly selected and a neighborhood radius is determined; the neighborhood radius can be set according to the extraction accuracy, that is, the higher the extraction accuracy, the smaller the neighborhood radius, and vice versa. Exemplarily, the neighborhood radius is 3.

[0052] According to the neighborhood radius, the neighborhood area corresponding to various sub-points is determined, and a new seed point is reselected in the neighborhood area; that is, for a single seed point, according to the preset neighborhood radius N, the neighborhood area corresponding to the current seed point is determined to be N*N; after determining the neighborhood area, the gradient values ​​corresponding to each pixel point in the neighborhood area are extracted; the maximum gradient value is extracted from the gradient values ​​corresponding to each pixel point, and the pixel point corresponding to the maximum gradient value is used as the new seed point to avoid the seed point falling on the contour boundary with a larger gradient, thereby affecting the accuracy of the subsequent analysis results (clustering effect); repeat this step until the position of the selected new seed point no longer changes, and the final seed point is used as the cluster center point.

[0053] S203, extracting the comprehensive distance between the center points of each cluster; based on the comprehensive distance, clustering each pixel point, and using the spots obtained after the clustering as the initial segmentation spots.

[0054] In this embodiment, the comprehensive distance includes color distance and spatial distance;

[0055] Specifically, the method for obtaining the color distance and the spatial distance includes:

[0056]

[0057] in, dc is the color distance; ds is the spatial distance;D’ is the comprehensive distance; Nc is the maximum color distance within the class; N is the maximum spatial distance within the class and satisfies Ns=S=sqrt(N / K) .

[0058] In some optional embodiments, Nc The value range is [1,40]; for example, Nc is 10.

[0059] S300, performing adjacent superpixel merging based on a merging scale on the initial segmented image spots to obtain a merging image spot distribution corresponding to the current merging scale for each of the multi-dimensional image data;

[0060] In an optional embodiment, step S300 is executed when a single multi-dimensional image data is processed. Figure 3 As shown, it includes the following sub-steps:

[0061] S301, constructing a regional adjacency graph of the current multi-dimensional image data based on the distribution of each of the initial segmentation patches;

[0062] Wherein, the regional adjacency graph includes nodes and connecting edges between two adjacent nodes;

[0063] The node is used to represent a superpixel, and the connecting edge represents an edge connecting two adjacent superpixels in space.

[0064] It should be noted that the weight of the edge can be determined based on color difference, spatial distance, or a combination of the two, which is not specifically limited here.

[0065] S302, according to the current merging scale, using a clustering method to perform neighboring superpixel merging on each of the initial segmentation spots to obtain a distribution of each of the merged spots.

[0066] The merged spot distribution includes each current segmentation spot, and the current segmentation spot is a temporary spot obtained after performing adjacent superpixel merging on each initial segmentation spot.

[0067] Specifically, each node in the region adjacency graph is traversed, and similar superpixels are merged according to a similarity metric and a current merging scale, so as to merge adjacent initial segmentation spots into a current segmentation spot.

[0068] The similarity metric is a metric used to determine whether to merge two adjacent superpixels; exemplarily, the similarity is a comprehensive metric constructed based on the color difference and the spatial distance.

[0069] In some other embodiments, after executing step S302, step S300 further includes:

[0070] S303, performing optimization processing on the merged current segmentation spots.

[0071] Specifically, the optimization process includes but is not limited to removing isolated areas, smoothing boundaries, etc., so as to improve the visual quality and analysis effect of the clustering area.

[0072] S400, based on the merged patch distribution corresponding to each of the multi-dimensional image data and the field distribution data, combined with the characteristic factor library of the target tree species, performing a model optimization based on characteristic factors on the current recognition model of the target tree species to update the recognition model;

[0073] The characteristic factor library of the target tree species is pre-constructed and used to identify the characteristic factors of the target tree species.

[0074] The target tree species recognition model is a classifier model, which is used to recognize the target tree species, that is, by calculating the probability that each image spot is determined to be the target tree species, thereby determining the recognition image spot corresponding to each image spot.

[0075] Exemplarily, the classifier model includes various machine learning models such as support vector machine (SVM) and random forest.

[0076] In this embodiment, the current recognition model is the recognition model obtained after executing the last model optimization, and the recognition model includes the target characteristic factors determined during the last model optimization; wherein the target characteristic factors are characteristic factors associated with the identification of the target tree species.

[0077] Specifically, obtaining the target characteristic factor contained in the current recognition model as the current target characteristic factor;

[0078] Based on the current target characteristic factor, a characteristic factor in the characteristic factor library that is different from the current target characteristic factor is used as a current candidate characteristic factor.

[0079] According to the calculation method of the current candidate feature factor, corresponding calculations are performed on each current segmentation spot of the multidimensional image data to obtain a corresponding feature distribution map, that is, the feature distribution map contains the features of each current segmentation spot corresponding to the current candidate feature factor.

[0080] Exemplarily, the current candidate characteristic factor is the NDVI factor, and for each current segmentation patch of the multidimensional image data, NDVI calculation is performed to obtain the NDVI value corresponding to the current segmentation patch; based on the NDVI values ​​corresponding to each current segmentation patch, an NDVI distribution map is constructed, and the NDVI distribution map is used as the current NDVI distribution map of the multidimensional image data.

[0081] Based on the feature distribution map corresponding to each of the multidimensional image data and the field distribution data of the target tree species corresponding to the multidimensional image data, a feature screening method is used to extract feature factors associated with target tree species identification from each of the current candidate feature factors as new target feature factors; based on the new target feature factors, the current target feature factors are updated to obtain a new classifier model.

[0082] Based on the new classifier model, target tree species identification is performed on the current distribution of each merged image patch; based on the identification image patches and the corresponding field distribution data, the new classifier model is obtained at the current identification accuracy; it is detected whether the current identification accuracy meets the accuracy threshold requirement, if so, the new classifier model is used as the updated new identification model, if not, the feature factor library is updated to re-acquire the current candidate feature factors based on the updated new feature factor library.

[0083] In an optional embodiment, when step S300 is executed, the Boruta method is used to identify all characteristic factors associated with the target tree species identification from each of the current candidate characteristic factors; wherein the core idea of ​​the Boruta algorithm is to evaluate the importance of each candidate characteristic factor for identifying the map by introducing pseudo-characteristic factors corresponding to the candidate characteristic factors, and determine whether to use the candidate characteristic factor as a new target characteristic factor based on the evaluation results.

[0084] Specifically, the implementation method of using the Boruta method to identify all characteristic factors associated with the target tree species from the current candidate characteristic factors is as follows: Figure 4 As shown, it includes the following sub-steps:

[0085] S401, based on the current candidate feature factors, obtaining pseudo feature factors corresponding to each current candidate feature factor; based on the current candidate feature factors and the corresponding pseudo feature factors, constructing a current candidate feature factor set;

[0086] Specifically, according to the current candidate characteristic factor, the characteristic factor value corresponding to the candidate characteristic factor of each current segmentation pattern spot is obtained; after the characteristic factor value corresponding to each current segmentation pattern spot is obtained, the factor value is randomly shuffled to obtain a pseudo characteristic factor corresponding to the current candidate characteristic factor;

[0087] The current candidate feature factor is combined with the pseudo feature factor to obtain a current candidate feature factor set.

[0088] S402, using a random forest model, performing a relevance evaluation on the current candidate feature factor set to obtain the importance of each current candidate feature factor corresponding to the identification pattern;

[0089] Specifically, each characteristic factor in the current candidate characteristic factor set (current candidate characteristic factor or pseudo characteristic factor) is used as the characteristic factor to be evaluated; for a single characteristic factor to be evaluated, the characteristic factor value corresponding to each current segmentation spot and the characteristic factor to be evaluated is calculated; the characteristic factor value corresponding to each current segmentation spot is used as sample data, and the field type corresponding to the current segmentation spot (i.e., whether it is the target tree species) is used as label data corresponding to the sample data;

[0090] A random forest classifier is trained based on the sample data and the corresponding label data.

[0091] The random forest can evaluate the contribution of each feature factor to be evaluated to the prediction by integrating multiple decision trees.

[0092] Random forest assigns each feature factor to be evaluated an importance score based on its contribution to the prediction. The importance score of each current candidate feature factor is compared with the maximum importance score of the pseudo feature factor. If the importance score of a current candidate feature factor is higher than the maximum importance score of the pseudo feature factor, the current candidate feature factor is judged to be a relevant feature factor, otherwise it is judged to be a non-relevant feature factor.

[0093] After the current iteration, the non-correlated feature factors that fail to exceed the maximum importance score of the pseudo feature factor are deleted, and the above process is repeated for the remaining feature factors to be evaluated until all the feature factors to be evaluated are evaluated as related feature factors or non-correlated feature factors.

[0094] The method provided in this embodiment uses the Boruta algorithm to perform relevance evaluation on each characteristic factor to be evaluated, so as to comprehensively evaluate the importance of each current candidate characteristic factor, and ensure that the current candidate characteristic factors finally retained all have actual contributions to the identification of the target tree species.

[0095] It should be noted that a key advantage of the Boruta algorithm is its ability to handle interactions between feature factors, that is, some feature factors are not important when evaluated alone, but if they can provide additional predictive power when combined with other feature factors, Boruta may also be able to identify the importance of these feature factors.

[0096] S500, based on the new recognition model, target tree species recognition is performed on each of the multidimensional image data to obtain a current correct recognition patch; edge matching detection is performed on the current correct recognition patch, and the merging scale is adjusted based on the edge matching detection result, so as to return to perform adjacent superpixel merging and subsequent steps based on the new merging scale to obtain an optimal recognition model and a corresponding optimal merging scale.

[0097] Among them, the correctly identified patch is the current segmented patch correctly determined by the recognition model as the target tree species, that is, the current segmented patch is recognized as the target tree species by the recognition model, and the true type of the patch is the patch of the target tree species.

[0098] Specifically, after obtaining the new recognition model, the current segmentation spots in the multidimensional image data are input into the new recognition model to obtain the recognition probability of each current segmentation spot corresponding to the target tree species; based on the recognition probability and in combination with a preset probability threshold, the current segmentation spots with a recognition probability greater than the probability threshold are extracted as the recognition spots of the target tree species;

[0099] Based on the field distribution data of the target tree species, the identification spots overlapping with the field distribution data are extracted from the identification spots as the current correct identification spots of the multi-dimensional image data.

[0100] After obtaining the multi-dimensional image data at each current correct identification spot, obtaining the field distribution spot corresponding to each correct identification spot based on the field distribution data of the target tree species;

[0101] After obtaining the current correct recognition spots and the field distribution spots corresponding to each of the correct recognition spots, edge matching detection is performed on the current correct recognition spots based on the edges of the field distribution spots; when the detection result of the edge matching detection does not meet the preset matching conditions, the current merging scale is adjusted based on the detection result, so as to return to perform the adjacent superpixel merging based on the adjusted merging scale, that is, return to execute step S300 based on the adjusted merging scale until the detection result of the edge matching detection meets the preset matching conditions; when the detection result of the edge matching detection meets the preset matching conditions, the current iterative process is exited, and the current recognition model is used as the optimal recognition model, and the current merging scale is used as the optimal merging scale.

[0102] In a specific implementation, the implementation method of performing edge matching detection on the current correctly identified spot based on the edge of the field distribution spot includes:

[0103] Extracting the spot boundary of the correctly identified spot and extracting the spot boundary of the corresponding field distribution spot; based on the spot boundary distribution of the correctly identified spot and the corresponding field distribution spot, using the edge matching detection method, extracting the edge matching degree between the correctly identified spot and the corresponding field distribution spot; detecting whether the edge matching degree is greater than a matching degree threshold;

[0104] If so, it is determined that the edges of the two are matched, and the current merging scale is determined to be the target merging scale, that is, the optimal merging scale; and the current recognition model is used as the target recognition model, that is, the optimal recognition model;

[0105] If not, the current merging scale is adjusted according to the edge distribution states of the two to optimize the superpixel segmentation process.

[0106] In a specific embodiment, adjusting the merging scale according to the edge distribution states of the two includes:

[0107] When the proportion of the edges of the correctly identified spots that exceed the edges of the corresponding actual distribution spots is greater than a threshold, that is, the edges of most of the correctly identified spots exceed the edges of the corresponding actual distribution spots, indicating that the recognition range of a single identified spot is too large, the current merging scale is reduced accordingly; when the proportion of the edges of the correctly identified spots that are located within the edges of the corresponding actual distribution spots is greater than a threshold, that is, the edges of most of the correctly identified spots do not exceed the edges of the corresponding actual distribution spots, indicating that the recognition range of a single identified spot is too small, the current merging scale is increased accordingly; based on the adjusted current merging scale, re-execute the above steps S300 to S500 until exit.

[0108] In an optional embodiment, the edge matching detection method includes a spatial superposition analysis method, a change vector analysis method, and the like.

[0109] The method for acquiring the target tree species remote sensing identification model provided in the present embodiment performs segmentation on the multidimensional image by using the superpixel segmentation method, and performs adjacent superpixel clustering and merging based on the merging scale on the spots obtained by the segmentation to obtain the current segmented spots corresponding to the current merging scale, and based on the distribution of each current segmented spot, constructs a feature distribution map corresponding to the feature factor to achieve screening of the feature factor, so as to optimize the recognition model of the target tree species based on the screened target recognition factor, and detect the boundary matching degree of the correct recognition spot of the optimized recognition model. When it is detected that the correct recognition spot does not match the boundary of the corresponding actual map spot, the current merging scale is adjusted, and the optimal merging scale matching the target tree species in the working area and the recognition model corresponding to the optimal merging scale are obtained through the above-mentioned iterative process, so as to ensure the recognition accuracy of the target tree species while effectively improving the accuracy of the spot boundary, so that the matching degree between the boundary of the segmented spot and the actual boundary of the target tree species is effectively improved.

[0110] In order to improve the model recognition accuracy of the target tree species, in an optional embodiment, the method for acquiring the target tree species remote sensing recognition model, before executing step S100, further includes:

[0111] S600, constructing characteristic factors of the target tree species, and constructing a characteristic factor library of the target tree species based on the characteristic factors.

[0112] Specifically, the characteristic factor is obtained as follows: Figure 5As shown, including:

[0113] S601, obtaining image data of the working area, surface feature spectral reflectance data and sample site survey data;

[0114] Wherein, the image data at least includes high-resolution remote sensing image data of the target tree species to be identified;

[0115] The sample plot survey data includes environmental data such as topographic data, meteorological data, soil data, etc. in the local area where the target tree species sample points are located;

[0116] The ground object spectral reflectance data is spectral reflectance data collected by using a ground object spectrometer.

[0117] Specifically, after acquiring the remote sensing image data of each scene in the working area, pre-processing such as cropping, radiation calibration and atmospheric correction is performed on each remote sensing image data to obtain image data corresponding to the target tree species; according to the acquisition time of the image data, the ground spectral data corresponding to the target tree species is collected within the same or similar acquisition time period to obtain the ground object spectral reflectance data corresponding to the target tree species;

[0118] And, according to the acquisition time of the image data, the sample plot survey data corresponding to the local area where the target tree species is located is collected within the same or similar acquisition time period to obtain the sample plot survey data corresponding to the target tree species.

[0119] After obtaining the image data, ground object spectral data and sample plot survey data corresponding to the target tree species, the ground control point data in the working area are combined to perform spatial registration processing between the ground spectral data and the remote sensing image data, and perform spatiotemporal registration processing between the sample plot survey data and the remote sensing image data to achieve spatiotemporal matching among the three.

[0120] It should be noted that when performing long-term monitoring of target tree species, the image data covers the area where the target tree species is located, and covers multispectral satellite images and hyperspectral satellite images of different seasons, so as to improve the coverage rate of remote sensing image data in the working area and increase the characteristic factors of the target tree species.

[0121] S602, constructing characteristic factors of target tree species based on the image data, the ground object spectral reflectance data and the sample plot survey data;

[0122] In this application, the characteristic factors include habitat factors, texture factors and spectral factors;

[0123] The habitat factor is a characteristic factor that characterizes the surrounding environment conditions such as topography, weather and soil in the local area where the target tree species is located;

[0124] The texture factor is a characteristic factor representing the image texture information corresponding to the target tree species;

[0125] The spectral factor is a characteristic factor that characterizes the changes in the spectral curve corresponding to the target tree species; wherein the spectral factor includes a phenological spectral factor and a band spectral factor; the phenological spectral factor is a characteristic factor used to characterize the changes in the spectral reflectance of the target tree species in different phenological periods, including but not limited to the peaks or troughs of the near-infrared band; the band spectral factor is a characteristic factor used to characterize the changes in the spectral reflectance of the target tree species corresponding to different bands, including but not limited to NDVI, DVI, etc.

[0126] Specifically, for the texture factor, image data containing the target tree species is acquired, and texture information corresponding to the target tree species is extracted from the image data, so as to construct the texture factor based on the texture information.

[0127] Using a ground object spectrometer to collect spectral reflectance data of target tree species in different phenological periods; extracting spectral reflectance information of the same band in different phenological periods, and constructing phenological spectrum factors based on the spectral reflectance information corresponding to the same band in different phenological periods; and extracting spectral reflectance information corresponding to different bands in the same phenological period based on the spectral reflectance data corresponding to the same phenological period, and constructing band spectrum factors based on the spectral reflectance information corresponding to different bands in the same phenological period;

[0128] And, based on the terrain data, constructing terrain characteristic factors for characterizing the topography and landforms in the local area where the target tree species are located, including but not limited to slope, slope aspect and concavity and convexity;

[0129] Based on the meteorological data, construct meteorological characteristic factors for characterizing the temperature and humidity degree in the local area where the target tree species is located, including but not limited to temperature and humidity;

[0130] Based on the soil data, soil characteristic factors are constructed to characterize the soil status in the local area where the target tree species is located, including but not limited to soil layer thickness, soil quality, surface soil temperature, etc.

[0131] In a specific embodiment, the characteristic factors constructed are shown in Table 1 below.

[0132] Table 1 Characteristic factors of target tree species

[0133]

[0134] The method for acquiring the remote sensing identification model of the target tree species provided in the embodiment of the present application constructs identification characteristic factors of the target tree species based on image data of the working area, spectral reflectance data of the ground objects and sample plot survey data; and uses machine learning methods to perform automatic screening on the identification characteristic factors, so that the target identification factors corresponding to the target tree species can be obtained quickly, efficiently and accurately. Then, the identification model constructed based on the target identification factors can accurately and efficiently obtain the identification patches corresponding to the target tree species, which not only improves the accuracy of the tree species identification patches, but also can be applied to the identification process of a variety of different target tree species. It has high adaptability and scalability, and provides an effective monitoring means for long-term monitoring of tree species in a large area.

[0135] In order to solve the technical problems existing in the prior art, the embodiment of the present invention further provides a target tree species remote sensing identification method based on air-ground collaboration (hereinafter referred to as "target tree species remote sensing identification method"), which is used to identify and obtain target tree species spots in a working area;

[0136] like Figure 6 As shown, the target tree species remote sensing identification method comprises the following steps:

[0137] S10, obtaining an optimal merging scale corresponding to the target tree species in the work area, and an optimal recognition model of the target tree species corresponding to the merging scale;

[0138] Among them, the optimal recognition model is obtained by adopting the method for acquiring the target tree species remote sensing recognition model provided in the above embodiment, and the optimal merging scale is obtained by executing a model training process on a preset target tree species recognition model based on the method for acquiring the target tree species remote sensing recognition model. The specific acquisition method is the same as the acquisition method in the above embodiment, which will not be repeated here.

[0139] S20, performing superpixel segmentation on the multi-dimensional image data of the working area to obtain an initial segmentation image patch after segmentation;

[0140] The multi-dimensional image data are remote sensing image data containing texture features and spectral features of target tree species.

[0141] In this embodiment, the implementation method of performing superpixel segmentation on the multi-dimensional image data of the working area is the same as the implementation method in the above embodiment, and will not be repeated here.

[0142] S30, based on the optimal merging scale, performing adjacent superpixel merging of the corresponding merging scale on each of the initial segmentation spots to obtain a corresponding merging spot;

[0143] Specifically, the implementation method of performing adjacent superpixel merging of the corresponding merging scale on each of the initial segmentation spots based on the optimal merging scale is the same as the implementation method in the above embodiment, which will not be repeated here.

[0144] S40, using the optimal recognition model to identify each of the merged image spots to obtain a recognition result of the target tree species.

[0145] The target tree species remote sensing identification method provided in the embodiment of the present application can not only accurately and efficiently obtain the identification patches corresponding to the target tree species, but also improve the accuracy of the identification boundaries corresponding to the target tree species, thereby greatly improving the recognition and extraction efficiency of the target tree species and providing convenience for subsequent analysis.

[0146] Based on the same technical concept, the method for acquiring the target tree species remote sensing identification model or the target tree species remote sensing identification method provided in the above embodiments of the present invention can be implemented on the terminal side or the server side.

[0147] See also Figure 7 , is an optional hardware structure diagram of an electronic terminal 50 provided in an embodiment of the present invention. The electronic terminal 50 may be a live broadcast machine, a camera, a mobile phone, a computer device, a tablet device, a personal digital processing device, a factory background processing device, etc. that integrates photo / video functions. The electronic terminal 50 includes: at least one processor 51, a memory 52, at least one network interface 53 and a user interface 54. The various components in the device are coupled together through a bus system 55. It can be understood that the bus system 55 is used to realize the connection and communication between these components. In addition to the data bus, the bus system 55 also includes a power bus, a control bus and a status signal bus.

[0148] The user interface 54 may include a display, a keyboard, a mouse, a trackball, a click gun, keys, buttons, a touch pad or a touch screen.

[0149] It is understood that the memory 52 can be a volatile memory or a non-volatile memory, and can also include both volatile and non-volatile memories. Among them, the non-volatile memory can be a read-only memory (ROM), a programmable read-only memory (PROM), which is used as an external cache. By way of example but not limitation, many forms of RAM are available, such as static random access memory (SRAM), synchronous static random access memory (SSRAM). The memory described in the embodiments of the present invention is intended to include but is not limited to these and any other suitable categories of memory.

[0150] The memory 52 in the embodiment of the present invention is used to store various categories of data to support the operation of the electronic terminal 50. Examples of these data include: any executable program used to operate on the electronic terminal 50, such as an operating system 521 and an application 522; the operating system 521 includes various system programs, such as a framework layer, a core library layer, a driver layer, etc., for implementing various basic services and processing hardware-based tasks. The application 522 can include various applications, such as a media player (MediaPlayer), a browser (Browser), etc., for implementing various application services. The method for acquiring the target tree species remote sensing identification model or the target tree species remote sensing identification method that implements the embodiment of the present invention can be included in the application 522.

[0151] The method disclosed in the above embodiment of the present invention can be applied to the processor 51, or implemented by the processor 51. The processor 51 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the hardware integrated logic circuit in the processor 51 or the instruction in the form of software. The above processor 51 may be a general processor, a digital signal processor (DSP, Digital Signal Processor), or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The processor 51 can implement or execute the methods, steps and logic block diagrams disclosed in the embodiment of the present invention. The general processor 51 may be a microprocessor or any conventional processor, etc. In combination with the steps of the accessory optimization method provided in the embodiment of the present invention, it can be directly embodied as a hardware decoding processor to execute, or it can be executed by a combination of hardware and software modules in the decoding processor. The software module can be located in a storage medium, which is located in a memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware.

[0152] In an exemplary embodiment, the electronic terminal 50 may be implemented by one or more application specific integrated circuits (ASIC), DSP, programmable logic device (PLD), complex programmable logic device (CPLD) to execute the aforementioned method.

[0153] An embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon. When the program is called by a processor, the method for acquiring the target tree species remote sensing identification model or the target tree species remote sensing identification method provided by the present invention is implemented.

[0154] Among them, the computer-readable storage medium can be a tangible device that can hold and store instructions used by the instruction execution device. The computer-readable storage medium can be, for example, (but not limited to) an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the above. More specific examples of computer-readable storage media (a non-exhaustive list) include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, and a mechanical encoding device.

[0155] The computer-readable program described herein can be downloaded from a computer-readable storage medium to each computing / processing device, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. A network adapter card or a network interface in each computing / processing device receives computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in a computer-readable storage medium in each computing / processing device.

[0156] It should be noted that, in the various embodiments of the present application, the serial numbers of the above steps do not indicate the order of execution. The execution order of the steps should be determined by their functions and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0157] The above embodiments are merely illustrative of the principles and effects of the present invention, and are not intended to limit the present invention. Anyone familiar with the art may modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by a person of ordinary skill in the art without departing from the spirit and technical concept disclosed by the present invention shall still be covered by the claims of the present invention.

Claims

1. A method for acquiring a target tree species remote sensing identification model, characterized in that: include: Constructing a sample data set of target tree species in the work area; the sample data set includes a sample data pair consisting of multidimensional image data and corresponding field distribution data of the target tree species; Performing superpixel segmentation on each of the multi-dimensional image data to obtain a corresponding initial segmentation pattern; Performing neighboring superpixel merging based on the current merging scale on each of the initial segmented spots to obtain a merged spot distribution corresponding to the current merging scale of the multi-dimensional image data; Based on the merged patch distribution corresponding to each of the multidimensional image data and the field distribution data, combined with the characteristic factor library of the target tree species, a model optimization based on characteristic factors is performed on the current recognition model of the target tree species to update the recognition model, including: Acquire the current target characteristic factor contained in the current recognition model; based on the current target characteristic factor, use the characteristic factor in the characteristic factor library that is different from the current target characteristic factor as the current candidate characteristic factor; perform corresponding calculations on each current segmentation pattern of the multidimensional image data according to the calculation method of the current candidate characteristic factor to obtain a corresponding characteristic distribution map; based on the characteristic distribution map corresponding to each multidimensional image data and the corresponding field distribution data, use the characteristic screening method to extract the characteristic factor associated with the target tree species identification from each current candidate characteristic factor as the new target characteristic factor; based on the new target characteristic factor, update the current target characteristic factor; Based on the updated new recognition model, target tree species recognition is performed on each of the multidimensional image data to obtain the current correct recognition patch; edge matching detection is performed on the current correct recognition patch, and the merging scale is adjusted based on the edge matching detection result to obtain the optimal recognition model and the corresponding optimal merging scale.

2. The method for acquiring the target tree species remote sensing identification model according to claim 1, characterized in that: The merging of neighboring superpixels based on the current merging scale, when executed, includes: Based on the distribution of each of the initial segmentation spots, construct a regional adjacency graph of the current multi-dimensional image data; According to the current merging scale, a clustering method is used to perform neighboring superpixel merging on each of the initial segmentation spots to obtain the distribution of each merged spot.

3. The method for acquiring the target tree species remote sensing identification model according to claim 1, characterized in that: The feature screening method includes the Boruta method.

4. The method for acquiring the target tree species remote sensing identification model according to claim 1, characterized in that: The performing edge matching detection on the current correctly identified pattern includes: Based on the field distribution data of the target tree species, obtaining the field distribution spots corresponding to each correctly identified spot; Extracting the spot boundary of each correctly identified spot and extracting the spot boundary of the corresponding field distribution spot; based on the boundary distribution of the two, using the edge matching detection method, extracting the edge matching degree between the correctly identified spot and the corresponding field distribution spot; Check whether the edge matching degree is greater than the matching degree threshold.

5. The method for acquiring the target tree species remote sensing identification model according to claim 1, characterized in that: The adjusting the merging scale based on the edge matching detection result includes: When the proportion of the edge of the correctly identified spot exceeding the edge of the target spot is greater than the proportion threshold, the current merging scale is reduced; and when the proportion of the edge of the correctly identified spot not exceeding the edge of the target spot is greater than the proportion threshold, the current merging scale is increased.

6. The method for acquiring a target tree species remote sensing identification model according to claim 1, characterized in that: Also includes: Construct each characteristic factor of the target tree species, and construct a characteristic factor library of the target tree species based on each characteristic factor; wherein the characteristic factors include habitat factors, texture factors and spectral factors.

7. A method for remote sensing identification of target tree species based on air-ground collaboration, characterized in that: include: Obtaining the optimal merging scale corresponding to the target tree species in the work area and the optimal recognition model of the target tree species corresponding to the merging scale; Perform superpixel segmentation on the multi-dimensional image data of the working area to obtain initial segmentation patches after segmentation; Based on the optimal merging scale, performing adjacent superpixel merging of the corresponding merging scale on each of the initial segmented spots to obtain a corresponding merging spot distribution; The optimal recognition model is used to identify each patch in the combined patch distribution to obtain the identification of the target tree species; wherein the optimal recognition model is obtained by using the method for obtaining the remote sensing recognition model of the target tree species as described in any one of claims 1 to 6.

8. A storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method for acquiring a target tree species remote sensing identification model as described in any one of claims 1 to 6 is implemented, or when the program is executed by a processor, the method for remote sensing identification of target tree species as described in claim 7 is implemented.

9. A terminal, characterized in that: It includes a processor and a memory, and the memory is communicatively connected to the processor; the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory, so that the terminal executes the method for acquiring the target tree species remote sensing identification model as described in any one of claims 1 to 6 or the target tree species remote sensing identification method as described in claim 7.

Images