Remote sensing tree species identification method and device

Abstract

The invention discloses a remote sensing tree species identification method and device, and relates to the field of artificial intelligence, and the method comprises the steps: obtaining and determining a canopy height model based on an image shot by an unmanned plane in a to-be-detected region and a digital surface model; carrying out segmentation processing on the canopy height model; performing feature extraction on the segmentation result to obtain image features and structural features; inputting the image features and the structural features into a hybrid expert model to obtain a tree identification result; analyzing the image features and the structural features through a gating network to obtain a weight combination; and identifying the image features and the structural features through an expert network according to the weight combination to obtain a tree identification result. Generating a canopy height model by fusing the image and the digital surface model, and accurately extracting a single tree crown boundary; secondly, a hybrid expert model is introduced, expert networks focusing on different tree species types are dynamically scheduled through a gating network, and refined classification is achieved; through the step-by-step progressive design, the tree species identification precision is obviously improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a remote sensing method and apparatus for tree species identification. Background Technology

[0002] High-precision tree species identification plays a crucial role in forest resource management. With the development of UAV remote sensing technology and artificial intelligence, automated tree species identification based on high-resolution imagery has become a research focus. By combining forest imagery data acquired by UAVs with image processing and machine learning algorithms, tree species identification can be achieved, providing support for forest management and conservation.

[0003] Currently, tree species identification methods based on traditional image processing and machine learning mainly use canopy height models generated by digital surface models to segment individual tree canopies. These methods often use algorithms such as watershed and region growing for canopy segmentation, then manually design and extract spectral, texture, and structural features, and finally use traditional classifiers to classify the tree species. However, these canopy height model-based segmentation methods perform poorly when dealing with complex situations such as understory vegetation interference, irregular or severely overlapping canopy shapes, and are prone to oversegmentation or undersegmentation. Because inaccurate segmentation results introduce background noise or interference information from nearby ground features into the samples, they seriously affect the feature learning of subsequent classification models, thus limiting the overall accuracy of tree species identification. Summary of the Invention

[0004] This application provides a remote sensing tree species identification method and apparatus, with the aim of improving the accuracy of tree species identification.

[0005] To achieve the above objectives, this application provides the following technical solution:

[0006] The first aspect of this application provides a remote sensing tree species identification method, including:

[0007] Acquire images and digital surface models of the area to be tested taken by drones;

[0008] Determine the canopy height model based on the image and the digital surface model;

[0009] The canopy height model is segmented to obtain segmentation results; the segmentation results include the canopy mask or canopy polygon for each tree;

[0010] Feature extraction is performed on the segmentation results to obtain image features and structural features;

[0011] The image features and structural features are input into a hybrid expert model to obtain tree recognition results; wherein, the hybrid expert model includes a gating network and an expert network; the gating network analyzes the image features and structural features to obtain a weight combination; the expert network identifies the image features and structural features according to the weight combination to obtain tree recognition results.

[0012] Optionally, determining the canopy height model based on the image and the digital surface model includes:

[0013] Align the digital surface model according to the image to obtain the aligned digital surface model;

[0014] The difference between the aligned digital surface model and the preset digital terrain model is calculated to obtain the canopy height model.

[0015] Optionally, the segmentation process of the canopy height model to obtain the segmentation result includes:

[0016] The canopy height model is filtered using a filtering algorithm to obtain the canopy apex of each tree;

[0017] The crown vertices of each tree and the image are input into the segmentation model to obtain the segmentation results.

[0018] Optionally, the expert network includes at least a plurality of initial expert networks;

[0019] The step of identifying the tree by combining the weights based on the segmentation results using the expert network includes:

[0020] Obtain the weights corresponding to each of the initial expert networks from the weight combinations;

[0021] Multiple initial tree identification results are obtained by identifying the image features and structural features through each initial expert network;

[0022] Based on the weights corresponding to each initial expert network, the multiple initial tree identification results are weighted and calculated to obtain the tree identification result.

[0023] Optionally, the training process of the hybrid expert model includes:

[0024] Obtain sample tree species information;

[0025] Based on the segmentation results, spatial matching is performed on the sample tree species information to obtain sample tree species labels;

[0026] Extract the sample image features and sample structure features from the sample tree species labels;

[0027] Based on the sample image features and the sample structural features, sample data is constructed;

[0028] The sample data is input into a hybrid expert model to obtain the sample tree identification results;

[0029] Calculate the loss function between the sample tree identification results and the actual tree identification results corresponding to the sample data;

[0030] If the loss function does not converge, the model parameters of the hybrid expert model are adjusted, and the process returns to the step of inputting the sample data into the hybrid expert model to obtain the sample tree recognition result.

[0031] If the loss function converges, then the training of the hybrid expert model is considered complete.

[0032] A second aspect of this application provides a remote sensing tree species identification device, comprising:

[0033] The acquisition unit is used to acquire images and digital surface models of the area to be tested taken by a drone.

[0034] A determining unit is configured to determine a canopy height model based on the image and the digital surface model;

[0035] A segmentation unit is used to segment the canopy height model to obtain a segmentation result; the segmentation result includes the canopy mask or canopy polygon for each tree;

[0036] An extraction unit is used to extract features from the segmentation results to obtain image features and structural features;

[0037] The recognition unit is used to input the image features and the structural features into a hybrid expert model to obtain tree recognition results; wherein, the hybrid expert model includes a gating network and an expert network; the gating network analyzes the image features and the structural features to obtain a weight combination; the expert network recognizes the image features and the structural features according to the weight combination to obtain tree recognition results.

[0038] Optionally, the determining unit is specifically used for:

[0039] Align the digital surface model according to the image to obtain the aligned digital surface model;

[0040] The difference between the aligned digital surface model and the preset digital terrain model is calculated to obtain the canopy height model.

[0041] Optionally, the segmentation unit is specifically used for:

[0042] The canopy height model is filtered using a filtering algorithm to obtain the canopy apex of each tree;

[0043] The crown vertices of each tree and the image are input into the segmentation model to obtain the segmentation results.

[0044] Optionally, the expert network includes at least a plurality of initial expert networks;

[0045] The identification unit includes:

[0046] A sub-unit is used to obtain the weights corresponding to each of the initial expert networks from the weight combination;

[0047] The identification subunit is used to identify the image features and structural features through each of the initial expert networks to obtain multiple initial tree identification results;

[0048] The calculation subunit is used to perform weighted calculation on multiple initial tree identification results according to the weights corresponding to each initial expert network, so as to obtain the tree identification result.

[0049] Optional, also includes:

[0050] The information acquisition unit is used to acquire information about the sample tree species.

[0051] A matching unit is used to perform spatial matching on the sample tree species information based on the segmentation result to obtain sample tree species labels;

[0052] The feature extraction unit is used to extract sample image features and sample structural features from the sample tree species labels;

[0053] A construction unit is used to construct sample data based on the sample image features and the sample structural features;

[0054] The input unit is used to input the sample data into the hybrid expert model to obtain the sample tree recognition result;

[0055] The loss function unit is used to calculate the loss function between the sample tree identification result and the actual tree identification result corresponding to the sample data;

[0056] The adjustment unit is used to adjust the model parameters of the hybrid expert model if the loss function does not converge, and then return to the step of inputting the sample data into the hybrid expert model to obtain the sample tree recognition result.

[0057] The training determination unit is used to determine that the training of the hybrid expert model is complete if the loss function converges.

[0058] The technical solution provided in this application acquires images and digital surface models of the area to be tested taken by drones; determines a canopy height model based on the images and digital surface models; segments the canopy height model; extracts features from the segmentation results to obtain image features and structural features; inputs the image features and structural features into a hybrid expert model to obtain tree identification results; analyzes the image features and structural features through a gating network to obtain a weight combination; and identifies the image features and structural features based on the weight combination through an expert network to obtain tree identification results. By fusing images and digital surface models to generate a canopy height model and accurately extracting the canopy boundaries of individual trees, and by introducing a hybrid expert model and dynamically scheduling expert networks focused on different tree species through a gating network, refined classification is achieved. This progressive design significantly improves the accuracy of tree species identification. Attached Figure Description

[0059] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0060] Figure 1 A flowchart illustrating a remote sensing tree species identification method provided in this application embodiment;

[0061] Figure 2 A schematic diagram illustrating a remote sensing tree species identification method provided in an embodiment of this application;

[0062] Figure 3 This is a schematic diagram of the architecture of a remote sensing tree species identification device provided in an embodiment of this application. Detailed Implementation

[0063] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0064] In this application, the terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0065] like Figure 1 The flowchart shown is a remote sensing tree species identification method provided in an embodiment of this application, including the following steps:

[0066] S101: Acquire images and digital surface models of the area to be tested taken by a drone.

[0067] The images captured by the drone are high-resolution images. A Digital Surface Model (DSM) is an elevation model that represents the top elevation of all objects on the Earth's surface.

[0068] S102: Determine the canopy height model based on images and digital surface models.

[0069] The Canopy Height Model (CHM) is a two-dimensional data model where the value of each pixel represents the height of the tree canopy from the ground at that location.

[0070] Optionally, in another embodiment of this application, the specific implementation of step S102 includes processes A1 to A2.

[0071] A1: Align the digital surface model according to the image to obtain the aligned digital surface model.

[0072] It is understandable that aligning the digital surface model with the image involves georeferencing and aligning the image and the digital surface model with coordinates to ensure that each pixel has a consistent geographical location in both the image and the DSM, thus obtaining the aligned digital surface model.

[0073] A2: Calculate the difference between the aligned digital surface model and the preset digital terrain model to obtain the canopy height model.

[0074] The preset digital terrain model is a three-dimensional model used to represent changes in the Earth's surface elevation. It records the height values ​​of various points on the terrain surface in a digital way, thus forming a digital representation of the ground surface.

[0075] S103: Segment the canopy height model to obtain the segmentation results.

[0076] The segmentation results include the canopy mask or canopy polygon for each tree.

[0077] Specifically, a canopy mask is a type of raster data, and a canopy polygon is a precise closed contour curve drawn along the boundary of the canopy mask.

[0078] Optionally, in another embodiment of this application, the specific implementation of step S103 includes processes B1 to B2.

[0079] B1: The canopy height model is filtered using a filtering algorithm to obtain the crown apex of each tree.

[0080] Optionally, the filtering algorithm includes, but is not limited to, the local maximum filtering algorithm.

[0081] It is understandable that filtering algorithms are used to process the canopy height model, which means using local maximum filtering algorithms to find the crown vertices of each tree in the canopy height model, i.e., the coordinates of the crown vertices.

[0082] B2: Input the crown vertex and image of each tree into the segmentation model to obtain the segmentation result.

[0083] Optionally, the segmentation model may include, but is not limited to, the SAM model.

[0084] Understandably, by inputting the crown vertex and image of each tree into the segmentation model, the powerful zero-shot segmentation capability of the segmentation model can generate pixel-level, high-precision segmentation results for each individual tree on the image.

[0085] It should be noted that the segmentation model is pre-trained based on the canopy vertices of the sample trees and the sample images as input, and the segmentation results of manually annotated sample images as the training target.

[0086] As can be seen, the SAM model can effectively segment complex tree canopies with blurred edges and large areas of adhesion, and the generated boundaries closely match the actual morphology. This provides extremely "clean" input data for subsequent classification tasks, eliminating feature contamination at the source.

[0087] S104: Extract features from the segmentation results to obtain image features and structural features.

[0088] Among them, image features are image slices cropped from the tree canopy mask, which are two-dimensional features. Structural features refer to various three-dimensional quantitative indicators corresponding to the tree canopy, such as maximum tree height, average tree height, canopy area, canopy volume, surface area, and elevation variation coefficient, which are three-dimensional features.

[0089] S105: Input the image features and structural features into the hybrid expert model to obtain the tree identification results.

[0090] The hybrid expert model includes a gating network and an expert network. The gating network analyzes image features and structural features to obtain a weight combination. The expert network identifies image features and structural features based on the weight combination to obtain tree identification results.

[0091] Specifically, a gating network is a small neural network. Its function is to "make decisions" or "route".

[0092] Optionally, the hybrid expert model can be a MoE network model.

[0093] Specifically, the tree identification results include the tree species category and confidence level. A tree species distribution map of the test area can be generated based on the tree identification results.

[0094] Optionally, in another embodiment of this application, the expert network includes at least a plurality of initial expert networks; the specific implementation of obtaining tree recognition results by recognizing image features and structural features through the expert networks according to weight combinations includes processes C1 to C3.

[0095] C1: Obtain the weights corresponding to each initial expert network from the weight combination.

[0096] Each initial expert network is a lightweight convolutional neural network. Each initial expert network is designed to be good at identifying one or more specific tree species (e.g., initial expert network 1 is responsible for Pinaceae, initial expert network 2 for Cupressaceae, initial expert network 3 for Populus, etc.).

[0097] C2: Multiple initial tree identification results are obtained by identifying image features and structural features through each initial expert network.

[0098] Each initial expert network is good at identifying different tree species, so the initial tree identification results may be the same or different.

[0099] Understandably, by identifying image features and structural features through each initial expert network, that is, by extracting and classifying image features and structural features through each initial expert network, multiple initial tree identification results are obtained.

[0100] C3: Based on the weights corresponding to each initial expert network, the multiple initial tree identification results are weighted and calculated to obtain the tree identification result.

[0101] Understandably, during the weighted calculation process, each expert network is assigned a dynamic weight. This weight depends on its expertise in identifying the current target tree species: experts proficient in this type of tree are given higher weights, while those less proficient are given lower weights. The final identification result is derived from the weighted sum of each initial identification result and its corresponding weight.

[0102] It can be seen that the hybrid expert model effectively improves recognition performance through a "specialization of expertise" strategy. The gating network in the model acts like an experienced triage desk, routing input data to the most specialized expert network (such as an expert network for coniferous trees or broad-leaved trees) based on features such as crown shape. This mechanism ensures that the model can learn highly targeted, fine-grained features, thereby achieving a significant improvement in overall recognition accuracy in areas with high tree species diversity.

[0103] Optionally, in another embodiment of this application, the specific implementation of training the hybrid expert model includes processes D1 to D8.

[0104] D1: Obtain sample tree species information.

[0105] The sample tree species information refers to the tree species information obtained from field surveys. This information must include at least the location coordinates of the sample trees.

[0106] D2: Based on the segmentation results, spatial matching is performed on the sample tree species information to obtain the sample tree species labels.

[0107] Understandably, spatial matching of sample tree species information is performed based on the canopy polygons in the segmentation results, assigning the segmented individual tree canopies with the true labels from the ground survey. In other words, using spatial connectivity techniques, a "point-within-a-polygon" overlay analysis is performed on the GPS coordinates of the ground survey sample trees and the canopy polygon layer. If a sample tree point falls inside a canopy polygon, the true tree species attribute of the sample tree is automatically assigned to that polygon based on this spatial relationship, thus completing the label construction for all samples.

[0108] D3: Extract sample image features and sample structure features from the sample tree species labels.

[0109] This involves extracting sample image features and sample structure features from the sample tree species labels, which is essentially extracting multimodal features.

[0110] Optionally, image features can be extracted from the original image corresponding to the sample tree species information, and sample structural features can be extracted from the digital surface model and canopy height model corresponding to the sample tree species information.

[0111] D4: Construct sample data based on sample image features and sample structure features.

[0112] D5: Input the sample data into the hybrid expert model to obtain the sample tree identification results.

[0113] The sample tree identification results include the type and confidence level of the sample tree.

[0114] D6: Calculate the loss function between the sample tree identification results and the actual tree identification results corresponding to the sample data.

[0115] Optionally, the mean squared error between the sample tree identification results and the actual tree identification results corresponding to the sample data can be calculated, and the mean squared error can be used as the loss function.

[0116] D7: If the loss function does not converge, adjust the model parameters of the hybrid expert model and return to the execution process D5.

[0117] It should be noted that if the loss function does not converge, it means that the loss function has not decreased as the training steps increase. In this case, adjust the model parameters of the hybrid expert model and return to the execution process D5 until the loss function converges.

[0118] D8: If the loss function converges, then the training of the hybrid expert model is complete.

[0119] Understandably, if the loss function converges, that is, if the loss function gradually decreases as the training steps increase and eventually stabilizes or no longer changes significantly, it indicates that the training of the hybrid expert model is complete.

[0120] For a better explanation of the above, please refer to [link / reference]. Figure 2 The diagram illustrates a remote sensing tree species identification method. First, high-resolution images, DSM data, and sample tree species information (i.e., forest resource survey data in the diagram) are acquired by a drone. Based on the high-resolution images and DSM data, a canopy height model is determined. The canopy height model is then segmented using a SAM model to obtain the segmentation results (i.e., individual tree canopy segmentation results in the diagram). Feature extraction is performed on the segmentation results to obtain image features and structural features. Individual tree label maps are determined based on the forest resource survey data and segmentation results. Training samples are then determined based on the individual tree label maps. The hybrid expert model is trained using these training samples. Once the hybrid expert model is trained, the image features and structural features are input into the hybrid expert model to obtain the tree identification results.

[0121] It should be noted that, based on the processes shown in S101-S105 above, this embodiment can achieve the following beneficial effects:

[0122] 1. By employing a segmentation model and combining it with a canopy height model to automatically generate prompts, pixel-level accurate segmentation of individual tree canopies is achieved. Compared with traditional watershed algorithms, this method significantly improves performance in areas where canopies overlap, effectively separating overlapping canopies and obtaining individual tree samples with complete morphology and clear boundaries. This method fundamentally solves the problem of subsequent classification features being "contaminated" due to inaccurate segmentation, laying a crucial data foundation for high-precision tree species identification.

[0123] 2. The hybrid expert model abandons the traditional "one-size-fits-all" single model and adopts a "gated routing, expert divide-and-conquer" strategy. This model can adaptively assign tree canopies with different characteristics to the most suitable experts for identification, thereby learning more targeted and finer-grained discriminative features. When faced with complex forest stands with diverse tree species and subtle interspecies differences, this method shows significant advantages in recognition accuracy, robustness, and generalization ability.

[0124] 3. The hybrid expert model deeply integrates three-dimensional features such as tree height, crown width, volume, and surface structure extracted from the digital surface model. This information serves as a key basis for distinguishing tree species with similar spectra but different three-dimensional morphologies. Through the effective fusion of multimodal data, a more discriminative feature space is constructed, thereby significantly enhancing the model's discriminative ability and effectively reducing the probability of confusion between categories.

[0125] 4. This application's embodiments construct a complete workflow solution from prompt generation, canopy segmentation, feature extraction to classification and recognition, significantly reducing reliance on manual intervention and parameter tuning. The method has a clear workflow, and its core modules possess excellent scalability, are easy to integrate and deploy, and can efficiently process large-scale remote sensing data, providing an innovative technical path that combines high precision and high efficiency for forestry resource surveys and ecological monitoring.

[0126] like Figure 3 The diagram shown is a schematic of the architecture of a remote sensing tree species identification device provided in an embodiment of this application. The identification device includes: an acquisition unit 100, a determination unit 200, a segmentation unit 300, an extraction unit 400, and an identification unit 500.

[0127] The acquisition unit 100 is used to acquire images and digital surface models of the area to be tested taken by a drone.

[0128] The determination unit 200 is used to determine the canopy height model based on the image and digital surface model.

[0129] The determining unit 200 is specifically used to: align the digital surface model according to the image to obtain the aligned digital surface model; calculate the difference between the aligned digital surface model and the preset digital terrain model to obtain the canopy height model.

[0130] Segmentation unit 300 is used to segment the canopy height model to obtain segmentation results; the segmentation results include the canopy mask or canopy polygon of each tree.

[0131] The segmentation unit 300 is specifically used to: filter the canopy height model using a filtering algorithm to obtain the canopy vertices of each tree; input the canopy vertices and images of each tree into the segmentation model to obtain the segmentation results.

[0132] The extraction unit 400 is used to extract features from the segmentation results to obtain image features and structural features.

[0133] The recognition unit 500 is used to input image features and structural features into a hybrid expert model to obtain tree recognition results. The hybrid expert model includes a gating network and an expert network. The gating network analyzes the image features and structural features to obtain a weight combination. The expert network recognizes the image features and structural features according to the weight combination to obtain tree recognition results.

[0134] The identification unit 500 includes: an expert network including at least multiple initial expert networks.

[0135] Obtain sub-units, used to extract the weights corresponding to each initial expert network from the weight combination.

[0136] The identification subunit is used to identify image features and structural features through each initial expert network, resulting in multiple initial tree identification results.

[0137] The computational subunit is used to perform weighted calculations on multiple initial tree identification results according to the weights corresponding to each initial expert network, so as to obtain the tree identification result.

[0138] In summary, a canopy height model is generated by fusing imagery and digital surface models, and the canopy boundaries of individual trees are accurately extracted. Secondly, a hybrid expert model is introduced, and a gating network dynamically schedules expert networks focused on different tree species to achieve refined classification. This progressive design significantly improves the accuracy of tree species identification.

[0139] Combination Figure 3 The identification device also includes, as shown, the following:

[0140] The information acquisition unit is used to acquire information about the sample tree species.

[0141] The matching unit is used to perform spatial matching of sample tree species information based on the segmentation results to obtain sample tree species labels.

[0142] The feature extraction unit is used to extract sample image features and sample structural features from the sample tree species labels.

[0143] The construction unit is used to construct sample data based on sample image features and sample structural features.

[0144] The input unit is used to input sample data into the hybrid expert model to obtain the sample tree recognition results.

[0145] The loss function unit is used to calculate the loss function between the sample tree identification result and the actual tree identification result corresponding to the sample data.

[0146] The adjustment unit is used to adjust the model parameters of the hybrid expert model if the loss function does not converge, and then return to the step of inputting sample data into the hybrid expert model to obtain the sample tree recognition results.

[0147] The training determines the unit, which is used to determine that the training of the hybrid expert model is complete if the loss function converges.

[0148] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. In particular, for system or system embodiments, since they are fundamentally similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The systems and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. Components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0149] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0150] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A remote sensing tree species identification method, characterized in that, include: Acquire images and digital surface models of the area to be tested taken by drones; Determine the canopy height model based on the image and the digital surface model; The canopy height model is segmented to obtain segmentation results; the segmentation results include the canopy mask or canopy polygon for each tree; Feature extraction is performed on the segmentation results to obtain image features and structural features; The image features and structural features are input into a hybrid expert model to obtain tree recognition results; wherein, the hybrid expert model includes a gating network and an expert network; the gating network analyzes the image features and structural features to obtain a weight combination; the expert network identifies the image features and structural features according to the weight combination to obtain tree recognition results.

2. The method according to claim 1, characterized in that, The determination of the canopy height model based on the image and the digital surface model includes: Align the digital surface model according to the image to obtain the aligned digital surface model; The difference between the aligned digital surface model and the preset digital terrain model is calculated to obtain the canopy height model.

3. The method according to claim 1, characterized in that, The segmentation process of the canopy height model to obtain the segmentation result includes: The canopy height model is filtered using a filtering algorithm to obtain the canopy apex of each tree; The crown vertices of each tree and the image are input into the segmentation model to obtain the segmentation results.

4. The method according to claim 1, characterized in that, The expert network includes at least multiple initial expert networks; The step of identifying the tree by combining the weights based on the segmentation results using the expert network includes: Obtain the weights corresponding to each of the initial expert networks from the weight combinations; Multiple initial tree identification results are obtained by identifying the image features and structural features through each initial expert network; Based on the weights corresponding to each initial expert network, the multiple initial tree identification results are weighted and calculated to obtain the tree identification result.

5. The method according to claim 1, characterized in that, The training process of the hybrid expert model includes: Obtain sample tree species information; Based on the segmentation results, spatial matching is performed on the sample tree species information to obtain sample tree species labels; Extract the sample image features and sample structure features from the sample tree species labels; Based on the sample image features and the sample structural features, sample data is constructed; The sample data is input into a hybrid expert model to obtain the sample tree identification results; Calculate the loss function between the sample tree identification results and the actual tree identification results corresponding to the sample data; If the loss function does not converge, the model parameters of the hybrid expert model are adjusted, and the process returns to the step of inputting the sample data into the hybrid expert model to obtain the sample tree recognition result. If the loss function converges, then the training of the hybrid expert model is considered complete.

6. A remote sensing tree species identification device, characterized in that, include: The acquisition unit is used to acquire images and digital surface models of the area to be tested taken by a drone. A determining unit is configured to determine a canopy height model based on the image and the digital surface model; A segmentation unit is used to segment the canopy height model to obtain a segmentation result; the segmentation result includes the canopy mask or canopy polygon for each tree; An extraction unit is used to extract features from the segmentation results to obtain image features and structural features; The recognition unit is used to input the image features and the structural features into a hybrid expert model to obtain tree recognition results; wherein, the hybrid expert model includes a gating network and an expert network; the gating network analyzes the image features and the structural features to obtain a weight combination; the expert network recognizes the image features and the structural features according to the weight combination to obtain tree recognition results.

7. The apparatus according to claim 6, characterized in that, The determining unit is specifically used for: Align the digital surface model according to the image to obtain the aligned digital surface model; The difference between the aligned digital surface model and the preset digital terrain model is calculated to obtain the canopy height model.

8. The apparatus according to claim 6, characterized in that, The segmentation unit is specifically used for: The canopy height model is filtered using a filtering algorithm to obtain the canopy apex of each tree; The crown vertices of each tree and the image are input into the segmentation model to obtain the segmentation results.

9. The apparatus according to claim 6, characterized in that, The expert network includes at least multiple initial expert networks; The identification unit includes: A sub-unit is used to obtain the weights corresponding to each of the initial expert networks from the weight combination; The identification subunit is used to identify the image features and structural features through each of the initial expert networks to obtain multiple initial tree identification results; The calculation subunit is used to perform weighted calculation on multiple initial tree identification results according to the weights corresponding to each initial expert network, so as to obtain the tree identification result.

10. The apparatus according to claim 6, characterized in that, Also includes: The information acquisition unit is used to acquire information about the sample tree species. A matching unit is used to perform spatial matching on the sample tree species information based on the segmentation result to obtain sample tree species labels; The feature extraction unit is used to extract sample image features and sample structural features from the sample tree species labels; A construction unit is used to construct sample data based on the sample image features and the sample structural features; The input unit is used to input the sample data into the hybrid expert model to obtain the sample tree recognition result; The loss function unit is used to calculate the loss function between the sample tree identification result and the actual tree identification result corresponding to the sample data; The adjustment unit is used to adjust the model parameters of the hybrid expert model if the loss function does not converge, and then return to the step of inputting the sample data into the hybrid expert model to obtain the sample tree recognition result. The training determination unit is used to determine that the training of the hybrid expert model is complete if the loss function converges.

Images