A low-altitude remote sensing virtual sample generation and intelligent investigation method based on real forest three-dimensional structure driving

Abstract

The application discloses a kind of low-altitude remote sensing virtual sample generation and intelligent investigation method based on real forest three-dimensional structure driving, belong to remote sensing intelligent interpretation and forest resource investigation field.Acquire unmanned aerial vehicle laser radar point cloud or photogrammetry point cloud, construct three-dimensional forest scene including terrain, individual and crown layer spatial structure of forest;According to sensor parameter, observation visual angle, illumination and imaging condition, generate remote sensing image with space label by rendering;Domain migration is used to optimize the feature difference between rendered image and real image, and a virtual sample dataset with real observation features is constructed;The intelligent interpretation model is trained using the dataset, and forest target detection, semantic segmentation, instance segmentation and intelligent investigation of trees are realized.The present application can automatically construct large-scale high-quality labeled samples, reduce the cost of manual sampling, and improve the adaptability of the model in complex forest environment.

Description

Technical Field

[0001] This invention belongs to the fields of remote sensing information technology, forest resource survey, artificial intelligence and digital forestry technology. Specifically, it relates to a method for generating and intelligently surveying low-altitude remote sensing virtual samples based on the three-dimensional structure of real forests. In particular, it relates to a method for automatically constructing forest remote sensing samples and conducting intelligent forest resource surveys by utilizing three-dimensional forest scene construction, virtual generation of remote sensing images and training of intelligent interpretation models. Background Technology

[0002] Currently, forest resource surveys require obtaining various structural information, such as canopy cover, stand density, spatial location of individual trees, canopy extent, and canopy width of individual trees. Traditional forest surveys mainly rely on ground plot measurements and manual interpretation methods. Although these methods can obtain relatively accurate survey results, they often suffer from low survey efficiency, limited spatial coverage, and difficulty in meeting the needs of large-scale, continuous forest monitoring.

[0003] In recent years, low-altitude unmanned aerial vehicle (UAV) remote sensing technology has provided a new technical means for refined forest surveys due to its high spatial resolution and flexible acquisition capabilities. Combining deep learning-based target detection, semantic segmentation, and instance segmentation methods, forest targets can be automatically identified and forest structure information extracted from UAV remote sensing imagery. However, the performance of deep learning models is highly dependent on large-scale, high-quality labeled samples. Because forest scenes possess highly complex three-dimensional structural features, significant differences exist between different regions in tree species composition, stand density, canopy structure, spatial distribution, and observation conditions. Obtaining manually labeled samples covering various forest types and environmental conditions requires substantial manual input, limiting the widespread application of intelligent survey models.

[0004] To reduce the reliance of intelligent remote sensing interpretation models on manually labeled data, existing technologies attempt to construct virtual remote sensing samples through computer simulation. For example, virtual forests are constructed using parametric tree models or regularized 3D scene generation methods, and then further simulated with remote sensing images and corresponding labels. However, forests, as typical complex natural scenes, are influenced by various factors such as individual tree differences, growth status, canopy layers, branch and leaf distribution, and spatial relationships within forest stands. Existing virtual forest scenes generated based on regularized models typically rely on preset parameters or simplified structural assumptions, making it difficult to fully reflect the complex 3D spatial heterogeneity of real forests. This results in differences between the generated simulated images and real low-altitude remote sensing images in terms of canopy morphology, occlusion relationships, shadow distribution, and background features, thus affecting the application effectiveness of intelligent interpretation models trained on simulated samples in real forest environments.

[0005] Furthermore, while some existing methods improve the visual difference between simulated and real images through image generation or style transfer techniques, they typically focus primarily on adjusting two-dimensional image features, making it difficult to simultaneously ensure image realism and consistency with the spatial structure of forest targets. When the appearance of an image changes, it may lead to shifts in tree canopy outlines, individual tree positions, and corresponding label information, reducing the effectiveness of virtual samples in tasks such as target detection and instance segmentation.

[0006] Therefore, how to utilize the three-dimensional structural information of real forests to drive the construction of virtual forest scenes, and combine it with the remote sensing imaging process to generate low-altitude remote sensing virtual samples with real observation characteristics and maintain spatial label consistency, so as to support the training of intelligent forest interpretation models and large-scale forest resource surveys, is a technical problem that urgently needs to be solved. Summary of the Invention

[0007] To achieve the above-mentioned objectives of this invention, this invention provides a method for generating and intelligently surveying low-altitude remote sensing virtual samples based on a real forest three-dimensional structure, comprising the following steps: (1) Obtain large-scale publicly available UAV lidar point cloud data or photogrammetric point cloud data, process the point cloud data, extract forest topography information, individual tree spatial location and canopy structure features, and construct a three-dimensional forest scene model with real spatial structure features. (2) Based on the three-dimensional forest scene model, according to the parameters of the low-altitude remote sensing sensor, the observation angle, the spatial resolution, the illumination conditions and the imaging parameters, the three-dimensional forest scene is virtually rendered to generate a remote sensing rendering image with real observation conditions; at the same time, according to the location, canopy structure and category information of individual trees in the three-dimensional forest scene, the target category label, instance label or pixel-level mask label corresponding to the space of the remote sensing rendering image is automatically generated. (3) Utilize generative artificial intelligence models, image generation models or other domain adaptation methods to optimize the feature distribution differences between remote sensing rendered images and real low-altitude remote sensing images, generate virtual simulation remote sensing images with real remote sensing observation features, maintain spatial consistency between virtual simulation remote sensing images and corresponding labels, and construct a large-scale low-altitude remote sensing virtual sample dataset with automatic annotation information. (4) Based on the aforementioned low-altitude remote sensing virtual sample dataset, the intelligent interpretation model is trained to obtain an intelligent interpretation model for forest remote sensing surveys; (5) Obtain low-altitude remote sensing images of the target survey area, and input the low-altitude remote sensing images into the trained intelligent interpretation model to realize forest target recognition, semantic segmentation, instance segmentation and automatic extraction of forest survey parameters.

[0008] Furthermore, the three-dimensional forest scene model is obtained using either a parametric construction method or a reconstruction method based on real observation data. The parametric construction method is based on a three-dimensional single-tree model, setting parameters such as tree species composition, spatial distribution of single trees, stand density, tree height, crown morphology, and posture according to the target forest type to generate virtual forest scenes with different structural features. The reconstruction method based on real observation data utilizes lidar point clouds or photogrammetric point clouds, employing spatial discretization, voxelization, or three-dimensional reconstruction methods to preserve the spatial distribution of trees, canopy structure, and spatial heterogeneity of vegetation in a real forest.

[0009] Furthermore, the low-altitude remote sensing virtual imaging process is set according to the remote sensing sensor parameters, flight platform parameters, and environmental observation conditions. These parameters include, but are not limited to, sensor type, flight altitude, observation angle, solar irradiance conditions, sky diffuse light ratio, image spatial resolution, and scene range, in order to generate remote sensing rendering images under different observation conditions.

[0010] Furthermore, the domain transfer method is used to reduce the feature distribution differences between remotely sensed rendered images and real low-altitude remotely sensed images, including but not limited to image feature transformation based on generative artificial intelligence models, adversarial learning models, diffusion models, or other domain adaptation methods.

[0011] Furthermore, during the domain migration process, target label information automatically generated based on the 3D forest scene is introduced to constrain the structural consistency of the image conversion process, so that the forest targets before and after the domain migration maintain spatial correspondence and ensure that the number of trees, the location of individual trees, the canopy outline and the instance boundary information do not change.

[0012] Furthermore, the intelligent interpretation model is trained based on the constructed virtual simulation remote sensing sample dataset, and can be combined with a small number of real remote sensing images of the target survey area and corresponding labels for transfer optimization to improve the model's adaptability to different forest types and different observation conditions.

[0013] Furthermore, the intelligent interpretation model includes a target detection model, a semantic segmentation model, an instance segmentation model, or other remote sensing information extraction models, used to obtain spatial distribution information of forest targets and further calculate forest survey parameters.

[0014] Furthermore, the forest survey parameters include, but are not limited to, canopy coverage, stand density, number of individual trees, location of individual trees, and canopy size information. Specifically, canopy coverage is determined based on the extracted canopy area and the survey area, stand density is determined based on the number of individual trees per unit area, and canopy size is calculated based on the canopy space range.

[0015] Compared with the prior art, the embodiments of the present invention have the following beneficial effects: (1) This invention utilizes real forest lidar point clouds or photogrammetric point clouds to construct a three-dimensional forest scene model with real spatial structure characteristics, so that the virtual forest scene can retain the spatial distribution of trees, canopy structure and spatial heterogeneity characteristics in the actual forest, thereby improving the consistency between virtual remote sensing samples and real forest scenes.

[0016] (2) The present invention performs virtual rendering of remote sensing images based on three-dimensional forest scenes and generates corresponding target label information simultaneously, realizing the automated construction of remote sensing images and labeled data, reducing the human cost required to obtain large-scale forest remote sensing training samples by traditional manual labeling methods.

[0017] (3) This invention optimizes the feature distribution difference between remote sensing rendered images and real low-altitude remote sensing images by using the domain transfer method. While maintaining the consistency of the spatial structure and label information of forest targets, it improves the authenticity and effectiveness of virtual samples and enhances the applicability of the intelligent interpretation model trained on virtual samples in real forest environments.

[0018] (4) The virtual sample generation method constructed in this invention can flexibly configure the three-dimensional scene and imaging parameters according to different forest structure types, sensor parameters and observation conditions, and can be extended to generate remote sensing training samples of different regions and types of forests, providing data support for the training and application of intelligent forest remote sensing survey models.

[0019] (5) The intelligent interpretation model based on virtual sample pre-training of the present invention can be applied to remote sensing survey tasks in different forest areas. While reducing the need for manual annotation of target areas, it realizes forest target recognition, semantic segmentation, instance segmentation and automatic extraction of forest survey parameters, thereby improving the automation and intelligence level of forest resource survey. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall process of the low-altitude remote sensing virtual sample generation and intelligent survey method driven by the real three-dimensional structure of a forest, as described in this invention.

[0021] Figure 2 This is a schematic diagram illustrating the construction of a 3D forest scene model based on the reconstruction method of real observation data and the parametric modeling method of this invention.

[0022] Figure 3 This is a schematic diagram of the low-altitude remote sensing virtual imaging rendering process based on a three-dimensional forest scene according to the present invention.

[0023] Figure 4 This is a schematic diagram illustrating the automatic generation of remote sensing rendered images and corresponding labels according to the present invention.

[0024] Figure 5This is a comparative diagram of the remote sensing rendered image, virtual simulation remote sensing image, real low-altitude remote sensing image, and corresponding labels of the present invention.

[0025] Figure 6 This is a schematic diagram of the results of intelligent interpretation of forest remote sensing and extraction of forest survey parameters according to the present invention. Detailed Implementation

[0026] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0027] Example 1: Low-altitude remote sensing virtual sample generation and intelligent survey based on real forest point cloud reconstruction Reference Figure 1 As shown, this embodiment sequentially performs the following steps: three-dimensional forest scene model construction, low-altitude remote sensing virtual imaging rendering and automatic label generation, virtual and real image domain migration, training of forest remote sensing intelligent interpretation model, and extraction of forest survey parameters for the target survey area.

[0028] Step (1): Refer to Figure 1 Step (1) and Figure 2 In method (a) of this embodiment, a 90 m × 130 m artificial broad-leaved forest plot is selected as the target forest area, and dense point cloud data of the plot is obtained from publicly available UAV photogrammetry data. First, the dense point cloud data is denoised, coordinate unified, and registered, and ground points and vegetation points are separated to extract plot topography information, individual tree spatial locations, and canopy structure features. Then, the vegetation point cloud is spatially discretized at a voxel scale of 1 m, and leaf area spatial distribution, leaf tilt angle distribution, and vegetation structure density parameters are set according to the point cloud distribution within each voxel. Subsequently, the processed point cloud data is imported into the 3D forest scene construction module to reconstruct a 3D forest scene model corresponding to the spatial structure of the sample point cloud. This model records the spatial location, canopy structure, and target category of each individual tree, and assigns a unique single-tree instance identifier to each individual tree for subsequent virtual imaging rendering and automatic tag generation.

[0029] Step (2): Refer to Figure 1 Step (2) and Figure 3 As shown, the sensor parameters, flight altitude, observation angle, spatial resolution, and imaging conditions of the real low-altitude remote sensing image corresponding to the target application scenario are read, and the observation zenith angle, observation azimuth angle, solar irradiance conditions, and sky diffuse light ratio of the three-dimensional forest scene are set accordingly. In this embodiment, the orthophoto observation method is adopted, the output image size is set to 1000×1000 pixels, and the three-dimensional radiative transfer model and virtual imaging rendering engine are used to perform virtual imaging of the three-dimensional forest scene obtained in step (1) to obtain a remote sensing rendering image containing information on the forest canopy, individual trees, and background.

[0030] While generating the remote sensing rendered image, the target categories and individual tree instance identifiers recorded in the 3D forest scene are projected onto the imaging plane according to the same imaging geometry, scene range, and spatial resolution. The corresponding target category and individual tree instance identifier are written to each output pixel, automatically generating target category labels, individual tree instance labels, semantic segmentation labels, and pixel-level mask labels that are registered pixel-by-pixel with the remote sensing rendered image. (Refer to...) Figure 4 As shown, the remote sensing rendered image and the corresponding colored single tree instance label have the same number of trees, tree location and crown outline. Different colors in the instance label correspond to different single tree instance identifiers.

[0031] Keeping the 3D forest scene structure unchanged, the observation angle, lighting conditions and spatial resolution are adjusted sequentially to generate remote sensing rendering images and their corresponding labels under different observation conditions in batches; the remote sensing rendering images generated in the same batch are associated with the labels according to the scene number and observation parameters, and their individual tree instance identifiers are recorded to form original virtual samples for subsequent domain migration.

[0032] Step (3): Refer to Figure 1 Step (3) and Figure 5 As shown, real low-altitude remote sensing images corresponding to the target application scenario of artificial broad-leaved forests are obtained. The remote sensing rendering image generated in step (2) is used as the source domain sample, and the real low-altitude remote sensing image is used as the target domain sample. A generative domain transfer model with structural consistency constraints is used to establish the mapping relationship between the source domain and the target domain. The remote sensing rendering image is input into the generative domain transfer model to optimize its spectral or color features, texture features, spatial distribution features, and imaging noise features, thereby generating a virtual simulation remote sensing image with the observation features of real low-altitude remote sensing images.

[0033] During the training of the domain transfer model, the labels automatically generated in step (2) are used as structural consistency constraint information to ensure that the number of trees, the spatial location of individual trees, the crown outline, the target category, and the instance boundary remain consistent before and after the domain transfer. After the domain transfer is completed, the labels corresponding to the remote sensing rendered images are directly used as the supervision labels for the virtual simulation remote sensing images, and the virtual simulation remote sensing images, corresponding labels, and imaging parameters are associated and stored to construct a low-altitude remote sensing virtual sample dataset with automatic annotation information. Real low-altitude remote sensing images are used to characterize the image features of the target domain and do not need to be registered pixel-by-pixel with the remote sensing rendered images.

[0034] Step (4): Refer to Figure 1In step (4), the virtual simulation remote sensing images and their corresponding labels from the low-altitude remote sensing virtual sample dataset are input into the forest remote sensing intelligent interpretation model for training. This embodiment employs an interpretation method combining semantic segmentation and instance segmentation: the semantic segmentation model uses target category labels or pixel-level mask labels as supervisory information to learn pixel-level classification features of forest areas and canopy areas; the instance segmentation model uses individual tree instance labels as supervisory information to learn the canopy instances of each individual tree and their contour features. After training, the model parameters and model weights are saved to obtain the forest remote sensing intelligent interpretation model used for forest target recognition, canopy segmentation, and individual tree extraction.

[0035] A small number of images are selected from real low-altitude remote sensing images of the target survey area and corresponding labels are created. These images are then used as fine-tuning samples to input into the trained forest remote sensing intelligent interpretation model. The model parameters are optimized to adapt the model to the image features and forest structure features of the target survey area.

[0036] Step (5): Refer to Figure 1 Step (5) and Figure 6 As shown, low-altitude remote sensing images of the target survey area are acquired and input into the forest remote sensing intelligent interpretation model obtained in step (4). The semantic segmentation model outputs the forest area and the canopy area, and the instance segmentation model outputs the distinguishable individual tree instances, the spatial location of the individual tree, and the outline of the individual tree canopy, thereby completing the forest target identification, semantic segmentation, and instance segmentation of the target survey area.

[0037] Based on the instance segmentation results, the number N of individual trees within the target survey area is counted. Then, according to the image ground spatial resolution, the crown outline of each individual tree is transformed to a ground coordinate scale to obtain the ground projection size of each tree crown outline and determine the crown width. Using the effective survey area A of the target survey area as a benchmark, according to... Calculate the stand density D; perform area statistics on the canopy region in the semantic segmentation results to obtain the ground projection area of ​​the canopy region. and in accordance with Calculate the canopy cover C. The final output includes the number of individual trees, their spatial location, canopy width, stand density, and canopy cover.

[0038] Example 2: Low-altitude remote sensing virtual sample generation and intelligent survey based on parametric modeling Reference Figure 1 Step (1) and Figure 2In method (b) of the above, this embodiment establishes a parametric forest scene of 55 m × 55 m. First, a three-dimensional single tree model library containing different tree species and different crown morphologies is established; then, according to the preset target forest structure, the tree species composition, tree height, crown morphology, single tree spatial distribution and stand density are set, 60 three-dimensional single tree models are selected from the three-dimensional single tree model library, the scale and posture of each three-dimensional single tree model are adjusted, and they are configured in a scene with terrain information to construct a three-dimensional forest scene model.

[0039] Each 3D tree model in the scene is configured with a unique tree instance identifier and target category information, and its spatial location and canopy structure are recorded. Based on this, the tree species composition, tree height, canopy morphology, tree spatial distribution and stand density are changed in sequence to generate multiple 3D forest scene models with different stand structure characteristics and tree instance identifiers.

[0040] For each three-dimensional forest scene model constructed by parameterization, steps (2) to (5) in Example 1 are executed in sequence to generate remote sensing rendering images and their corresponding spatial labels simultaneously, complete the virtual and real image domain migration and the construction of low-altitude remote sensing virtual sample dataset, train the forest remote sensing intelligent interpretation model, and use the model to interpret the low-altitude remote sensing images of the target survey area to obtain the number of individual trees, the spatial location of individual trees, the crown width of individual trees, the stand density and the canopy coverage.

[0041] The above are merely preferred embodiments of the present invention. Any equivalent substitutions or transformations made by those skilled in the art to the 3D forest scene construction method, virtual imaging model, domain migration method, intelligent interpretation model, or parameter settings without departing from the technical concept of the present invention should fall within the protection scope of the present invention.

Claims

1. A method for generating virtual samples and conducting intelligent surveys based on real forest 3D structure driven by low-altitude remote sensing, comprising the following steps: (1) Obtain large-scale publicly available UAV lidar point cloud data or photogrammetric point cloud data, process the point cloud data, and construct a three-dimensional forest scene model containing terrain information, individual tree spatial location and canopy structure features. (2) Based on the three-dimensional forest scene model, according to the preset low-altitude remote sensing observation parameters, sensor parameters and imaging conditions, the three-dimensional forest scene is virtually imaged and rendered to generate a remote sensing rendering image with spatial correspondence, and the target category label, instance label or pixel-level mask label corresponding to the remote sensing rendering image is generated simultaneously. (3) Optimize the feature distribution differences between the remote sensing rendered image and the real low-altitude remote sensing image by using a generative artificial intelligence model or domain adaptation method, generate a virtual simulation remote sensing image with the observation features of the real remote sensing image, and construct a low-altitude remote sensing virtual sample dataset with automatic annotation information. (4) Based on the aforementioned low-altitude remote sensing virtual sample dataset, the intelligent interpretation model is trained to obtain an intelligent interpretation model for forest remote sensing surveys; (5) Obtain low-altitude remote sensing images of the target survey area, and input the low-altitude remote sensing images into the intelligent interpretation model to realize forest target recognition, semantic segmentation, instance segmentation and forest survey parameter extraction.

2. The method for generating virtual samples and conducting intelligent surveys based on the three-dimensional structure of a real forest using low-altitude remote sensing, as described in claim 1, is characterized in that... The three-dimensional forest scene model in step (1) is constructed by either a reconstruction method based on real observation data or a parametric modeling method. Specifically, the reconstruction method based on real observation data is based on existing forest lidar point cloud data or photogrammetric point cloud data. The point cloud data is spatially discretized or voxelized, and the parameters of leaf area spatial distribution, leaf tilt angle distribution, voxel scale, and vegetation structure density are adjusted according to the forest structure characteristics to reconstruct a three-dimensional forest scene model with real spatial structure characteristics. The parametric modeling method is based on a three-dimensional single tree model. By setting single tree structure parameters and stand spatial parameters, the tree height, crown morphology, single tree spatial distribution, stand density, and composition structure are controlled to generate three-dimensional forest scene models that conform to the target forest structure characteristics in batches.

3. The method for generating virtual samples and conducting intelligent surveys based on the three-dimensional structure of a real forest using low-altitude remote sensing, as described in claim 1, is characterized in that... In step (2), based on the three-dimensional forest scene model, virtual imaging rendering is performed on the three-dimensional forest scene according to the parameters of the low-altitude remote sensing sensor, observation viewpoint, spatial resolution, illumination conditions, and imaging parameters to generate a remote sensing rendered image corresponding to the actual low-altitude remote sensing observation process. During the virtual imaging rendering process, the radiometric response process of ground objects under different observation conditions is simulated based on the spatial structural characteristics and optical attribute parameters of different components in the three-dimensional forest scene, obtaining a remote sensing rendered image containing information on the forest canopy, individual trees, and background. Simultaneously, based on the pre-acquired spatial location, canopy structure, and target category information of individual trees in the three-dimensional forest scene model, label data spatially registered with the remote sensing rendered image is automatically generated through spatial projection relationships. This includes, but is not limited to, canopy location labels, individual tree instance labels, semantic segmentation labels, or pixel-level mask labels, thereby batch-constructing a low-altitude remote sensing virtual sample dataset containing remote sensing images and corresponding label information.

4. The method for generating virtual samples and conducting intelligent surveys based on the three-dimensional structure of a real forest using low-altitude remote sensing, as described in claim 1, is characterized in that... In step (3), real low-altitude remote sensing image data is acquired, and a virtual-real image domain mapping relationship is constructed based on the real low-altitude remote sensing image data and the remote sensing rendered image obtained in step (2). The domain transfer method is used to reduce the feature distribution difference between the remote sensing rendered image and the real low-altitude remote sensing image. The domain transfer method includes optimizing the spectral features, texture features, spatial distribution features, and imaging noise features of the remote sensing rendered image based on generative artificial intelligence models, image generation models, feature alignment models, or other domain adaptation methods to generate a virtual simulation remote sensing image with the observation features of the real low-altitude remote sensing image. During the domain transfer process, the spatial structure information and label information generated by the three-dimensional forest scene model in the remote sensing rendered image are kept unchanged, so that the number of trees, the spatial location of individual trees, the crown outline, and the target category information are kept consistent between the virtual simulation remote sensing image and the corresponding label, thereby constructing a large-scale low-altitude remote sensing virtual sample dataset with automatic annotation information.

5. The method for generating virtual samples and conducting intelligent surveys based on the three-dimensional structure of a real forest using low-altitude remote sensing, as described in claim 1, is characterized in that... In step (4), the intelligent interpretation model is trained based on the constructed low-altitude remote sensing virtual sample dataset to obtain a remote sensing intelligent interpretation model suitable for different forest structure types and survey tasks. The low-altitude remote sensing virtual sample dataset includes remote sensing image data and their corresponding label information generated under different forest types, stand structures, spatial scales, and observation conditions. The intelligent interpretation model includes, but is not limited to, target detection models, semantic segmentation models, instance segmentation models, and other remote sensing information extraction models based on deep learning or machine learning. By using the virtual sample dataset for model training, model parameters or model weights are obtained for forest target identification, individual tree extraction, canopy segmentation, and forest survey parameter extraction.

6. The method for generating virtual samples and conducting intelligent surveys based on the three-dimensional structure of a real forest using low-altitude remote sensing, as described in claim 1, is characterized in that... In step (5), a forest remote sensing intelligent interpretation model is trained based on the constructed low-altitude remote sensing virtual sample dataset. The low-altitude remote sensing image of the area to be investigated is input into the intelligent interpretation model to achieve automated extraction of forest information in the target area. The intelligent interpretation model is pre-trained based on virtual simulation samples of different forest structure types, different observation conditions, and different survey tasks. It can intelligently interpret new forest area remote sensing images without relying on a large number of manually labeled samples in the target area for retraining. Through the intelligent interpretation model, semantic segmentation, target detection, instance segmentation, or other remote sensing information extraction methods can be used to automatically obtain information on tree canopy range, number of individual trees, spatial location of individual trees, and crown width of individual trees in the target forest area, and further calculate forest survey factors such as canopy coverage, stand density, and individual tree structure parameters. Simultaneously, based on a small amount of real remote sensing images and corresponding label data of the target survey area, the intelligent interpretation model can be optimized or subjected to transfer learning to further improve the survey accuracy of the model in specific areas and specific forest types.

Images