A method and system for extracting forest stand boundaries

Abstract

The application discloses a kind of forest stand boundary extraction method and system, the method includes the following steps: obtaining data, including forest remote sensing image data and ground data;The data obtained is preprocessed;Double-path CBAM-UNet network model is constructed;The data after pre-processing is input into double-path CBAM-UNet network model to carry out tree species classification prediction and obtain tree species classification chart;Image mosaic is carried out to tree species classification chart to obtain tree species classification grid chart;Vectorization processing and vector boundary optimization are carried out to tree species classification grid chart to extract forest stand boundary and generate forest stand boundary vector chart.The forest stand boundary extraction method provided in the application can realize multi-scale feature capture and feature fusion of multi-source heterogeneous data through the double-path CBAM-UNet network model constructed, can solve the problem of automatic extraction of forest stand boundary according to high-precision identification of forest stand tree species under complex forest conditions, and provide high-quality data for forest management.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of forestry resource monitoring, and particularly relates to a forest stand boundary extraction method and system. BACKGROUND

[0002] Forest stand boundary division discretizes continuous forest space into quantifiable management units, facilitating the statistics of forest area, volume and other indicators, clarifying the responsibility scope of management subjects, and providing spatial basis for harvesting, tending and other management activities. Through forest stand boundary division, spatial structures such as ecological corridors and buffer zones can be constructed to improve the stability of forest ecosystems. For example, China's forest resource inventory takes small plots as the basic unit, and through forest stand boundary division, dynamic monitoring of national forest resources is realized.

[0003] Forest stand boundary usually needs to be divided according to forest tree species identification. Tree species identification reveals the spatial heterogeneity of forests, providing an ecological basis for boundary demarcation. Different tree species have significant differences in community structure, functional role and environmental adaptability. If areas with different ecological characteristics are divided into the same management unit, it may lead to resource utilization conflicts or ecological function degradation. Therefore, the accuracy of forest tree species identification directly affects the accuracy of forest stand boundary division results.

[0004] Currently, high-precision forest tree species identification and forest stand boundary division mainly rely on manual field investigation, which requires professional personnel to enter the forest area and visually interpret remote sensing images to record the tree species composition and boundary information of each sample plot. This method is labor-intensive, costly, and difficult to implement in complex terrain areas. It is difficult to obtain continuous forest resource data on a large scale, and it is difficult to provide timely and accurate decision support for forestry management.

[0005] The automatic tree species identification methods based on remote sensing technology mainly include pixel-based supervised classification (PSC) (Franklin et al., 2000; Dalponte et al., 2012), object-based image analysis (OBIA) (Blaschke, 2010; Immitzer et al., 2016), machine learning classification (MLC) (Maxwell et al., 2018; Liu et al., 2018), and deep learning classification (DLC) developed in recent years (Hamraz et al., 2019; Onishi and Ise, 2021).

[0006] The PSC method compares the spectral features of each pixel in a remotely sensed image with known tree species samples, and determines the tree species category of the pixel using algorithms such as maximum likelihood and minimum distance. This method is simple to implement and computationally efficient, but it does not consider spatial context information, resulting in "salt-and-pepper noise" in the classification results and insufficient ability to distinguish between tree species with similar spectral features.

[0007] The OBIA method first segments the image into homogeneous objects, and then classifies them based on the objects' spectral, shape, and texture features. This method can utilize the spatial structure information of the tree canopy and reduce "salt-and-pepper noise," but the segmentation parameters are complex to adjust, and it is difficult to determine the optimal segmentation scale in heterogeneous forest stands, leading to over-segmentation or under-segmentation.

[0008] The MLC method primarily employs algorithms such as random forests, support vector machines, and artificial neural networks to establish classification models by extracting spectral and texture features. This method can handle high-dimensional data and has a strong ability to model nonlinear relationships between features; however, it heavily relies on manually designed feature engineering, has poor scalability, and limited generalization ability across different regions or different sensor data.

[0009] The DLC method has developed rapidly in recent years, mainly employing convolutional neural networks (CNNs) to automatically extract hierarchical features. The U-Net architecture and its variants, through encoder-decoder structures and skip connections, preserve spatial detail and are widely used in forest remote sensing segmentation tasks. However, most existing deep learning methods utilize only single optical remote sensing data, failing to adequately mine information on vertical structure and seasonal variations.

[0010] Regarding forest stand boundary delineation, the main methods include edge-based methods (EM) (Turker and Kok, 2013; Yan and Roy, 2014), region-based methods (RM) (Evans et al., 2002; Salman, 2006), and hybrid methods (HM) (Graesser and Ramankutty, 2017; Chen et al., 2020). These methods perform reasonably well with low spatial resolution or simple forest stand structures, but there is still significant room for improvement in automation and accuracy when dealing with high-resolution imagery and complex forest stand structures.

[0011] The above methods are proposed under specific research areas and specific stand conditions, and it is difficult to effectively distinguish tree species with similar spectral characteristics, and it is difficult to support large-scale operations. In the case of high canopy density, complex stand structure, and mixed natural forest and artificial forest, the existing method is difficult to realize high-precision tree species recognition and automatic division of stand boundary.

[0012] Therefore, developing a deep learning method using high-resolution optical remote sensing and canopy height model to realize high-precision recognition of tree species and automatic extraction of stand boundary in complex forest conditions has become a technical problem to be solved in modern fine management of forest. SUMMARY

[0013] In view of the above deficiencies of the existing forest tree species recognition technology, the purpose of the present application is to provide a stand boundary extraction method for solving the problem of automatic extraction of stand boundary according to high-precision recognition of stand tree species in complex forest conditions.

[0014] To achieve the above-mentioned purpose, the first aspect of the present application provides a stand boundary extraction method, which comprises the following steps:

[0015] Obtaining data, the data comprising forest remote sensing image data and ground data;

[0016] Preprocessing the obtained data;

[0017] Constructing a double-path CBAM-UNet network model;

[0018] Inputting the preprocessed data into the double-path CBAM-UNet network model for tree species classification prediction and obtaining a tree species classification map;

[0019] Image stitching is performed on the tree species classification map to obtain a tree species classification raster map;

[0020] Vectorization processing and vector boundary optimization are performed on the tree species classification raster map to extract the stand boundary and generate a stand boundary vector map.

[0021] According to the first aspect of the present application, the stand boundary extraction method comprises: preparing deep learning training samples according to the obtained data to train the double-path CBAM-UNet network model.

[0022] According to the first aspect of the present application, the preparation of deep learning training samples according to the obtained data to train the double-path CBAM-UNet network model comprises:

[0023] Preparing training samples, including image slicing and label preparation, and normalizing, feature engineering and data enhancement processing on the samples;

[0024] The model is initialized with a configuration including an AdamW optimizer, a cosine annealing schedule dynamic learning rate strategy, and a weighted cross-entropy loss function as a main loss function.

[0025] The model is trained using a mixed precision training strategy and a gradient clipping method to improve the training efficiency of the model and applying a random dropout strategy to enhance the generalization ability of the model.

[0026] According to the first aspect of the present application, the dual-path CBAM-UNet network model includes an input layer and an encoder, the input layer includes a left path and a right path, the encoder includes a left path encoder and a right path encoder, and the input layer is used to receive original image data and simultaneously input the original image data to the left path encoder and the right path encoder for parallel processing through the left path and the right path.

[0027] The left path encoder includes a multi-level feature extraction layer constructed based on down-sampling and a BasicConv module.

[0028] The right path encoder includes a multi-level feature extraction layer constructed based on down-sampling and a BasicConv module.

[0029] According to the first aspect of the present application, the left path encoder is used to process optical remote sensing image data, the right path encoder is used to process CHM canopy height model data, the optical remote sensing image data includes RGB bands or multispectral data, and the CHM canopy height model data provides tree height information, and the complementary fusion of spectral features and structural features is realized through dual-path feature extraction.

[0030] According to the first aspect of the present application, the dual-path CBAM-UNet network model sets a mode switch to control the working mode of the BasicConv module, and the working mode includes a standard mode and an ASPP enhanced mode.

[0031] When in the standard mode, the input features are processed through the basic convolution unit and then outputted.

[0032] When in the ASPP enhanced mode, the input features are processed through the basic convolution unit and the ASPP sub-module in parallel, the processing results of the two paths are fused through element-by-element addition, the fused features are processed through batch normalization and ReLU activation function, and then outputted after applying Dropout regularization.

[0033] According to the first aspect of the present application, the dual-path CBAM-UNet network model includes a decoder, an output layer, and a CBAM module.

[0034] The decoder is a four-layer structure, each layer of which comprises a transposed convolution module for up-sampling and a DoubleConv module for processing fused features;

[0035] The output layer comprises a convolution layer that maps the output features of the decoder to predefined tree species categories and applies a Softmax activation function to generate a probability distribution for each tree species category;

[0036] The CBAM module is used to enhance the fused features after each feature extraction of the left and right path encoders and feature fusion, and the enhanced features are fused with the output features of the transposed convolution module in the corresponding layer of the decoder through a skip connection, and the fused features are input into the DoubleConv module for processing.

[0037] According to the first aspect of the application, the double-path CBAM-UNet network model comprises a bottleneck layer, which comprises:

[0038] a left path DoubleConv module for processing the feature map output by the left path encoder;

[0039] a right path DoubleConv module for processing the feature map output by the right path encoder;

[0040] a bottleneck fusion layer that concatenates the feature maps output by the left path DoubleConv module and the right path DoubleConv module in the channel dimension to generate a concatenated feature map;

[0041] a terminal enhancement module that converts the concatenated feature into a highly nonlinear representation;

[0042] a CBAM enhancement module that applies attention mechanisms to the output of the bottleneck layer for enhancement processing.

[0043] According to the first aspect of the application, the vectorization processing and vector boundary optimization of the tree species classification raster map comprises the following steps:

[0044] Vectorizing the tree species classification raster map using GIS software, and converting the continuous area boundary of the same tree species into a forest stand boundary through spatial clustering of tree species categories;

[0045] Mapping the attribute values of the vectorized forest stand polygon to the tree species categories to establish a forest stand polygon attribute table;

[0046] Topological relationship processing to ensure that adjacent polygons share boundaries;

[0047] Applying geometric simplification processing to optimize the vector boundary of the forest stand;

[0048] Small area filtering is performed on the stand vector polygon, and the graph spot smaller than the threshold value is merged into the surrounding main stand;

[0049] The stand vector boundary is smoothed and optimized by using a boundary smoothing algorithm, so that the boundary is closer to the natural forest edge.

[0050] To achieve the above purpose, the second aspect of the present application provides a stand boundary extraction system, comprising:

[0051] A data acquisition module is used to collect forest remote sensing image data and ground data.

[0052] A data preprocessing module is used to preprocess the forest remote sensing image data and ground data collected by the data acquisition module.

[0053] A prediction module is used to perform tree species classification prediction according to the input image slice and obtain a tree species classification map.

[0054] An image processing module is used to perform splicing processing on the tree species classification map to obtain a complete tree species classification raster map.

[0055] A boundary extraction module is used to perform vectorization processing and vector boundary optimization on the tree species classification raster map to extract the stand boundary and generate a stand boundary vector map.

[0056] The advantages of the present application are as follows: first, the dual-path CBAM-UNet network model of the present application captures multi-scale features, fuses GF-7 multispectral images and high-resolution CHM data, and uses a deep learning model to obtain multi-scale features such as spectrum, spatial texture and vertical height, thereby improving the ability to distinguish similar tree species and the classification accuracy, and further improving the accuracy of stand boundary extraction. Second, in the area with mixed coniferous and broad-leaved trees and large terrain undulations, the classification accuracy is more than 81%, and the tree species and boundary can be stably identified for artificial pure forest and natural mixed forest, and the adaptability to complex forest land is strong. Third, a fully automated processing flow is constructed, manual intervention and field operation are reduced, and the classification map and stand boundary are automatically output, so that large-scale stand mapping update can be quickly completed, and an automated and efficient process is realized. Finally, through multi-source data feature fusion and algorithm improvement, the classification accuracy is better than that of traditional methods, and the Kappa coefficient reaches 0.78, which provides high-quality data for forest management. At the same time, under the condition of NVIDIA RTX4090 system and pytorch2.2 framework, it only takes 2 to 3 hours to process 12,300 hectares of test area, the calculation performance is excellent, and it is suitable for large-scale forest information automatic processing. BRIEF DESCRIPTION OF DRAWINGS

[0057] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings described below only illustrate some of the embodiments of the present application, and all other drawings obtained by those of ordinary skill in the art without creative effort based on these drawings also belong to the scope of protection of the present application.

[0058] Figure 1 A flowchart of the forest stand boundary extraction method according to the present application is shown in the figure.

[0059] Figure 2 A flowchart of the deep learning model sample making method according to the present application is shown in the figure.

[0060] Figure 3 A schematic diagram of the dual-channel CBAM-Unet network structure according to the present application is shown in the figure.

[0061] Figure 4 A schematic diagram of the BasicConv module structure according to the present application is shown in the figure.

[0062] Figure 5 A schematic diagram of the ASPP module structure according to the present application is shown in the figure.

[0063] Figure 6 A schematic diagram of the CBAM module structure according to the present application is shown in the figure.

[0064] Figure 7 A schematic diagram of the model training flow according to the present application is shown in the figure.

[0065] Figure 8 A schematic diagram of the tree species classification and prediction flow according to the present application is shown in the figure.

[0066] Figure 9 A schematic diagram of the forest stand boundary map automatically extracted by the method according to the present application and the existing forest resource small plot is shown in the figure.

[0067] Figure 10 A schematic diagram of the forest stand boundary extraction system according to the present application is shown in the figure. DETAILED DESCRIPTION

[0068] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings of the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort also belong to the scope of protection of the present application.

[0069] Before the present application is further described, the terms and nomenclature involved in the embodiments of the present application will be explained, which are applicable to the following explanations:

[0070] Canopy Height Model (CHM): A digital model reflecting the relative height of trees above the ground, calculated by subtracting the Digital Elevation Model (DEM) from the Digital Surface Model (DSM) obtained through LiDAR or photogrammetry techniques. It provides pixel-level tree height information and is widely used in forestry remote sensing, ecological research, and resource management, especially in tree species classification, forest boundary extraction, and biomass estimation.

[0071] ArcMap: Developed by the Environmental Systems Research Institute (ESRI), it is a core component of the GIS system and one of the main applications of the ArcGIS Desktop suite. It has powerful map making, spatial analysis, and data management functions, and can display, edit, and analyze map data. It provides data views and layout views, supports import and export of multiple data formats. Users can perform geographic processing tasks, publish map services, customize user experiences, and is widely used in urban planning, environmental protection, resource management, and other fields.

[0072] ASPP (Atrous Spatial Pyramid Pooling): A module for deep learning (especially computer vision) that aims to enhance the model's ability to perceive targets of different scales through multi-scale feature extraction and dilated convolution techniques.

[0073] Figure 1 The flowchart of the forest boundary extraction method of the present application is shown, which includes the following steps:

[0074] Step S1: Obtain data.

[0075] The data includes forest remote sensing image data and ground data.

[0076] The forest remote sensing image data used in the present application is sub-meter high spatial resolution multispectral data, which includes airborne digital images, satellite images (high-resolution satellite images), and canopy height model (CHM) data. Airborne digital images are used to create labels for deep learning model training, while satellite images are used for tree species classification and forest boundary delineation.

[0077] In the embodiments of the present application, the main data sources are GF-7 satellite multispectral optical images and canopy height model (CHM) data:

[0078] GF-7 satellite image: acquired in the vegetation growth period on May 23, 2023, the original multispectral band has a spatial resolution of 2.6 meters, and after fusing the panchromatic band, the spatial resolution is improved to about 0.5 meters. The image is subjected to radiation calibration, atmospheric correction, orthorectification and panchromatic fusion and other preprocessing processes to generate high-quality multispectral orthographic images, and resampled to 0.5-meter resolution.

[0079] CHM data: collected by airborne laser radar on September 3, 2023, with a resolution of 0.5 meters. By accurately registering and superimposing the CHM data and the GF-7 spectral image, the canopy height information of each pixel can be provided, realizing the fusion of spectral characteristics and spatial structure characteristics.

[0080] The ground data of the present application includes forest tree species sample point data collected by real-time kinematic differential (RTK) technology, and national forest resource second-class survey subcompartment data. Among them, the forest tree species sample point data records the tree species type and spatial coordinate position in detail, mainly used for model checking and verification; the national forest resource second-class survey subcompartment data can be used as a reference basis for label making.

[0081] In the embodiment of the present application, field sample plot investigation was carried out in Mulan test area in September 2024, and 100 ground sample points were collected for verification. At the same time, the forest stand subcompartment division data formed by the national forest resource second-class survey in 2021 was obtained, which can be used for precision verification and auxiliary labeling. Based on the above data, the tree species classification system of the test area was finally determined, and the target categories included 7 categories, namely larch, Chinese pine, scots pine, white birch, oak, bare land and background.

[0082] These remote sensing image data and ground data after preprocessing can be used as training data of deep learning neural network model.

[0083] Step S2: preprocessing the obtained data.

[0084] In the embodiment of the present application, different preprocessing methods are adopted according to the data type difference for the obtained original remote sensing image data and ground data:

[0085] For airborne and satellite data, the preprocessing procedure is as follows:

[0086] Radiation correction: for the panchromatic and multispectral images of GF-7 satellite, the absolute radiation correction method is adopted, and the correction coefficient is selected according to the sensor model to convert the DN value into apparent reflectivity;

[0087] Atmospheric correction: for GF-7 multispectral image, FLAASH model is used to eliminate atmospheric influence;

[0088] Orthorectification: The panchromatic and multispectral images are geometrically corrected based on the 0.5m resolution DEM and RPC parameters obtained by LiDAR, and the control point accuracy is better than 0.5 pixels.

[0089] Panchromatic fusion: The panchromatic image is fused by the Gram-Schmidt sharpening algorithm, which improves the spatial resolution to 0.5m while maintaining the spectral characteristics.

[0090] The CHM canopy height model data and optical images are registered, and the registration process includes:

[0091] Image registration: The GF-7 image and CHM data are unified to a 0.5m resolution grid using the nearest neighbor resampling method.

[0092] Noise suppression: A 7x7 window and a local adaptive Frost filter with a K coefficient of 2.0 are applied to the GF-7 image and CHM data to suppress abnormal high value noise.

[0093] Edge preservation: The edge preservation parameter a is set to 0.4 during filtering to preserve the forest boundary features.

[0094] Data normalization: The CHM height values and GF-7 multispectral values are linearly normalized to the range of 0-1.

[0095] Through the above preprocessing, the multi-source data fused with multispectral and height information is generated, providing a basis for subsequent analysis.

[0096] Step S3: According to the preprocessed data, deep learning training samples are prepared.

[0097] Figure 2 The preparation process of the deep learning training samples and the division method of the data set in the embodiment of the application are shown in the schematic diagram.

[0098] Firstly, using ArcMap software, a representative area in the test area is selected, and in the laboratory, 0.1m resolution aerial images are used to visually interpret and draw the initial draft of different tree species forest boundary.

[0099] Then, the RTK is used to collect forest and tree species sample points in the field, record the tree species type and patch boundary coordinates, verify and correct the forest vector map patch, and obtain high-quality forest boundary vector map. It is rasterized to the pixel grid of GF-7 image (grid length is 0.5m), and the tree species classification label map is obtained.

[0100] Meanwhile, in order to adapt to the input requirements of the deep learning model, the labels and corresponding images are cropped into small slices. In the embodiment of the present application, the size of the small slice image picture is 256x256 pixels, and 14,159 labeled image samples are made for model training and evaluation.

[0101] It can be understood that, in order to enhance the training effect and generalization ability of the model, the training sample data is respectively normalized and processed by feature engineering in the embodiment of the present application.

[0102] In terms of normalization processing, firstly, the mean and standard deviation of each feature dimension are calculated based on the entire data set, and then linear transformation is performed on each data point to map the original data to a standard normal distribution space with a mean of 0 and a standard deviation of 1, which can eliminate the dimensional differences between different features and preserve the distribution form of the data, so that the model training based on gradient descent converges more efficiently and avoids optimization bias caused by feature scale differences.

[0103] In terms of feature engineering processing, for the spatial structure and semantic information of image data, texture features, shape features and color histograms can be extracted to enhance the representation ability of the data and provide more discriminative inputs for the model. In addition, by combining the advantages of deep learning automatic feature extraction, high-level semantic features of the image are abstracted layer by layer through convolutional neural networks and other structures, which can reduce the subjectivity of manual feature design and further improve the understanding and generalization ability of the model for image scenes.

[0104] In addition, in order to increase the diversity of training data, avoid overfitting and enhance the generalization ability of the model, data augmentation techniques are used to expand the training data set. In the embodiment of the present application, the data augmentation techniques used include:

[0105] Geometric enhancement: random horizontal / vertical flipping (probability of 0.5 each), random rotation (rotation amplitude of ±15°, probability of 0.3).

[0106] Spectral enhancement: random brightness adjustment (adjustment amplitude of ±10%, probability of 0.3), random contrast adjustment (adjustment amplitude of ±10%, probability of 0.3).

[0107] Spatial enhancement: random cropping, the cropped picture size is 256x256 and more than 80% of the effective pixels are retained.

[0108] Finally, the prepared sample data is divided into training set, validation set and test set according to a certain proportion to ensure the fairness and reliability of the model training process and precision evaluation. In the embodiment of the present application, the proportions of the training set, the validation set and the test set are 6:2:2.

[0109] Step S4: constructing a double-path CBAM-UNet network model.

[0110] In terms of neural network design, the improved double-path CBAM-UNet deep convolutional neural network architecture is adopted in the embodiment of the present application, which can accurately identify different tree species by using multispectral and canopy height data.

[0111] Figure 3 The overall structure schematic diagram of the double-path CBAM-UNet network model of the present application is shown. As can be seen from the figure, the neural network includes an input layer, an encoder, a feature fusion layer, a bottleneck layer, a decoder and an output layer, and the modules of the neural network will be described in detail below in combination with the schematic diagram.

[0112] Input layer: The input layer provides double-path input channels, including a left-path input channel and a right-path input channel. The left-path input channel is used to receive GF-7 multispectral data, and the right-path input channel is used to receive CHM canopy height data.

[0113] In the embodiment of the present application, the channel number of the GF-7 multispectral data is 4, and the size is 256*256 pixels; the channel number of the CHM canopy height data is 1, and the size is also 256*256 pixels.

[0114] Encoder: The encoder includes parallel left-path and right-path encoders, which respectively extract features from the feature data input by the left-path and right-path input channels. The left-path and right-path encoders are both multi-level feature extraction layers based on down-sampling and BasicConv modules, which utilize a multi-level structure to achieve multi-layer step-by-step feature extraction. In the embodiment of the present application, the left-path and right-path encoders are both four-level structures. The input layer simultaneously inputs the original image data to the left-path and right-path encoders for parallel processing.

[0115] The left-path encoder is used to process the optical remote sensing image data (including GF-7 multispectral image data) input by the left-path input channel, and the optical remote sensing image data includes RGB bands or multispectral data, and the network structure is as follows:

[0116] First-level feature extraction layer: including BasicConv module. In the embodiment of the present application, the BasicConv module outputs a 96-channel feature map with a size of 256*256 pixels.

[0117] The second-level feature extraction layer includes a max-pooling layer for down-sampling and a BasicConv module, and after max-pooling, Dropout processing can be performed. In the embodiment of the application, the size of the pooling kernel is 2*2, and a 128-channel feature map with a size of 128*128 pixels is output by the BasicConv module.

[0118] The third-level feature extraction layer includes a max-pooling layer for down-sampling and a BasicConv module, and after max-pooling, Dropout processing can be performed. In the embodiment of the application, the size of the pooling kernel is 2*2, and a 256-channel feature map with a size of 64*64 pixels is output by the BasicConv module.

[0119] The fourth-level feature extraction layer includes a max-pooling layer for down-sampling and a BasicConv module, and after max-pooling, Dropout processing can be performed. In the embodiment of the application, the size of the pooling kernel is 2*2, and a 512-channel feature map with a size of 32*32 pixels is output by the BasicConv module.

[0120] The right path encoder is used for processing the CHM canopy height data input in the right path input channel, and the CHM canopy height model data provide forest height information, and the network structure is as follows:

[0121] The first-level feature extraction layer includes a BasicConv module. In the embodiment of the application, a 64-channel feature map with a size of 256*256 is output by the BasicConv module.

[0122] The second-level feature extraction layer includes a max-pooling layer for down-sampling and a BasicConv module, and after max-pooling, Dropout processing can be performed. In the embodiment of the application, a 128-channel feature map with a size of 128*128 pixels is output by the BasicConv module.

[0123] The third-level feature extraction layer includes a max-pooling layer for down-sampling and a BasicConv module, and after max-pooling, Dropout processing can be performed. In the embodiment of the application, a 256-channel feature map with a size of 64*64 pixels is output by the BasicConv module.

[0124] The fourth-level feature extraction layer includes a max-pooling layer for down-sampling and a BasicConv module, and after max-pooling, Dropout processing can be performed. In the embodiment of the application, a 512-channel feature map with a size of 32*32 pixels is output by the BasicConv module.

[0125] In the embodiment of the present application, the feature extraction and fusion of multi-source heterogeneous data are realized through double-path multi-level parallel feature extraction, including complementary fusion of spectral features and structural features.

[0126] The Dropout processing in each feature extraction layer of the left path encoder and the right path encoder is an optional operation, and Dropout can randomly discard neurons to prevent overfitting of the neural network and improve the generalization ability of the model. In the embodiment of the present application, the Dropout Rate is 0.3.

[0127] In the embodiment of the present application, the BasicConv module used in each feature extraction layer of the left path encoder and the right path encoder is a basic convolutional layer for feature extraction. Figure 4 The network structure diagram of the BasicConv module is shown. As can be seen from the diagram, the BasicConv module of the present application has two working modes, Figure 4 The left side shows the network structure diagram of the BasicConv module in the standard mode; Figure 4 The right side shows the network structure diagram of the BasicConv module in the ASPP enhanced mode.

[0128] As can be seen from the left side, Figure 4 As can be seen from the left side, the BasicConv module in the standard mode includes two convolutional layers connected in series, each followed by a BatchNorm and a ReLU activation function. In order to prevent overfitting of the model, a Dropout layer can be connected after each layer for regularization. At this time, the input features are only processed by the basic convolution unit and then output. In the embodiment of the present application, the convolution kernel size of the two convolutional layers is 3x3, and the Dropout Rate is 0.5.

[0129] As can be seen from the right side, Figure 4 As can be seen from the right side, the BasicConv module in the ASPP enhanced mode integrates the ASPP module in the network structure. At this time, the input features are processed in parallel by the basic convolution unit and the ASPP sub-module, the processing results of the two paths are fused by element-wise addition, and the fused features are processed by batch normalization and ReLU activation function, and then output after applying Dropout regularization. In this way, the BasicConv module has better multi-scale feature extraction capability.

[0130] In the embodiment of the present application, the dual-path CBAM-UNet network model has a preset aspp mode switch, which is used to determine the working mode of the BasicConv module. In a specific application scenario, when the model needs to balance the calculation efficiency and the basic feature extraction capability (such as processing small size targets or low complexity scenes), the aspp mode switch parameter can be set to the standard mode; and when the task requirement emphasizes more on multi-scale context information capture (such as large size target segmentation or complex scene classification), the aspp mode switch parameter can be set to the ASPP enhanced mode to improve the feature expression capability of the model.

[0131] Figure 5 The network structure diagram of the ASPP module is shown, and as can be seen from the diagram, the input layer of the ASPP module receives a feature map as the input data processed by the module. After the input layer, there are five parallel branches, and the five specific branches include:

[0132] The first convolution branch: in the embodiment of the present application, this branch uses a 1x1 convolution kernel to perform convolution operation on the input feature map, which is used to extract local detail features and capture the correlation between feature map channels.

[0133] The second dilated convolution branch: in the embodiment of the present application, this branch uses a 3x3 convolution kernel and the dilated rate of dilated convolution is 6, which is used to expand the receptive field and capture medium scale features.

[0134] The third dilated convolution branch: in the embodiment of the present application, this branch uses a 3x3 convolution kernel and the dilated rate of dilated convolution is 12, which is used to further expand the receptive field and obtain larger scale features.

[0135] The fourth dilated convolution branch: in the embodiment of the present application, this branch uses a 3x3 convolution kernel and the dilated rate of dilated convolution is 18, which is used to make the model have a larger receptive field and capture global scale features.

[0136] The fifth pooling and convolution branch: first, a pooling layer with a 1x1 pooling kernel is used to perform global average pooling operation on the feature map; then, a 1x1 convolution is performed; then, BatchNorm and ReLU activation function processing are performed; finally, the feature map size is adjusted by Bilinear Interpolate.

[0137] After the feature map is processed by the five parallel branches respectively, the fusion layer is used to perform concatenation (Concat) operation on the feature maps processed by the above parallel branches in the channel dimension, so as to fuse multi-scale feature information. Then, a 1x1 convolution kernel is used to perform convolution operation on the concatenated feature map, and the channel number of the feature map is adjusted.

[0138] Feature fusion layer: the dual-path CBAM-UNet network of the application has a plurality of feature fusion layers, which provide a multi-scale feature fusion mechanism, and implement a progressive cascaded feature fusion strategy at each encoding level to fuse the left and right path features.

[0139] It can be understood that, in the multi-level feature fusion process, in order to meet the conditions of feature fusion, the feature map sizes of the left path and the right path need to be consistent. At the same time, the feature maps are stacked in a way to increase the number of channels of the feature maps so as to enrich the expression of the features and avoid information loss. The specific process of multi-level feature fusion is as follows:

[0140] First-level fusion: cascade the left path features and the right path features to form an enhanced representation of the features. In the embodiment of the application, the channel number of the feature map input by the left path is 96, the channel number of the feature data input by the right path is 64, and the size of the feature map is 256x256 pixels. After fusion, a 256x256 feature map with 160 channels is obtained.

[0141] Second-level fusion: cascade the left path features and the right path features to form an enhanced representation of the features. In the embodiment of the application, the channel number of the feature map input by the left path and the right path is 128, and the size of the feature map is 128x128 pixels. After fusion, a 128x128 feature map with 256 channels is obtained.

[0142] Third-level fusion: cascade the left path features and the right path features to form an enhanced representation of the features. In the embodiment of the application, the channel number of the feature map input by the left path and the right path is 256, and the size of the feature map is 64x64 pixels. After fusion, a 64x64 feature map with 512 channels is obtained.

[0143] Fourth-level fusion: cascade the left path features and the right path features to form an enhanced representation of the features. In the embodiment of the application, the channel number of the feature map input by the left path and the right path is 512, and the size of the feature map is 32x32 pixels. After fusion, a 32x32 feature map with 1024 channels is obtained.

[0144] Bottleneck layer: the dual-path CBAM-UNet network model of the application includes a bottleneck layer, which is a key component connecting the encoder and the decoder. Through the bottleneck layer, the fused high-level features can be further enhanced. In the embodiment of the application, the bottleneck layer includes the following modules:

[0145] DoubleConv module: including a left DoubleConv module and a right DoubleConv module. The left and right DoubleConv modules process the feature maps output by the left and right encoders respectively. In the embodiment of the application, as known from the above-described encoder, the left and right encoders output 512-channel feature maps with a size of 32x32 pixels.

[0146] Bottleneck fusion module: used to concatenate the features of the left and right paths, generating a 1024-channel feature map with a size of 32x32 pixels.

[0147] Terminal enhancement module: converts the multi-channel feature map output by the bottleneck fusion module into a highly nonlinear representation through an additional DoubleConv module.

[0148] CBAM enhancement module: applies an attention mechanism to the output of the bottleneck layer to highlight key features. It can be understood that the output of the bottleneck layer here is the output of the terminal enhancement module.

[0149] The bottleneck layer effectively solves the problems of feature redundancy and boundary ambiguity in multi-source data integration through the implementation of multiple strategies such as cross-modal fusion, attention enhancement, and parameter compression. This makes the model more efficient while greatly improving the boundary segmentation accuracy compared to traditional UNet network models.

[0150] Decoder: the dual-path CBAM-UNet network model of the present application includes a decoder, which is used to restore the spatial resolution of the feature map. In the embodiments of the present application, the decoder also has a four-layer structure, and unlike the encoder, each layer of the decoder includes a transposed convolution module for upsampling and a DoubleConv module for processing fused features. In the embodiments of the present application, the convolution kernel size of the transposed convolution module is 2x2.

[0151] Meanwhile, the enhanced features obtained by the CBAM module enhancement of each layer of the left and right path encoders are connected to the corresponding layers of the decoder through a skip connection. The upsampled features and the enhanced features corresponding to the skip connection are merged through a concatenation operation. During the decoding process of the decoder, the number of channels of the feature map decreases (512→256→128→64) while the size of the feature map increases (32→64→128→256).

[0152] Output layer: the dual-path CBAM-UNet network model of the present application also includes an output layer, which maps the output features of the decoder to the predefined classes of forest tree species through a convolution layer with a convolution kernel size of 1x1, specifically including 7 predefined classes. Finally, the probability distribution of each class is generated by applying a Softmax activation function. Here, the size of the output feature map of the output layer is 256x256 pixels.

[0153] In conclusion, the double-path CBAM-UNet network of the present application contains two parallel encoder paths, which respectively process spectral information (from GF-7 four-channel multispectral data) and height information (from single-channel CHM data), through the "separation-fusion" mechanism, the neural network can independently extract features of different data types, avoid mutual interference, and perform feature fusion at four resolution levels, fully exerting the complementarity of spectral and height information. Through multi-level feature extraction and feature fusion, high-precision recognition of different tree species is realized.

[0154] Firstly, the network architecture combines CBAM attention module and ASPP multi-scale feature extraction technology. CBAM enhances the key features of regions such as tree species boundaries through dual attention mechanisms of channels and spaces. At the same time, the ASPP module uses parallel convolution with multiple dilation rates (6, 12, 18), which can capture tree crown features of different scales, thereby improving the recognition ability of different scale tree species.

[0155] Secondly, the network is optimized in the design of skip connection. On the basis of the traditional U-Net network model, the double-path feature fusion after CBAM enhancement is connected with the decoder, thereby improving the recognition ability of the network model in high canopy density and mixed forest area boundary fuzzy area.

[0156] Thirdly, through path differentiation design and parameter sharing, the total parameter quantity of the model is controlled within 126 million, effectively improving the training and application efficiency.

[0157] In the embodiment of the present application, the model input is 4 multispectral bands of GF-7 and 1 CHM height band, and the output is the tree species category corresponding to the image element. The classification categories cover the main dominant tree species (such as larch, Chinese pine, Korean pine, white birch, oak, etc.) and other secondary tree species or non-forest land (marked as "other" category), a total of 7 categories.

[0158] As shown in Figure 6 The CBAM module (convolutional block attention mechanism) of the present application contains channel attention mechanism and spatial attention mechanism. Among them, the channel attention is used to process the maximum pooling and average pooling features in parallel and apply shared MLP to extract channel attention weights; the spatial attention uses the maximum and average values of the channel dimension to generate a feature map, and applies a 7x7 convolution to generate a spatial attention map. In the embodiment of the present application, the CBAM module is applied after the bottleneck layer and at each skip connection, which can significantly enhance the important feature representation.

[0159] The DoubleConv module of the present application realizes feature layer-by-layer abstraction, non-linear modeling and channel adaptation through the combination of "two convolutions + non-linear activation + normalization", and is a basic component for improving feature expression capability in the double-path CBAM-UNet model. In the embodiment of the present application, the module is combined with ASPP, CBAM and the like, which significantly improves the distinguishing ability of similar tree species in complex forest stands (such as the classification accuracy of larch and oil pine is improved by about 5%), and optimizes the delicacy of boundary extraction. The DoubleConv module includes two 3x3 convolution layers, each of which is followed by a batch normalization (BatchNorm) and a ReLU activation function. This design can extract more complex features while keeping the spatial size of the feature map unchanged (by setting padding = 1) by stacking convolution layers. In specific implementation, the input and output channel numbers of DoubleConv can be adjusted according to network requirements, and the output channel number of the second convolution layer is usually consistent with the output channel number of the module.

[0160] In the embodiment of the present application, DoubleConv can enhance the feature expression capability of the network. Through two convolution operations, the network can learn more abstract features, and the introduction of batch normalization (BatchNorm) and ReLU activation function helps to speed up training and alleviate the gradient vanishing problem. This modular design makes DoubleConv can be flexibly embedded into the encoder and decoder of UNet, becoming an important part of the network structure.

[0161] Step S5: training the double-path CBAM-UNet network model using the preprocessed data.

[0162] Figure 7 The specific process of training the double-path CBAM-UNet network model of the present application is shown, which specifically includes the following steps:

[0163] Step S51: model initialization configuration.

[0164] The double-path CBAM-UNet network model constructed in the embodiment of the present application aims to realize effective semantic segmentation of multi-channel remote sensing images. The model is designed as a double-path structure, in which the left path inputs 4-channel multispectral image data, the right path inputs 1-channel CHM image data, and the overall input is 5-channel data. The final output of the model is an 8-channel classification map, corresponding to different tree species categories and background areas.

[0165] In the model training initialization stage, the optimizer and learning rate scheduling strategy are configured. Specifically, the AdamW optimizer is used for parameter update, the momentum parameters β1 is set to 0.9, β2 is set to 0.999, and weight decay = 0.01 is set to prevent overfitting. In terms of learning rate, the initial learning rate is set to 0.001, the minimum learning rate limit is 1e-6, and the cosine annealing scheduling algorithm (CosineAnnealingLR) is used to dynamically adjust the learning rate, thereby accelerating the model convergence speed and improving the training stability.

[0166] In order to improve the adaptability of the model to the unbalanced class distribution, the loss function design adopts the combination of main loss and auxiliary loss:

[0167] Main loss function: the weighted cross-entropy loss function is used, and the weight of each class is calculated according to the inverse ratio of its sample number to alleviate the class imbalance problem.

[0168] Auxiliary loss function: combine Dice loss and cross-entropy loss, weighted combination according to the proportion of 0.4:0.6, enhance the model's ability to identify the boundary of the target region.

[0169] Specific class balance enhancement mechanism: for the target class with fewer samples (such as oak trees), introduce FocalLoss (γ=2.0) to further improve the model's recognition effect on minority classes.

[0170] The organic combination of the above model structure and training strategy makes the present application have stronger expression ability and higher classification accuracy in complex vegetation classification tasks.

[0171] Step S52: use the prepared training sample data to train the model.

[0172] During the model training process, the data loader loads the training data and validation data with appropriate batch size. Among them, the training data is randomly shuffled to improve the generalization ability of the model, while the validation data maintains a fixed order to stably evaluate the model performance, and the data set size is recorded to provide a basis for subsequent performance index calculation. In the embodiment of the present application, the batch size in the training stage is set to 16, and the prediction stage is 30.

[0173] It can be understood that model training needs multiple cycles to be continuously optimized, and each cycle is mainly divided into two stages of training and verification. In the training stage, the model parameters are adjusted through the loss function to enable the model to learn the features in the data; in the verification stage, the model performance is evaluated by using the verification data that does not participate in the training to monitor the generalization ability of the model. In order to improve the training efficiency and stability, the mixed precision training strategy and the gradient clipping method are adopted in the embodiment of the application, and the dropout is applied to enhance the generalization ability of the model, and the dropout rate can be set according to actual needs, and the application does not make special limitation.

[0174] In terms of model performance evaluation, the model performance is comprehensively measured by calculating the loss, pixel-level accuracy and intersection over union (IoU, based on the overlap ratio of prediction and true label), and the model weight with the highest accuracy on the verification set is saved. In addition, the training progress will be recorded during the training process, and the computing resources will be regularly cleaned to optimize the memory usage and ensure the smooth progress of the training process. In the embodiment of the application, the total number of training cycles is set to 100 epochs, and if the verification set accuracy does not improve for 10 consecutive epochs during the training process, the training will be terminated in advance to avoid waste of resources.

[0175] Step S53: model evaluation and optimization.

[0176] In the model training stage, the training set images are input into the double-path CBAM-UNet network for iterative training, and the verification set is used to evaluate the model performance and adjust the hyperparameters to monitor the performance of the model. Specifically, the loss value, accuracy and intersection over union (IoU) are output for each epoch to evaluate the classification and segmentation ability of the model. At the same time, the time consumption of the entire training process is recorded to analyze the efficiency and effect of the model.

[0177] Once the evaluation result of the model reaches the expected standard, we will end the training and export the model; if it does not reach the expected, the parameters will be adjusted according to the evaluation feedback and the training will continue. In the training stage, the training set images are input into the network model for iterative training, and the verification set is used to guide the adjustment of the hyperparameters.

[0178] In the embodiment of the application, after multiple rounds of training and parameter optimization, the model achieves an accuracy of 87.66% on the training set and an accuracy of 80.07% on the verification set, which indicates that the model not only has good convergence, but also exhibits strong generalization ability. Finally, the best-performing model is selected for the test set, achieving an overall classification accuracy of 81.83%, an average intersection over union (mIoU) of 70.39%, and a Kappa coefficient of 0.7803, which proves the effectiveness of the model.

[0179] When the model training meets the expected performance, the model is exported and deployed to the target device for subsequent tree species classification tasks.

[0180] Step S6: Tree species classification prediction is performed through the deployed double-path CBAM-UNet network model, and a tree species classification map is obtained.

[0181] In the embodiment of the present application, the trained double-path CBAM-UNet model is applied to the whole remote sensing image of the Magnolia test area to realize pixel-by-pixel tree species classification prediction in a large area. Due to the large area of the research area, the GF-7 and CHM images can be cut into continuous and non-overlapping small image blocks of 256x256 pixels in the implementation process, and the double-path CBAM-UNet model is input block by block for forward prediction to obtain the corresponding small slice classification results.

[0182] As shown in Figure 8 , tree species classification prediction and obtaining a tree species classification map specifically includes the following steps:

[0183] Step S61: Data preparation.

[0184] Collect the GF-7 and CHM images of the test area, divide the collected image data of the test area into several continuous and non-overlapping image slices, and then pre-process the sliced images. The pre-processing process has been described in detail in step S2, which will not be repeated here. At the same time, the pre-processed image data needs to be sliced, which is the same as the processing method in the training sample. The size of the image slice is also 256x256 pixels.

[0185] Step S62: Load the trained model.

[0186] Read the weight file (such as.pth or.h5 format) with the highest validation set accuracy from the model save path, and load the weight parameters into the double-path CBAM-UNet network model through the deep learning framework (such as PyTorch / TensorFlow). To improve the prediction efficiency, deploy the model to a computing device equipped with a high-performance GPU such as NVIDIA RTX4090, and use CUDA to accelerate parallel computing.

[0187] After loading, set the model to evaluation mode through the model.eval() interface, and disable gradient calculation (such as using the torch.no_grad() context manager in PyTorch) to avoid unnecessary parameter update operations, reduce memory occupation, and speed up the forward propagation process.

[0188] Step S63: Slice prediction.

[0189] The pre-processed image slices are loaded into the model input queue in batches, and the batch size is set to 30 in this embodiment to balance GPU memory occupation and calculation efficiency. The model performs forward propagation on each slice, outputting a probability distribution tensor with dimensions of 256x256x7 (7 channels corresponding to the predicted probabilities of 6 target tree species and background). By applying the argmax maximum value operation, the maximum probability class index of each pixel is extracted along the channel dimension, and the continuous probability distribution is converted into a discrete pixel-level tree species classification map.

[0190] Finally, the predicted tree species classification map is saved according to the original slice position information for subsequent stitching processing.

[0191] Step S7: Image stitching of tree species classification map to obtain tree species classification raster map.

[0192] The prediction results of all slices are stitched and restored according to their geographical positions in the original image. During the stitching process, the results of adjacent blocks are smoothed to eliminate stitching marks and ensure the continuity and consistency of the classification map. Finally, a complete test area tree species distribution classification map is generated, and the result map is stored in raster format, with each pixel assigned a corresponding tree species category code. The specific process of image stitching includes the following steps:

[0193] Step S71: Geographical coordinate mapping and positioning.

[0194] When performing image stitching of the tree species classification map, first, the spatial reference information recorded during the slicing process, such as latitude and longitude range, row and column number, etc., is used to establish the position correspondence between each predicted slice and the original image. At the same time, a global grid index is constructed to determine the starting and ending coordinates of each slice in the complete grid map, and the overlapping area is marked to provide clear boundary information for subsequent fusion processing. This step is the basis for ensuring the accuracy of the stitching position.

[0195] Step S72: Feature fusion in overlapping area.

[0196] When processing feature fusion in the overlapping area, there is usually a certain slice size overlap between adjacent slices. For this part of the area, a weighted average strategy can be used to achieve smooth transition. The weight is calculated by a spatial distance decay function, for example, the weight is equal to the product of the negative alpha of the natural exponential function and the square of the distance, where alpha is the adjustment parameter and distance refers to the distance from the pixel to the slice boundary. For pixels with inconsistent classification, the prediction result with higher confidence is preferred, such as the maximum probability value of the Softmax output, which makes the classification result of the overlapping area more reasonable.

[0197] Step S73: Boundary feathering and gradient smoothing.

[0198] In the process of boundary feathering and gradient smoothing operation, a Gaussian blur filter is applied to the edge of the slice, generally with sigma set to 2-3 pixels, so as to generate a transition zone with a width of 5-10 pixels. For areas with sharp gradient changes, such as forest canopy edges, morphological dilation or erosion operations are introduced to optimize the continuity of the boundary. Moreover, the terrain information in the CHM data is used to perform elevation-weighted fusion in areas with large terrain undulations, thereby reducing the deviation caused by mountain shadows on the classification and making the boundaries of the spliced images more natural and continuous.

[0199] Step S74: raster encoding and metadata integration.

[0200] Finally, raster encoding and metadata integration need to be completed, the fused classification result is converted into a standard GeoTIFF raster format, so that each pixel value corresponds to a specific tree species code, such as 1 representing larch and 2 representing oil pine. At the same time, spatial reference system information is embedded, including projection method and geographic coordinate system, etc., and a classification attribute table is also attached, recording the tree species name, confidence threshold and other metadata corresponding to each code, finally generating a tree species distribution raster map that is spatially continuous, has consistent attributes and has complete geographic positioning information.

[0201] Finally, the generated tree species distribution raster map is exported and saved, and relevant metadata information is attached, facilitating subsequent application and analysis.

[0202] Step S8: vectorization processing and vector boundary optimization are performed on the tree species classification raster map to extract forest stand boundaries and generate a forest stand boundary vector map. Specifically, it includes:

[0203] Step S81: vectorize the tree species classification raster map using GIS software, and convert the boundaries of continuous areas of the same tree species into forest stand boundaries through spatial clustering of tree species categories.

[0204] In the embodiments of the present application, the tree species classification raster map can be vectorized using GIS software. The GIS software is a computer system tool for collecting, storing, managing, analyzing and visualizing geographic spatial data. Common GIS software includes ArcGIS series, QGIS, SuperMap, MapInfo, etc. When used, the selection can be made according to the needs, and the present application does not make special limitations.

[0205] Step S82: map the attribute values of the vectorized forest stand polygons with the tree species categories to establish a forest stand polygon attribute table.

[0206] In the embodiments of the present application, there are 7 tree species categories, which are mapped with the attribute values of the vectorized forest stand polygons to establish a forest stand polygon tree species attribute table.

[0207] Step S83: topological relationship processing is performed to ensure that adjacent polygons share boundaries.

[0208] Ensure that adjacent polygons share boundaries to avoid topological errors.

[0209] Step S84: Apply geometric simplification processing to optimize the stand vector boundary.

[0210] Apply geometric simplification processing to optimize the stand vector boundary. In the embodiment of the present application, the tolerance parameter of the Douglas-Peucker algorithm is set to 0.5 m.

[0211] Step S85: Filter small areas for the stand vector polygon, and merge the polygons smaller than the threshold into the surrounding main stand.

[0212] Set a minimum area threshold to merge the noise of the tree species classification raster map into the background. In the embodiment of the present application, the minimum area threshold is 0.5 ha (5000 m 2 ).

[0213] Step S86: Smooth and optimize the stand vector boundary using a boundary smoothing algorithm to make the boundary closer to the natural forest edge.

[0214] In the embodiment of the present application, the PAEK smoothing algorithm is applied to smooth the tree species classification raster map to make its boundary closer to the natural forest edge. In the embodiment of the present application, the tolerance of the PAEK smoothing algorithm is 2 m.

[0215] The PAEK (Parametric Exponential Kernel) smoothing algorithm is an algorithm for smoothing line features in GIS. It reduces data noise or detail fluctuations through mathematical calculations to generate a more concise curve, and adjusts the coordinates of the polygon vertices to reduce the jagged fluctuations of the boundary, making the line or contour closer to the natural form.

[0216] After the above vectorization processing and vector boundary optimization operation, a high-precision stand boundary vector map is finally generated, realizing the automatic delineation of the Mulan test area stand.

[0217] Finally, the high-precision stand boundary vector map generated by the model is stored in the corresponding forest resource management database, thus completing the work of stand species recognition and boundary extraction.

[0218] As for the utilization of the obtained forest stand boundary vector map, professional GIS software (such as ArcGIS, QGIS), online map platforms (such as Google Earth Pro, Mapbox, Cesium), visualization tools (such as Tableau, Power BI, DataV) and the like can be used to display the forest stand boundary vector map to relevant personnel. In combination with three-dimensional and dynamic effects, for example, a digital elevation model (DEM) can be loaded in Cesium or ArcGIS Earth, and the two-dimensional boundary can be superimposed on the three-dimensional terrain to intuitively display the relationship between the boundary and the terrain (for example, the boundary in a valley may need to be adjusted regularly due to water erosion). The boundary changes (such as boundary shrinkage due to logging in the past 10 years) can be displayed through a time axis function, or future scenarios (such as the impact of tree species distribution migration under climate change on the boundary) can be simulated, and the like.

[0219] In actual application, the forest stand species identification and boundary extraction method of the present application obtains high-precision tree species classification maps and forest stand boundary maps in Mulan test area, and is compared with existing forest resource subplots. Figure 9 The comparison results are shown, in which the red lines represent the forest stand boundaries automatically delineated by the method, and the background is the high-resolution remote sensing image base map (true color synthesis) of the study area. At the same time, the blue-green lines in the figure superimpose the subplots boundaries of the national forest resources second-class survey as a reference. It can be seen that the automatically extracted forest stand boundaries subdivide some large forest stands into multiple more homogeneous small patches, and more finely depict the differences in tree species composition within the forest stand. Compared with the traditional survey subplots, the automatically delineated forest stand units are more consistent with the actual distribution of forests, and the spatial accuracy is significantly improved.

[0220] As shown in Figure 10 The forest stand boundary extraction system 600 in the present application includes a graph data acquisition module 601, a data preprocessing module 602, a prediction module 603, an image processing module 604, and a boundary extraction module 605.

[0221] The data acquisition module 601 is responsible for collecting forest remote sensing image data (including airborne digital images, satellite images, CHM canopy height model data) and ground data (RTK collected tree species sample points, national forest resources second-class survey subplot data), among which the GF-7 satellite multispectral image and CHM data need to be acquired during the vegetation growth period and the laser radar acquisition period, and the ground data is used for model training and verification to ensure that the data covers the representative areas of the test area.

[0222] The data preprocessing module 602 is responsible for radiometric correction, atmospheric correction, orthorectification and panchromatic fusion of remote sensing images, so that the resolution of the GF-7 image is improved to 0.5 meters; the CHM data is subjected to noise suppression, edge preservation and normalization processing, and is registered with the optical image to unify the spatial coordinate system, and finally a multi-source data fused with spectral and height information is generated, which is prepared for model input.

[0223] The prediction module 603 is constructed based on a double-path CBAM-UNet network model, the left path processes GF-7 multispectral data to extract spectral features, the right path processes CHM data to extract structural features, features are extracted in parallel through a four-level encoder, key features are enhanced by using a CBAM module, multi-scale feature capture is optimized by combining an ASPP mode switch, spatial resolution is restored by a decoder after fusion through a bottleneck layer, and finally a tree species classification probability distribution is output.

[0224] In operation, the prediction module 603 slices the preprocessed image into 256x256 pixels, loads the trained model weight to the GPU, performs forward propagation with a batch size of 30, converts the output probability tensor into a pixel-level classification map through argmax, records the geographic coordinates when saving the slice results, and provides location information for subsequent splicing.

[0225] The image processing module 604 performs geographic coordinate mapping based on the slice spatial reference information, fuses the overlapping areas by using distance weighted average and confidence priority strategy, smoothes the boundary through Gaussian blur and morphological operation, finally converts the classification result into GeoTIFF format, embeds the spatial reference system and integrates the metadata, and generates a complete tree species classification raster map.

[0226] The boundary extraction module 605 is used for vectorization processing and vector boundary optimization of the tree species classification raster map to extract the forest boundary and generate a forest boundary vector map.

[0227] In summary, the forest boundary extraction method and system has the following advantages:

[0228] Firstly, multi-scale feature capture: using GF-7 multispectral images and high-resolution CHM data, combined with a deep learning model with attention mechanism and a hollow pyramid pooling, multi-scale feature information such as spectral, spatial texture and vertical height of the forest stand can be obtained at the same time. This enables the model to identify subtle differences between different tree species, even in tree species with similar spectral characteristics (such as Pinus sylvestris and Pinus tabulaeformis), which can be distinguished by crown height or texture, thereby improving the classification accuracy and further improving the accuracy of forest boundary extraction.

[0229] Second, complex forest adaptability: The method achieved more than 81% classification accuracy in the complex and large-scale terrain Mulan test area with mixed coniferous and broad-leaved forests, indicating its good adaptability to complex and heterogeneous forest. The results show that whether it is a large-area contiguous artificial pure forest or a natural mixed-age forest, this method can achieve stable tree species recognition and forest boundary extraction.

[0230] Third, automated and efficient process: From data preprocessing, model training to result generation, a complete automated processing flow is formed, greatly reducing manual intervention and field work. The model automatically outputs tree species classification map and forest boundary through remote sensing data, without the time-consuming manual sketching. Especially in the forest boundary extraction stage, the vectorization and algorithm optimization are used to automatically complete the small plot division, improving the production efficiency and data processing consistency. Compared with manual investigation, this method can complete the mapping update of large-scale forest in a short time, which helps to obtain timely forest resource change information.

[0231] Fourth, precision improvement: Thanks to the feature fusion of multi-source data and the improved model algorithm, the method of the present application has obvious improvement in tree species classification accuracy and forest boundary refinement compared with traditional methods. The automatically delineated forest boundary matches well with the actual forest patch, and can distinguish the heterogeneous forest which is not subdivided in traditional two-class survey. The Kappa coefficient of the classification result can reach 0.78, indicating a high degree of conformity with the true ground situation. This high-precision data product can be used as high-quality basic data for forest resource management, providing strong support for forest management decision-making and ecological research.

[0232] Fifth, improved computing performance: In actual application test, the method of the present application runs on an NVIDIA RTX4090 system, uses the PyTorch 2.2 deep learning framework, and the complete processing time for a test area of 12,300 hectares is about 2 to 3 hours, which has excellent computing efficiency and economic benefits, and can support large-scale forest information automation processing tasks.

[0233] In summary, the forest boundary extraction method of the present application successfully realizes high-precision and automated acquisition of forest resource information based on high-resolution optical remote sensing and canopy height model. The method fully utilizes the advantages of multi-source remote sensing data and realizes accurate description of complex forest by combining deep learning, which has important practical significance for forest resource survey and monitoring.

[0234] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any changes or replacements within the scope of the disclosed technology can be easily thought of by those skilled in the art, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method of stand boundary extraction, characterized by, The forest stand boundary extraction method comprises the following steps: acquiring data comprising forest remote sensing image data and ground data; preprocessing the acquired data; constructing a double-path CBAM-UNet network model; the double-path CBAM-UNet network model comprises an input layer and an encoder, the input layer comprises a left path and a right path, the encoder comprises a left path encoder and a right path encoder, and the input layer is used to receive original image data and simultaneously input the original image data to the left path encoder and the right path encoder for parallel processing through the left path and the right path; wherein the left path encoder comprises a multi-level feature extraction layer constructed based on down-sampling and a BasicConv module; the right path encoder comprises a multi-level feature extraction layer constructed based on down-sampling and a BasicConv module; the double-path CBAM-UNet network model sets a mode switch to control the working mode of the BasicConv module, and the working mode comprises a standard mode and an ASPP enhanced mode; the BasicConv module in the standard mode comprises two convolution layers connected in series, and each layer is followed by a BatchNorm and a ReLU activation function; the left path encoder is used to process optical remote sensing image data, and the right path encoder is used to process CHM canopy height model data; the optical remote sensing image data comprises RGB bands or multispectral data, and the CHM canopy height model data provides tree height information; through double-path feature extraction, spectral features and structural features are complementarily fused; the double-path CBAM-UNet network model comprises a decoder, an output layer and a CBAM module; the decoder is a four-layer structure, each layer of which comprises a transposed convolution module for up-sampling and a DoubleConv module for processing fused features; the output layer comprises a convolution layer that maps the output features of the decoder to predefined tree species categories and applies a Softmax activation function to generate a probability distribution of each tree species category; the CBAM module is used to enhance the fused features after feature fusion of each feature extraction of the left path encoder and the right path encoder, and the enhanced features are fused with the output features of the transposed convolution module of the corresponding layer of the decoder through a skip connection, and the fused features are input into the DoubleConv module for processing; the double-path CBAM-UNet network model comprises a bottleneck layer, and the bottleneck layer comprises: a left path DoubleConv module for processing feature maps output by the left path encoder; a right path DoubleConv module for processing feature maps output by the right path encoder; a bottleneck fusion layer for concatenating the feature maps output by the left path DoubleConv module and the right path DoubleConv module in the channel dimension to generate concatenated feature maps; a terminal enhancement module for converting the concatenated feature maps into a highly nonlinear representation; The CBAM enhancement module enhances the output of the bottleneck layer by applying an attention mechanism; The pre-processed data is input into the double-path CBAM-UNet network model for tree species classification prediction and a tree species classification map is obtained; An image mosaic is performed on the tree species classification map to obtain a tree species classification raster map; Vectorization processing and vector boundary optimization are performed on the tree species classification raster map to extract the forest stand boundary and generate a forest stand boundary vector map.

2. The stand boundary extraction method according to claim 1, characterized by, The forest stand boundary extraction method comprises: preparing a deep learning training sample according to the obtained data to train the double-path CBAM-UNet network model.

3. The stand boundary extraction method according to claim 2, characterized by, The deep learning training sample is prepared according to the obtained data to train the double-path CBAM-UNet network model, which comprises: The training sample is prepared, including image slicing and label preparation, and the sample is normalized, feature engineered and data enhanced; The model is initialized and configured, including using the AdamW optimizer, the cosine annealing scheduling dynamic learning rate strategy and the weighted cross-entropy loss function as the main loss function; During the model training, the mixed precision training strategy and the gradient clipping method are used to improve the training efficiency of the model, and the random dropout strategy is applied to enhance the generalization ability of the model.

4. The stand boundary extraction method according to claim 1, characterized by, The double-path CBAM-UNet network model sets a mode switch to control the working mode of the BasicConv module, and the working mode includes a standard mode and an ASPP enhancement mode; When in the standard mode, the input features are output after being processed by the basic convolution unit; When in the ASPP enhancement mode, the input features are processed in parallel by the basic convolution unit and the ASPP sub-module, the two processing results are fused by element-wise addition, the fused features are processed by batch normalization and ReLU activation function, and are output after applying Dropout regularization.

5. The stand boundary extraction method according to claim 1, characterized by, The vectorization processing and vector boundary optimization of the tree species classification raster map comprise the following steps: The tree species classification raster map is vectorized using GIS software, and the boundaries of continuous regions of the same tree species are converted into forest stand boundaries through spatial clustering of tree species categories; The attribute values of the vectorized forest stand polygons are mapped with the tree species categories to establish a forest stand polygon attribute table; Topological relationship processing is performed to ensure that adjacent polygons share boundaries; Geometric simplification processing is applied to optimize the forest stand vector boundaries; Small area filtering is performed on the forest stand vector polygons, and the polygons smaller than the threshold value are merged into the surrounding main forest stands; A boundary smoothing algorithm is used to smooth and optimize the forest stand vector boundaries to make the boundaries closer to the natural forest edge.

6. A system for implementing the stand boundary extraction method of any one of claims 1 to 5, characterized in that, The method comprises: a data acquisition module for collecting forest remote sensing image data and ground data; a data preprocessing module for preprocessing the forest remote sensing image data and ground data collected by the data acquisition module; a prediction module for tree species classification prediction based on the input image slices and obtaining a tree species classification map; an image processing module for performing mosaic processing on the tree species classification map to obtain a complete tree species classification raster map; The boundary extraction module is used for vectorization processing and vector boundary optimization on the tree species classification raster map to extract the forest stand boundary and generate a forest stand boundary vector map.

Images