Method and system for identifying forest stand tree species based on multi-source remote sensing data fusion

Abstract

The invention discloses a forest stand tree species identification method and system based on multi-source remote sensing data fusion, and the method comprises the following steps: obtaining data which comprises forest land remote sensing image data and ground data; carrying out preprocessing on the obtained data; a dual-path CBAM-UNet network model is constructed; inputting the preprocessed data into the double-path CBAM-UNet network model to carry out tree species classification prediction and obtain a tree species classification chart; performing image splicing on the tree species classification map to obtain a tree species classification grid map; the system comprises a data acquisition module, a data preprocessing module, a prediction module and an image processing module. According to the forest stand tree species identification method and system based on multi-source remote sensing data fusion provided by the invention, multi-scale feature capture and feature fusion of multi-source heterogeneous data can be realized through the constructed dual-path CBAM-UNet network model, high-precision identification can be performed on forest stand tree species under complex links, and high-quality data can be provided for forest management.

Description

Technical Field

[0001] The present invention relates to the technical field of forestry resource monitoring, and in particular to a forest stand tree species identification method and system based on multi-source remote sensing data fusion. Background Art

[0002] Currently, high-precision forest tree species identification relies primarily on manual field surveys. This requires professionals to enter forest areas and visually interpret remote sensing imagery to record the tree species composition and boundaries of each plot. This is labor-intensive and costly, and is difficult to implement in areas with complex terrain. It also makes it difficult to obtain large-scale, continuous forest resource data, hindering timely and accurate decision-making support for forestry management.

[0003] The main methods for automatic tree species identification based on remote sensing technology 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 the recently developed deep learning classification (DLC) (Hamraz et al., 2019; Onishi and Ise, 2021).

[0004] The PSC method compares the spectral signature of each pixel in a remote sensing image with known tree species samples and determines the tree species to which the pixel belongs 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, resulting in "salt and pepper noise" in the classification results and insufficient ability to distinguish between tree species with similar spectral signatures.

[0005] The OBIA method first segments the image into homogeneous objects and then classifies them based on their spectral, shape, and texture characteristics. This method can leverage the spatial structure of tree crowns and reduce "salt and pepper noise," but segmentation parameters are complex to adjust and determining the optimal segmentation scale in heterogeneous forest stands can be difficult, leading to over-segmentation or under-segmentation.

[0006] The MLC method primarily uses algorithms such as random forests, support vector machines, and artificial neural networks to build classification models by extracting spectral and texture features. This method can handle high-dimensional data and is highly capable of modeling nonlinear relationships between features. However, it relies heavily on manually designed feature engineering, has poor scalability, and has limited generalization capabilities across different regions or sensor data.

[0007] The DLC method has developed rapidly in recent years, primarily employing convolutional neural networks (CNNs) to automatically extract hierarchical features. The U-Net architecture and its variants, through an encoder-decoder structure and skip connections, preserve spatial details and are widely used in forest remote sensing segmentation tasks. However, most existing deep learning methods utilize only single optical remote sensing data and lack the ability to exploit vertical structure and seasonal variation.

[0008] The above methods were developed for specific study areas and forest stand conditions. They struggle to effectively distinguish tree species with similar spectral characteristics and are difficult to support large-scale operations. Existing methods struggle to achieve high-precision tree species identification in areas with high canopy density, complex forest structures, and a mix of natural and planted forests.

[0009] Therefore, developing a deep learning method that utilizes high-resolution optical remote sensing and canopy height models to achieve high-precision identification of tree species under complex forest conditions has become a technical problem that needs to be urgently solved in modern forest precision management. Summary of the Invention

[0010] In view of the above-mentioned deficiencies in current forest tree species identification technology, the object of the present invention is to provide a forest stand tree species identification technology for solving the problem of high-precision identification of forest stand tree species under complex forest conditions.

[0011] To achieve the above-mentioned object, a first aspect of the present invention provides a method for identifying tree species in a stand based on multi-source remote sensing data fusion, the method comprising the following steps:

[0012] Acquiring data, wherein the data includes forest land remote sensing image data and ground data;

[0013] Preprocess the acquired data;

[0014] Construct a dual-path CBAM-UNet network model;

[0015] The preprocessed data is input into the dual-path CBAM-UNet network model to perform tree species classification prediction and obtain a tree species classification map;

[0016] The tree species classification map is image stitched to obtain a tree species classification grid map.

[0017] According to a first aspect of the present invention, the forest stand tree species identification method based on multi-source remote sensing data fusion includes: producing deep learning training samples based on the acquired data to train the dual-path CBAM-UNet network model.

[0018] According to the first aspect of the present invention, the step of producing deep learning training samples based on the acquired data to train the dual-path CBAM-UNet network model includes:

[0019] Producing training samples, including image slicing and labeling, and performing normalization, feature engineering, and data augmentation on the samples;

[0020] Model initialization configuration, including using the AdamW optimizer, a cosine annealing scheduler dynamic learning rate strategy, and a weighted cross entropy loss function as the main loss function;

[0021] Model training: During the model training process, a mixed precision training strategy and a gradient clipping method are used to improve the training efficiency of the model and a random dropout strategy is applied to enhance the generalization ability of the model.

[0022] According to the first aspect of the present invention, 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, the input layer is used to receive original image data and simultaneously input the original image data into the left path encoder and the right path encoder through the left path and the right path for parallel processing;

[0023] Wherein, the left path encoder comprises:

[0024] The first-level feature extraction layer, including the BasicConv module;

[0025] The second-level feature extraction layer includes a maximum pooling layer for downsampling and a BasicConv module;

[0026] The third-level feature extraction layer includes a maximum pooling layer and a BasicConv module for downsampling, and Dropout processing is performed after maximum pooling;

[0027] The fourth-level feature extraction layer includes a maximum pooling layer and a BasicConv module for downsampling, and Dropout processing is performed after maximum pooling;

[0028] The right path encoder comprises:

[0029] The first-level feature extraction layer, including the BasicConv module;

[0030] The second-level feature extraction layer includes a maximum pooling layer and a BasicConv module for downsampling, and Dropout processing is performed after maximum pooling;

[0031] The third-level feature extraction layer includes a maximum pooling layer and a BasicConv module for downsampling, and Dropout processing is performed after maximum pooling;

[0032] The fourth-level feature extraction layer includes a maximum pooling layer and a BasicConv module for downsampling, and Dropout processing is performed after maximum pooling.

[0033] According to the first aspect of the present invention, 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 includes RGB bands or multispectral data, and the CHM canopy height model data provides tree height information. The complementary fusion of spectral features and structural features is achieved through dual-path feature extraction.

[0034] According to the first aspect of the present invention, 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;

[0035] When in the standard mode, the input features are only processed by the basic convolution unit and then output;

[0036] When in the ASPP enhancement mode, the input features are processed in parallel by the basic convolution unit and the ASPP submodule at the same time, and the two processing results are fused by element-by-element addition. The fused features are processed by batch normalization and ReLU activation function, and are output after applying Dropout regularization.

[0037] According to the first aspect of the present invention, the dual-path CBAM-UNet network model includes a decoder, an output layer and a CBAM module;

[0038] The decoder has a four-layer structure, each of which includes a transposed convolution module for upsampling and a DoubleConv module for processing fusion features;

[0039] The output layer includes a convolutional layer, which 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;

[0040] The CBAM module is used to enhance the fused features after feature fusion after each feature extraction of the left path encoder and the right path encoder, and after the enhanced processing, perform feature fusion with the output features of the transposed convolution module of the corresponding layer in the decoder through jump connection, and the fused features are input into the DoubleConv module for processing.

[0041] According to the first aspect of the present invention, the dual-path CBAM-UNet network model includes a bottleneck layer, and the bottleneck layer includes:

[0042] A left-path DoubleConv module, configured to process the feature map output by the left-path encoder;

[0043] A right-path DoubleConv module, configured to process the feature map output by the right-path encoder;

[0044] The bottleneck fusion layer 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.

[0045] a terminal enhancement module, which converts the concatenated features into a highly nonlinear representation;

[0046] The CBAM enhancement module applies the attention mechanism to the output of the bottleneck layer for enhancement.

[0047] According to the first aspect of the present invention, the forest remote sensing image data includes airborne digital images, satellite-borne images and CHM canopy height model data, and the preprocessing of the acquired data includes:

[0048] Performing radiometric correction, atmospheric correction, orthorectification, and panchromatic fusion on the airborne digital image and the satellite-borne image;

[0049] The CHM data is subjected to noise suppression, edge preservation and data normalization processing, and the processed CHM data is subjected to image registration processing with the optical image to achieve unified alignment of the spatial coordinate system.

[0050] To achieve the above-mentioned object, the second aspect of the present invention provides a forest stand tree species identification system based on multi-source remote sensing data fusion, comprising:

[0051] Data acquisition module, used to collect forest remote sensing image data and ground data;

[0052] A data preprocessing module, used for preprocessing the forestland remote sensing image data and ground data collected by the data acquisition module;

[0053] The prediction module predicts tree species classification based on the input image slices and obtains a tree species classification map;

[0054] Image processing module, which splices the tree species classification map to obtain a complete tree species classification grid map

[0055] Advantages of this invention: First, the dual-path CBAM-UNet network model captures multi-scale features by fusing GF-7 multispectral imagery with high-resolution CHM data. Leveraging a deep learning model, it captures multi-scale features such as spectrum, spatial texture, and vertical height, improving the ability to distinguish similar tree species and classification accuracy. Second, in areas with mixed coniferous and broad-leaved trees and highly undulating terrain, the classification accuracy exceeds 81%. It can reliably identify tree species in both pure planted and naturally mixed forests, demonstrating strong adaptability to complex woodlands. Third, a fully automated processing workflow reduces manual intervention and fieldwork, automatically outputs classification maps, and rapidly updates large-scale forest stand mapping, creating an automated and efficient process. Finally, through multi-source data feature fusion and algorithmic improvements, the classification accuracy surpasses traditional methods, reaching a Kappa coefficient of 0.78, providing high-quality data for forest management. Furthermore, using an NVIDIA RTX4090 system and the PyTorch 2.2 framework, processing a 12,300-hectare test area took only 2 to 3 hours, demonstrating excellent computing performance and suitability for automated processing of large-scale forest information. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0057] Figure 1 This is a flow chart of the method for identifying tree species in a stand based on multi-source remote sensing data fusion according to the present invention;

[0058] Figure 2 This is a schematic diagram of the deep learning model sample production process described in the present invention;

[0059] Figure 3 This is a schematic diagram of the dual-channel CBAM-Unet network structure of the present invention;

[0060] Figure 4 This is a schematic diagram of the BasicConv module structure diagram of the present invention;

[0061] Figure 5 This is a schematic diagram of the ASPP module structure of the present invention;

[0062] Figure 6This is a schematic diagram of the CBAM module structure of the present invention;

[0063] Figure 7 This is a schematic diagram of the model training process described in the present invention;

[0064] Figure 8 This is a schematic diagram of the tree species classification prediction process of the present invention;

[0065] Figure 9 It is a structural schematic diagram of the stand tree species identification system based on multi-source remote sensing data fusion of the present invention. DETAILED DESCRIPTION

[0066] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0067] Before further explaining the present invention in detail, the nouns and terms involved in the embodiments of the present invention are explained. The nouns and terms involved in the embodiments of the present invention are subject to the following interpretations:

[0068] A canopy height model (CHM) is a digital model that reflects the relative height of trees relative to the ground. It is calculated by subtracting a digital elevation model from a digital surface model obtained using lidar or photogrammetry. It provides pixel-level tree height information. CHM has important applications in forestry remote sensing, ecological research, and resource management, particularly in tree species classification, stand boundary extraction, and biomass estimation.

[0069] ArcMap, developed by the Environmental Systems Research Institute (ESRI), is a core component of the GIS system and one of the key applications in the ArcGIS Desktop suite. It offers powerful mapping, spatial analysis, and data management capabilities, enabling the display, editing, and analysis of map data. It provides data and layout views and supports importing and exporting data in various formats. Users can perform geoprocessing tasks, publish map services, and customize the user experience. It is widely used in a variety of fields, including urban planning, environmental protection, and resource management.

[0070] ASPP (Atrous Spatial Pyramid Pooling): is a module used in deep learning (especially computer vision) that aims to enhance the model's perception of objects of different scales through multi-scale feature extraction and dilated convolution techniques.

[0071] Figure 1 The present invention provides a flowchart of a stand tree species identification method based on multi-source remote sensing data fusion, which includes the following steps:

[0072] Step S1: Acquire data.

[0073] Among them, the data includes forest remote sensing image data and ground data.

[0074] The forest remote sensing imagery used in this study is multispectral data with sub-meter high spatial resolution. It comes from airborne digital imagery, satellite imagery (high-resolution satellite imagery), and canopy height model (CHM) data. Airborne digital imagery is used to generate labels for deep learning model training, while satellite imagery is used for tree species classification and stand boundary delineation.

[0075] In the embodiment of the present invention, the main data sources are the Gaofen-7 (GF-7) satellite multispectral optical images and canopy height model (CHM) data:

[0076] GF-7 satellite imagery: Acquired on May 23, 2023, during the vegetation growth period, the original multispectral bands have a spatial resolution of 2.6 meters. Fusion with panchromatic bands increases the spatial resolution to approximately 0.5 meters. The imagery undergoes preprocessing, including radiometric calibration, atmospheric correction, orthorectification, and panchromatic fusion, to generate a high-quality multispectral orthoimage, which is then resampled to 0.5-meter resolution.

[0077] CHM data: collected on September 3, 2023, using aerial lidar, with a resolution of 0.5 meters. By precisely registering and overlaying CHM data with GF-7 spectral imagery, canopy height information is provided for each pixel, integrating spectral and spatial structural features.

[0078] The ground data used in this paper includes forest tree species sample point data collected using real-time kinematic (RTK) technology, as well as small-scale data from the National Forest Resources Category II Survey. The forest tree species sample point data, which details tree species types and spatial coordinate locations, is primarily used for model verification and validation, while the small-scale data from the National Forest Resources Category II Survey serves as a reference for labeling.

[0079] In an embodiment of the present invention, a field sample plot survey was carried out in the Mulan experimental area in September 2024, and 100 ground sample points were collected for verification. At the same time, the stand sub-class division data formed by the 2021 National Forest Resources Class II Survey was obtained, which can be used for accuracy verification and auxiliary labeling. Based on the above data, the tree species classification system of the experimental area was finally determined, and the target categories included 7 categories, namely larch, Chinese pine, Mongolian pine, white birch, oak, bare land and background.

[0080] After preprocessing, these remote sensing image data and ground data can be used as training data for deep learning neural network models.

[0081] Step S2: Preprocess the acquired data.

[0082] In the embodiment of the present invention, different pre-processing methods are used for the acquired original remote sensing image data and ground data according to the difference in data types:

[0083] For airborne and satellite-borne data, the preprocessing process is as follows:

[0084] Radiometric correction: For the panchromatic and multispectral images of the GF-7 satellite, the absolute radiometric correction method is used. The correction coefficient is selected according to the sensor model to convert the DN value into the apparent reflectance.

[0085] Atmospheric correction: For GF-7 multispectral images, the FLAASH model is used to eliminate atmospheric effects;

[0086] Orthorectification: Panchromatic and multispectral images are geometrically rectified based on the 0.5m resolution DEM and RPC parameters obtained by LiDAR, with control point accuracy better than 0.5 pixel;

[0087] Panchromatic fusion: Panchromatic images are fused using the Gram-Schmidt sharpening algorithm to increase the spatial resolution to 0.5 m while maintaining spectral characteristics.

[0088] Image registration processing is performed on the CHM canopy height model data and the optical image. The registration process includes:

[0089] Image registration: The GF-7 image and CHM data were unified to a 0.5 m resolution grid using the nearest neighbor resampling method;

[0090] Noise suppression: A local adaptive Frost filter with a 7×7 window and a K factor of 2.0 was applied to the GF-7 image and CHM data to suppress abnormally high-value noise.

[0091] Edge preservation: When filtering, the edge preservation parameter α is set to 0.4 to preserve the forest stand boundary features;

[0092] Data normalization: The CHM height values ​​and GF-7 multispectral values ​​were linearly normalized to the range of 0-1.

[0093] Through the above preprocessing, multi-source data integrating multispectral and height information is generated, providing a basis for subsequent analysis.

[0094] Step S3: Create deep learning training samples based on the preprocessed data.

[0095] Figure 2A schematic diagram showing the process of producing deep learning training samples and the division method of the data set in an embodiment of the present invention is shown.

[0096] First, ArcMap software was used to select representative areas within the experimental area. 0.1-meter-resolution aerial images were used indoors, and the second-category forest resource survey small-scale data were referred to to visually interpret and outline the preliminary draft boundaries of forest stands of different tree species.

[0097] We then used RTK to collect forest plot and tree species sample points, recorded tree species types and patch boundary coordinates, verified and corrected stand vector patches, and obtained a high-quality stand boundary vector map. This map was then rasterized to the pixel grid of the GF-7 image (with a grid length of 0.5 meters) to produce a tree species classification label map.

[0098] At the same time, to meet the input requirements of the deep learning model, the labels and corresponding images were cropped into small slices. In this embodiment of the present invention, the size of the cropped small slice images was 256×256 pixels, and a total of 14,159 labeled image samples were produced for model training and evaluation.

[0099] It can be understood that in order to enhance the training effect and generalization ability of the model, in the embodiment of the present invention, the training sample data are respectively normalized and feature-engineered.

[0100] In terms of normalization, the mean of each feature dimension is first calculated based on the entire data set. and standard deviation , then for each data point Execute formula This operation maps the original data to a standard normal distribution space with a mean of 0 and a standard deviation of 1. This not only eliminates the dimensional differences between different features, but also preserves the distribution of the data, making gradient descent-based model training more efficient and avoiding optimization bias caused by feature scale differences.

[0101] In terms of feature engineering, we can extract manual features such as texture, shape, and color histograms based on the spatial structure and semantic information of image data, enhancing the data's representational capabilities and providing more discriminative input for the model. Furthermore, by combining the advantages of automatic feature extraction through deep learning, we can gradually abstract the high-level semantic features of images through structures such as convolutional neural networks. This not only reduces the subjectivity of manual feature design, but also explores the complex connections implicit in the data, further improving the model's understanding and generalization capabilities of image scenes.

[0102] In addition, in order to increase the diversity of training data, avoid overfitting and enhance the generalization ability of the model, data enhancement technology is used to expand the training data set. In the embodiment of the present invention, the data enhancement technology used includes:

[0103] Geometric augmentation: random horizontal / vertical flip (probability 0.5), random rotation (rotation amplitude ±15°, probability 0.3).

[0104] Spectral enhancement: random brightness adjustment (adjustment range is ±10%, probability is 0.3), random contrast adjustment (adjustment range is ±10%, probability is 0.3).

[0105] Spatial enhancement: Random cropping, the cropped image size is 256×256 and more than 80% of valid pixels are retained.

[0106] Finally, the prepared sample data is divided into a training set, a validation set, and a test set according to a certain ratio to ensure the fairness and reliability of the model training process and accuracy evaluation. In the embodiment of the present invention, the ratio of the training set, validation set, and test set is 6:2:2.

[0107] Step S4: Construct a dual-path CBAM-UNet network model.

[0108] In terms of neural network design, the embodiment of the present invention adopts an improved dual-path CBAM-UNet deep convolutional neural network architecture, which uses multispectral and canopy height data to accurately identify different tree species.

[0109] Figure 3 The overall structure of the dual-path CBAM-UNet network model of the present invention is shown in the figure. 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. The following will be combined with the schematic diagram to explain each module of the neural network in detail.

[0110] Input layer: The input layer provides dual-path input channels. 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.

[0111] In the embodiment of the present invention, the GF-7 multispectral data has 4 channels and a size of 256×256 pixels; the CHM canopy height data has 1 channel and a size of 256×256 pixels.

[0112] Encoder: The encoder includes a parallel left-path encoder and a right-path encoder, which perform feature extraction on feature data inputted by the left-path input channel and the right-path input channel, respectively. Both the left-path encoder and the right-path encoder have multi-layer structures to implement multi-layer, step-by-step feature extraction. In this embodiment of the present invention, both the left-path encoder and the right-path encoder have a four-layer structure. The input layer simultaneously inputs the original image data into the left-path encoder and the right-path encoder for parallel processing.

[0113] Among them, 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. The optical remote sensing image data includes RGB bands or multispectral data. Its network structure is as follows:

[0114] The first-level feature extraction layer includes a BasicConv module. In the embodiment of the present invention, the BasicConv module outputs a 96-channel feature map with a size of 256×256 pixels.

[0115] The second-level feature extraction layer includes a max pooling layer for downsampling and a BasicConv module. Dropout processing can be performed after max pooling. In this embodiment of the present invention, the pooling kernel size is 2×2, and the BasicConv module outputs a 128-channel feature map of size 128×128 pixels.

[0116] The third-level feature extraction layer includes a max pooling layer for downsampling and a BasicConv module. Dropout processing can be performed after max pooling. In this embodiment of the present invention, the pooling kernel size is 2×2, and the BasicConv module outputs a 256-channel feature map of size 64×64 pixels.

[0117] The fourth-level feature extraction layer includes a max pooling layer for downsampling and a BasicConv module, followed by dropout processing. In this embodiment of the present invention, the pooling kernel size is 2×2, and the BasicConv module outputs a 512-channel feature map of size 32×32 pixels.

[0118] Among them, the right path encoder is used to process the CHM canopy height data input by the right path input channel. The CHM canopy height model data provides tree height information. Its network structure is as follows:

[0119] The first-level feature extraction layer includes a BasicConv module. In the embodiment of the present invention, the BasicConv module outputs a 64-channel feature map of size 256×256.

[0120] The second-level feature extraction layer includes a maximum pooling layer for downsampling and a BasicConv module. Dropout processing can be performed after maximum pooling. In this embodiment of the present invention, the BasicConv module outputs a 128-channel feature map of size 128×128 pixels.

[0121] The third-level feature extraction layer includes a max pooling layer for downsampling and a BasicConv module. Dropout processing can be performed after max pooling. In this embodiment of the present invention, the BasicConv module outputs a 256-channel feature map of 64×64 pixels.

[0122] The fourth-level feature extraction layer includes a maximum pooling layer for downsampling and a BasicConv module. Dropout processing can be performed after maximum pooling. In this embodiment of the present invention, the BasicConv module outputs a 512-channel feature map of size 32×32 pixels.

[0123] In the embodiment of the present invention, feature extraction and fusion of multi-source heterogeneous data are achieved through dual-path and multi-level parallel feature extraction, including the complementary fusion of spectral features and structural features.

[0124] The Dropout process in each feature extraction layer of the left and right path encoders is an optional operation. Dropout can randomly drop neurons to prevent overfitting of the neural network and improve the generalization ability of the model. In this embodiment of the present invention, the Dropout Rate is set to 0.3.

[0125] In the embodiment of the present invention, the BasicConv module used in each feature extraction layer in 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 figure, the BasicConv module of the present invention has two working modes: Figure 4 The left side shows the network structure diagram of the BasicConv module in standard mode; Figure 4 The right side shows the network structure diagram of the BasicConv module in ASPP enhanced mode.

[0126] Depend on Figure 4 As shown on the left, the BasicConv module in standard mode consists of two convolutional layers connected in series, each followed by BatchNorm and ReLU activation functions. To prevent overfitting, each layer can be followed by a Dropout layer for regularization. In this case, the input features are processed only by the basic convolutional units before being output. In this embodiment of the present invention, the convolution kernel size of both convolutional layers is 3×3, and the Dropout Rate is 0.5.

[0127] Depend on Figure 4 As can be seen on the right, the BasicConv module in ASPP enhancement mode integrates the ASPP module into the network structure. Input features are processed in parallel by both the basic convolutional unit and the ASPP submodule. The results of these two processes are fused by element-by-element addition. The fused features are then batch normalized and activated with the ReLU activation function, and then regularized with Dropout before being output. This gives the BasicConv module better multi-scale feature extraction capabilities.

[0128] In this embodiment of the present invention, the dual-path CBAM-UNet network model has a preset aspp mode switch, which determines the operating mode of the BasicConv module. In specific application scenarios, when the model needs to balance computational efficiency and basic feature extraction capabilities (such as processing small objects or low-complexity scenes), the aspp mode switch parameter can be set to standard mode. When the task requirements place greater emphasis on capturing multi-scale contextual information (such as large object segmentation or complex scene classification), the aspp mode switch parameter can be set to ASPP enhanced mode to improve the model's feature expression capabilities.

[0129] Figure 5 The network structure diagram of the ASPP module is shown. As can be seen from the figure, the input layer of the ASPP module receives the feature map as the input data for module processing. The input layer is followed by five parallel branches, which are:

[0130] First convolution branch: In an embodiment of the present invention, this branch performs a convolution operation on the input feature map with a 1×1 convolution kernel to extract local detail features and capture the correlation between feature map channels.

[0131] Second dilated convolution branch: In the embodiment of the present invention, this branch uses a 3×3 convolution kernel and a dilation rate of 6 for dilated convolution, which is used to expand the receptive field and capture medium-scale features.

[0132] The third dilated convolution branch: In the embodiment of the present invention, this branch uses a 3×3 convolution kernel and a dilation rate of 12 for dilated convolution, which is used to further increase the receptive field and obtain larger-scale features.

[0133] Fourth dilated convolution branch: In the embodiment of the present invention, this branch uses a 3×3 convolution kernel and a dilation rate of 18 for the dilated convolution, which is used to give the model a larger receptive field and capture global scale features.

[0134] The fifth pooling and convolution branch first performs a global average pooling operation on the feature map through a pooling layer with a pooling kernel of 1×1; then performs a 1×1 convolution; then it is processed by batch normalization (BatchNorm) and ReLU activation function; finally, the feature map size is adjusted using bilinear interpolation (Bilinear Interpolate).

[0135] After the feature map is processed by five parallel branches, it is concatenated in the channel dimension by the fusion layer to fuse multi-scale feature information. The concatenated feature map is then convolved using a 1x1 convolution kernel, and the number of channels in the feature map is adjusted.

[0136] Feature fusion layer: The dual-path CBAM-UNet network of the present invention has multiple feature fusion layers, which provide a multi-scale feature fusion mechanism and implement a progressive cascade feature fusion strategy at each encoding level to fuse the left and right path features.

[0137] It can be understood that in the multi-level feature fusion process, in order to meet the conditions for feature fusion, the feature map sizes of the left and right paths must be consistent. At the same time, the feature maps are superimposed to increase the number of channels of the feature maps, thereby enriching the feature expression and avoiding information loss. The specific process of multi-level feature fusion is:

[0138] First-level fusion: Concatenates the left and right path features to form an enhanced representation. In this embodiment of the present invention, the feature map input from the left path has 96 channels, while the feature map input from the right path has 64 channels. Both feature maps are sized at 256×256 pixels, resulting in a 160-channel, 256×256 feature map after fusion.

[0139] Second-level fusion: Concatenate the left and right path features to form an enhanced representation of the features. In this embodiment of the present invention, the number of channels of the left and right path input feature maps is 128, and the size of the feature maps is 128×128 pixels. After fusion, a 128×128 feature map with 256 channels is obtained.

[0140] Third-level fusion: Concatenate the left and right path features to form an enhanced representation of the features. In this embodiment of the present invention, the number of channels of the left and right path input feature maps is 256, and the size of the feature maps is 64×64 pixels. After fusion, a 64×64 feature map with 512 channels is obtained.

[0141] Fourth-level fusion: Cascade the left and right path features to form an enhanced feature representation. In this embodiment of the present invention, the number of channels of the left and right path input feature maps is 512, and the size of the feature maps is 32×32 pixels. After fusion, a 32×32 feature map with 1024 channels is obtained.

[0142] Bottleneck layer: The dual-path CBAM-UNet network model of the present invention includes a bottleneck layer. The bottleneck layer is a key component connecting the encoder and decoder. The bottleneck layer can further enhance the fused high-level features. In this embodiment of the present invention, the bottleneck layer includes the following modules:

[0143] DoubleConv module: This module includes a left-path DoubleConv module and a right-path DoubleConv module. The left and right DoubleConv modules process the feature maps output by the left and right encoders, respectively. In this embodiment of the present invention, as described above, the encoders for both left and right encoders output 512-channel feature maps of 32×32 pixels.

[0144] Bottleneck layer fusion module: used to cascade the features of the left and right paths to generate a 1024-channel feature map of size 32×32 pixels.

[0145] Terminal enhancement module: The multi-channel feature map output by the bottleneck layer fusion module is converted into a highly nonlinear representation through an additional DoubleConv module.

[0146] CBAM Enhancement Module: This 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.

[0147] The bottleneck layer effectively addresses the issues of feature redundancy and boundary ambiguity in multi-source data integration by implementing multiple strategies, including cross-modal fusion, attention enhancement, and parameter compression. This allows the model to significantly improve boundary segmentation accuracy compared to traditional UNet models while maintaining efficiency.

[0148] Decoder: The dual-path CBAM-UNet network model of the present invention includes a decoder, which is used to restore the spatial resolution of the feature map. In this embodiment of the present invention, the decoder also has a four-layer structure. Unlike the encoder, each layer of the decoder includes a transposed convolution module for upsampling and a DoubleConv module for processing fused features. In this embodiment of the present invention, the convolution kernel size of the transposed convolution module is 2×2.

[0149] Simultaneously, the enhanced features obtained from the CBAM modules in each layer of the left and right encoders are connected to the corresponding layers in the decoder via skip connections. The upsampled features and the enhanced features corresponding to the skip connections are combined through a cascade operation. During the decoder, the number of channels in the feature map decreases (512 → 256 → 128 → 64) while the size of the feature map increases (32 → 64 → 128 → 256).

[0150] Output Layer: The dual-path CBAM-UNet network model of the present invention also includes an output layer. This layer maps the decoder's output features to predefined categories of stand tree species, specifically seven predefined categories, using a 1×1 convolutional layer. Finally, a Softmax activation function is applied to generate a probability distribution for each category. Here, the output feature map of the output layer is 256×256 pixels.

[0151] In summary, the dual-path CBAM-UNet network of the present invention comprises two parallel encoder paths, one for processing spectral information (from GF-7 four-channel multispectral data) and the other for processing height information (from single-channel CHM data). Through a "separation-fusion" mechanism, the neural network independently extracts features from different data types, avoiding mutual interference. Feature fusion is then performed at four resolution levels, leveraging the complementarity of spectral and height information. This multi-level feature extraction and fusion enables high-precision identification of different tree species.

[0152] First, the network architecture combines the CBAM attention module and ASPP multi-scale feature extraction technology. CBAM uses a dual channel and spatial attention mechanism to enhance key features in areas such as tree species boundaries. Simultaneously, the ASPP module uses parallel convolutions with multiple dilation rates (6, 12, and 18) to capture crown features at different scales, thereby improving the ability to identify tree species of varying sizes.

[0153] Secondly, the network optimized its skip connection design. Based on the traditional U-Net network model, the dual-path features were fused and enhanced with CBAM before being connected to the decoder. This improved the network model's ability to recognize blurred boundaries between high-density and mixed forests.

[0154] Thirdly, through path-differentiated design and parameter sharing, the total number of model parameters is controlled at 126 million, effectively improving training and application efficiency.

[0155] In this embodiment, the model inputs are the four GF-7 multispectral bands and one CHM height band, and the output is the tree species classification for the corresponding pixel. The classification covers seven categories, including dominant tree species (such as larch, Chinese pine, Korean pine, white birch, and oak), as well as other minor tree species or non-forested land (denoted as "Other").

[0156] like Figure 6 As shown, the CBAM module (Convolutional Block Attention Mechanism) of the present invention includes both channel-wise and spatial-wise attention mechanisms. Channel-wise attention is used to process max-pooled and average-pooled features in parallel and extract channel-wise attention weights using a shared MLP. Spatial attention generates feature maps using the maximum and average values ​​of the channel dimensions, and applies a 7×7 convolution to generate the spatial attention map. In this embodiment of the present invention, the CBAM module is applied after the bottleneck layer and at each skip connection, significantly enhancing the representation of important features.

[0157] The DoubleConv module of this invention achieves layer-by-layer feature abstraction, nonlinear modeling, and channel adaptation through a combination of "double convolution + nonlinear activation + normalization." It is a fundamental component for improving feature representation in the dual-path CBAM-UNet model. In embodiments of this invention, this module, combined with ASPP, CBAM, and other methods, significantly improves the ability to distinguish spectrally similar tree species in complex forest stands (for example, the classification accuracy of larch and Chinese pine increased by approximately 5%) and optimizes the refinement of boundary extraction. The DoubleConv module consists of two 3×3 convolutional layers, each followed by batch normalization (BatchNorm) and a ReLU activation function. This design, by stacking convolutional layers, enables the extraction of more complex features while maintaining the spatial size of the feature map (by setting padding = 1). In specific implementations, the number of input and output channels of DoubleConv can be adjusted according to network requirements, and the number of output channels of the second convolutional layer is typically consistent with the number of output channels of the module.

[0158] In this embodiment of the present invention, DoubleConv enhances the network's feature representation capabilities. Through two convolution operations, the network can learn more abstract features. The introduction of batch normalization (BatchNorm) and the ReLU activation function also helps accelerate training and alleviate the vanishing gradient problem. This modular design allows DoubleConv to be flexibly embedded in the encoder and decoder of a UNet, becoming a crucial component of the network architecture.

[0159] Step S5: Use the preprocessed data to train the dual-path CBAM-UNet network model.

[0160] Figure 7 The specific process of dual-path CBAM-UNet network model training of the present invention is demonstrated, which specifically includes the following steps:

[0161] Step S51: Model initialization configuration.

[0162] The dual-path CBAM-UNet network model constructed in this embodiment of the present invention is designed to achieve effective semantic segmentation of multi-channel remote sensing imagery. The model is designed as a dual-path structure, with the left path inputting 4-channel multispectral image data and the right path inputting 1-channel CHM image data, for a total of 5 channels of input data. The model ultimately outputs an 8-channel classification map, corresponding to different tree species and background areas.

[0163] During the initialization phase of model training, an optimizer and learning rate scheduling strategy were configured. Specifically, the AdamW optimizer was used for parameter updates, with momentum parameters β1 set to 0.9 and β2 set to 0.999. Weight decay (weight_decay) was set to 0.01 to prevent overfitting. Regarding the learning rate, the initial learning rate was set to 0.001, with a minimum learning rate limit of 1e-6. The cosine annealing scheduling algorithm (CosineAnnealingLR) was used to dynamically adjust the learning rate, accelerating model convergence and improving training stability.

[0164] In order to improve the model's adaptability to imbalanced category distribution, the loss function design adopts a combination of main loss and auxiliary loss:

[0165] Main loss function: Using the weighted cross entropy loss function, the weight of each category is calculated according to the inverse proportion of its sample number to alleviate the problem of class imbalance.

[0166] Auxiliary loss function: Combine Dice loss and cross entropy loss, and perform a weighted combination in a ratio of 0.4:0.6 to enhance the model's ability to recognize the boundaries of the target area.

[0167] Enhanced balancing mechanism for specific categories: For target categories with a small number of samples (such as oak trees), FocalLoss (γ=2.0) is introduced to further improve the model's recognition effect on minority classes.

[0168] The organic combination of the above model structure and training strategy enables the present invention to have stronger expression ability and higher classification accuracy in complex vegetation classification tasks.

[0169] Step S52: Perform model training using the prepared training sample data.

[0170] During model training, the data loader loads training and validation data in appropriate batch sizes. The training data is randomly shuffled to improve the model's generalization capabilities, while the validation data is kept in a fixed order to facilitate stable evaluation of model performance. The dataset size is also recorded to provide a basis for subsequent calculation of performance metrics. In this embodiment of the present invention, the batch size is set to 16 during the training phase and 30 during the prediction phase.

[0171] As you can understand, model training requires multiple cycles of continuous optimization, each of which is primarily divided into two phases: training and validation. During the training phase, a loss function is used to adjust model parameters, allowing the model to learn features from the data. During the validation phase, validation data, which was not used in training, is used to evaluate model performance and monitor the model's generalization capabilities. To improve training efficiency and stability, the embodiments of this invention utilize a mixed-precision training strategy and gradient clipping methods, along with random dropout (Dropout) to enhance the model's generalization capabilities. The Dropout Rate can be set based on actual needs and is not specifically limited in this invention.

[0172] Model performance is comprehensively measured by calculating loss, pixel-level accuracy, and intersection-over-union (IoU, based on the overlap ratio between predicted and true labels). The weight of the model with the highest accuracy on the validation set is saved. Furthermore, training progress is recorded during the training process, and computing resources are regularly cleared to optimize memory usage and ensure smooth training. In this embodiment of the present invention, the training cycle is set to a total of 100 epochs. During training, if the validation set accuracy does not improve for 10 consecutive epochs, training is terminated early to avoid wasting resources.

[0173] Step S53: Model evaluation and optimization.

[0174] During the model training phase, the training set images are fed into the dual-path CBAM-UNet network for iterative training. Simultaneously, the validation set is used to evaluate model performance and adjust hyperparameters to monitor model performance. Specifically, the loss value, accuracy, and intersection-over-union (IoU) are output for each epoch to evaluate the model's classification and segmentation capabilities. The training time is also recorded to analyze the model's efficiency and effectiveness.

[0175] Once the model evaluation results meet the expected standards, we will end training and export the model. If they do not meet expectations, we will adjust the parameters based on the evaluation feedback and continue training. During the training phase, the training set images are input into the network model for iterative training, while the validation set is used to guide the adjustment of hyperparameters.

[0176] In this embodiment of the present invention, after multiple rounds of training and parameter optimization, the model achieved an accuracy of 87.66% on the training set and 80.07% on the validation set, demonstrating both good convergence and strong generalization. Ultimately, the best-performing model was selected and applied to the test set, achieving an overall classification accuracy of 81.83%, a mean intersection-over-union (mIoU) of 70.39%, and a Kappa coefficient of 0.7803, demonstrating the model's effectiveness.

[0177] When the model training meets the expected performance, the model is exported and deployed to the target device for subsequent tree species classification tasks.

[0178] Step S6: Use the deployed dual-path CBAM-UNet network model to perform tree species classification prediction and obtain a tree species classification map.

[0179] In this embodiment of the present invention, a trained dual-path CBAM-UNet model was applied to the entire remote sensing imagery of the Mulan experimental area, achieving pixel-by-pixel tree species classification prediction across a large area. Due to the large study area, a sliding window was used to crop the GF-7 and CHM images into contiguous, non-overlapping 256×256 pixel blocks. These blocks were then fed into the dual-path CBAM-UNet model for forward prediction, yielding the corresponding small-slice classification results.

[0180] like Figure 8 As shown, the following steps are specifically included to predict tree species classification and obtain tree species classification maps:

[0181] Step S61: Data preparation.

[0182] Collect GF-7 and CHM images of the test area and segment them into several continuous, non-overlapping image slices. Preprocess the sliced ​​images. The preprocessing process has been described in detail in step S2 and will not be repeated here. The preprocessed image data also needs to be sliced, using the same method as used in the training sample generation. The image slice size is also 256×256 pixels.

[0183] Step S62: Load the trained model.

[0184] 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 dual-path CBAM-UNet network model using a deep learning framework (such as PyTorch / TensorFlow). To improve prediction efficiency, deploy the model to a computing device equipped with a high-performance GPU such as the NVIDIA RTX4090, using CUDA to accelerate parallel computing.

[0185] After loading is complete, 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 updates, reduce memory usage, and speed up the forward propagation process.

[0186] Step S63: Slice prediction.

[0187] The preprocessed image slices are loaded into the model input queue in batches. In this example, the batch size is set to 30 to balance GPU memory usage and computational efficiency. The model performs a forward pass on each slice, outputting a probability distribution tensor of 256×256×7 dimensions (with seven channels corresponding to the predicted probabilities of the six target tree species and background). An argmax operation is applied to extract the maximum probability class index for each pixel along the channel dimension, converting the continuous probability distribution into a discrete pixel-level tree species classification map.

[0188] Finally, the predicted tree species classification map is saved according to the position information of the original slices for subsequent splicing processing.

[0189] Step S7: Perform image stitching on the tree species classification map to obtain a tree species classification raster map.

[0190] The prediction results of all slices are stitched together according to their geographic locations in the original image. During the stitching process, the adjacent block results are smoothed to eliminate stitching marks and ensure the continuity and consistency of the classification map. Finally, a complete tree species distribution classification map of the test area is generated. The result map is stored in raster format, and each pixel is assigned a corresponding tree species category code. The specific process of image stitching includes the following steps:

[0191] Step S71: Geographic coordinate mapping and positioning.

[0192] When stitching tree species classification maps, the first step is to establish the positional correspondence between each predicted slice and the original image based on the spatial reference information recorded during the slicing process, such as latitude and longitude ranges, row and column numbers, etc. A global raster index is also constructed to determine the start and end coordinates of each slice within the complete raster map. Overlapping areas are also marked to provide clear boundary information for subsequent fusion processing. This step is fundamental to ensuring accurate stitching positions.

[0193] Step S72: Fusion of overlapping area features.

[0194] When fusing features in overlapping regions, adjacent slices often overlap to a certain extent. For these regions, a weighted averaging strategy can be used to achieve a smooth transition. Weights are calculated using a spatial distance decay function. For example, the weight is equal to the product of the negative α of the natural exponential function and the square of the distance, where α is a tuning parameter and the distance refers to the distance from the pixel to the slice boundary. For pixels with inconsistent classifications, a prediction with a higher confidence level, such as the maximum probability value output by Softmax, is prioritized. This ensures a more reasonable classification result for overlapping regions.

[0195] Step S73: Boundary feathering and gradient smoothing processing.

[0196] During the boundary feathering and gradient smoothing process, a Gaussian blur filter is applied to the slice edges, typically with a σ setting of 2-3 pixels, to create a transition band 5-10 pixels wide. For areas with dramatic gradient changes, such as forest canopy edges, morphological dilation or erosion operations are introduced to optimize boundary continuity. Furthermore, utilizing the topographic information in the CHM data, elevation-weighted fusion is performed on areas with significant topographical fluctuations to reduce the bias caused by hillshade on classification and ensure a more natural and continuous boundary in the stitched image.

[0197] Step S74: integrating raster coding with metadata.

[0198] Finally, the raster encoding and metadata integration process is completed. The fused classification results are converted into a standard GeoTIFF raster format, so that each pixel value corresponds to a specific tree species code, such as 1 for larch and 2 for Chinese pine. Spatial reference system information, including projection and geographic coordinate system, is also embedded. A classification attribute table is also added to record metadata such as the tree species name and confidence threshold corresponding to each code. The result is a spatially continuous, attribute-consistent tree species distribution raster map with complete geographic location information.

[0199] Finally, the generated tree species distribution raster map is exported and saved, and relevant metadata information is attached to facilitate subsequent application and analysis.

[0200] Figure 9 The schematic diagram of the structure of a stand tree species identification system based on multi-source remote sensing data fusion in one embodiment of the present application is shown. The system 600 in the present application includes a graph data acquisition module 601, a data preprocessing module 602, a prediction module 603 and an image processing module 604.

[0201] Data acquisition module 601 is responsible for collecting forest remote sensing image data (including airborne digital images, satellite images, and CHM canopy height model data) and ground data (tree species sample points collected by RTK and national forest resources Class II survey small group data). GF-7 satellite multispectral images and CHM data must be acquired during the vegetation growth period and the lidar acquisition period. Ground data are used for model training and verification to ensure that the data covers representative areas of the experimental area.

[0202] The data preprocessing module 602 is responsible for performing radiometric correction, atmospheric correction, orthorectification, and panchromatic fusion on remote sensing images, increasing the resolution of GF-7 images to 0.5 meters; performing noise suppression, edge preservation, and normalization on CHM data, and aligning it with optical images to unify the spatial coordinate system, ultimately generating multi-source data that fuses spectral and altitude information in preparation for model input.

[0203] The prediction module 603 is constructed based on a dual-path CBAM-UNet network model. The left path processes GF-7 multispectral data to extract spectral features, and the right path processes CHM data to extract structural features. Features are extracted in parallel through a four-level encoder, and the CBAM module is used to enhance key features. The ASPP mode switch is combined to optimize multi-scale feature capture. After bottleneck layer fusion, the decoder restores the spatial resolution and finally outputs the tree species classification probability distribution.

[0204] During operation, the prediction module 603 slices the preprocessed image into 256×256 pixels, loads the trained model weights to the GPU, performs forward propagation with a batch size of 30, outputs the probability tensor and converts it into a pixel-level classification map through argmax, and records the geographic coordinates when saving the slicing results to provide location information for subsequent splicing.

[0205] The image processing module 604 performs geographic coordinate mapping based on the spatial reference information of the slices, fuses the overlapping areas using distance-weighted averaging and confidence-first strategies, smoothes the boundaries through Gaussian blur and morphological operations, and finally converts the classification results into GeoTIFF format, embeds the spatial reference system and integrates metadata to generate a complete tree species classification raster map.

[0206] In summary, the forest stand tree species identification method and system based on multi-source remote sensing data fusion of the present invention has the following advantages:

[0207] First, multi-scale feature capture: Utilizing GF-7 multispectral imagery and high-resolution CHM data, combined with a deep learning model using an attention mechanism and dilated pyramid pooling, the model simultaneously captures multi-scale feature information about the forest stand, including its spectrum, spatial texture, and vertical height. This enables the model to discern subtle differences between tree species, even between species with similar spectral characteristics (such as Scots pine and Chinese pine) by distinguishing them by canopy height or texture, thereby improving classification accuracy.

[0208] Second, adaptability to complex woodlands: The method achieved classification accuracy exceeding 81% in the Magnolia test area, which features mixed coniferous and broadleaf trees and a highly undulating terrain, demonstrating its adaptability to complex, heterogeneous woodlands. The results demonstrate that the method can reliably identify tree species in both large, contiguous areas of pure planted forests and naturally mixed, mixed-age stands.

[0209] Third, an automated and efficient process: From data preprocessing and model training to result generation, a complete automated process has been established, significantly reducing manual intervention and fieldwork. The model automatically outputs tree species classification maps and stand boundaries based on remote sensing data, eliminating the traditional, time-consuming manual delineation. Compared to manual surveys, this method can complete large-scale stand mapping and updates in a shorter time, facilitating timely information on forest resource changes.

[0210] Fourth, improved accuracy: Thanks to the fusion of multi-source data features and an improved model algorithm, the proposed method significantly improves the accuracy of tree species classification and identification compared to traditional methods. This high-precision data product can serve as high-quality basic data for forest resource management, providing strong support for forest management decision-making and ecological research.

[0211] Fifth, improved computing performance: In actual application tests, the method of the present invention was run on an NVIDIA RTX4090 system using the PyTorch 2.2 deep learning framework. The complete processing time for a 12,300-hectare test area was approximately 2 to 3 hours. It has excellent computing efficiency and economic benefits and can support large-scale automated forest information processing tasks.

[0212] In summary, the present method for identifying tree species, based on high-resolution optical remote sensing and a canopy height model, successfully achieves high-precision, automated acquisition of forest resource information. This method leverages the advantages of multi-source remote sensing data and, combined with deep learning, enables accurate identification of tree species in complex woodlands, possessing significant practical significance for forest resource surveys and monitoring.

[0213] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A method for identifying tree species in a stand based on multi-source remote sensing data fusion, characterized in that: The stand tree species identification method based on multi-source remote sensing data fusion comprises the following steps: Acquiring data, wherein the data includes forest land remote sensing image data and ground data; Preprocess the acquired data; Construct a dual-path CBAM-UNet network model; The preprocessed data is input into the dual-path CBAM-UNet network model to perform tree species classification prediction and obtain a tree species classification map; The tree species classification map is image stitched to obtain a tree species classification grid map.

2. The method for identifying tree species in a stand based on multi-source remote sensing data fusion according to claim 1, characterized in that: The stand tree species identification method based on multi-source remote sensing data fusion includes: preparing deep learning training samples based on the acquired data to train the dual-path CBAM-UNet network model.

3. The method for identifying tree species in a stand based on multi-source remote sensing data fusion according to claim 2, characterized in that: The step of producing a deep learning training sample based on the acquired data to train the dual-path CBAM-UNet network model includes: Producing training samples, including image slicing and labeling, and performing normalization, feature engineering, and data augmentation on the samples; Model initialization configuration, including using the AdamW optimizer, a cosine annealing scheduler dynamic learning rate strategy, and a weighted cross entropy loss function as the main loss function; Model training: During the model training process, a mixed precision training strategy and a gradient clipping method are used to improve the training efficiency of the model and a random dropout strategy is applied to enhance the generalization ability of the model.

4. The method for identifying tree species in a stand based on multi-source remote sensing data fusion according to claim 1, characterized in that: The dual-path CBAM-UNet network model includes an input layer and an encoder, wherein the input layer includes a left path and a right path, and the encoder includes a left path encoder and a right path encoder. The input layer is used to receive original image data and simultaneously input the original image data into the left path encoder and the right path encoder through the left path and the right path for parallel processing; Wherein, the left path encoder comprises: The first-level feature extraction layer, including the BasicConv module; The second-level feature extraction layer includes a maximum pooling layer for downsampling and a BasicConv module; The third-level feature extraction layer includes a maximum pooling layer and a BasicConv module for downsampling, and Dropout processing is performed after maximum pooling; The fourth-level feature extraction layer includes a maximum pooling layer and a BasicConv module for downsampling, and Dropout processing is performed after maximum pooling; The right path encoder comprises: The first-level feature extraction layer, including the BasicConv module; The second-level feature extraction layer includes a maximum pooling layer and a BasicConv module for downsampling, and Dropout processing is performed after maximum pooling; The third-level feature extraction layer includes a maximum pooling layer and a BasicConv module for downsampling, and Dropout processing is performed after maximum pooling; The fourth-level feature extraction layer includes a maximum pooling layer and a BasicConv module for downsampling, and Dropout processing is performed after maximum pooling.

5. The method for identifying tree species in a stand based on multi-source remote sensing data fusion according to claim 4, characterized in that: 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 includes RGB bands or multispectral data, and the CHM canopy height model data provides tree height information. The complementary fusion of spectral features and structural features is achieved through dual-path feature extraction.

6. The method for identifying tree species in a stand based on multi-source remote sensing data fusion according to claim 4, characterized in that: 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; When in the standard mode, the input features are only processed by the basic convolution unit and then output; When in the ASPP enhancement mode, the input features are processed in parallel by the basic convolution unit and the ASPP submodule at the same time, and the two processing results are fused by element-by-element addition. The fused features are processed by batch normalization and ReLU activation function, and are output after applying Dropout regularization.

7. The method for identifying tree species in a stand based on multi-source remote sensing data fusion according to claim 4, characterized in that: The dual-path CBAM-UNet network model includes a decoder, an output layer and a CBAM module; The decoder has a four-layer structure, each of which includes a transposed convolution module for upsampling and a DoubleConv module for processing fusion features; The output layer includes a convolutional layer, which 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 after each feature extraction of the left path encoder and the right path encoder, and after the enhanced processing, perform feature fusion with the output features of the transposed convolution module of the corresponding layer in the decoder through jump connection, and the fused features are input into the DoubleConv module for processing.

8. The method for identifying tree species in a stand based on multi-source remote sensing data fusion according to claim 4, characterized in that: The dual-path CBAM-UNet network model includes a bottleneck layer, which includes: A left-path DoubleConv module, configured to process the feature map output by the left-path encoder; A right-path DoubleConv module, configured to process the feature map output by the right-path encoder; The bottleneck fusion layer 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. a terminal enhancement module, which converts the concatenated features into a highly nonlinear representation; The CBAM enhancement module applies the attention mechanism to the output of the bottleneck layer for enhancement.

9. The method for identifying tree species in a stand based on multi-source remote sensing data fusion according to any one of claims 1 to 8, characterized in that: The forest remote sensing image data includes airborne digital images, satellite-borne images and CHM canopy height model data, and the preprocessing of the acquired data includes: Performing radiometric correction, atmospheric correction, orthorectification, and panchromatic fusion on the airborne digital image and the satellite-borne image; The CHM data is subjected to noise suppression, edge preservation and data normalization processing, and the processed CHM data is subjected to image registration processing with the optical image to achieve unified alignment of the spatial coordinate system.

10. A forest stand tree species identification system based on multi-source remote sensing data fusion, characterized in that: include: Data acquisition module, used to collect forest remote sensing image data and ground data; A data preprocessing module, used for preprocessing the forestland remote sensing image data and ground data collected by the data acquisition module; The prediction module predicts tree species classification based on the input image slices and obtains a tree species classification map; The image processing module performs splicing processing on the tree species classification map to obtain a complete tree species classification grid map.

Images