Construction method and system of fine tree species identification model, identification method and system of fine tree species identification model

Abstract

The invention discloses a construction method, a construction system, a recognition method and a recognition system for a fine tree species recognition model, belongs to the field of forest management and ecological protection, and solves the problems that the overall contribution of the whole group of time sequence features to model performance cannot be comprehensively examined, and due to the resolution limitation of remote sensing images and the randomness of forest species distribution space and quantity, the model performance cannot be comprehensively examined in the prior art. And the performance of the remote sensing tree species classification model is obviously reduced. The method comprises the steps of 1, constructing a high-dimensional multi-source feature data set; 2, performing feature optimization based on the high-dimensional multi-source feature data set to obtain an optimal classification feature set; 3, collecting actual measurement sample data, and constructing a deep learning sample library in combination with the optimal classification feature set; and step 4, presetting a fine tree species identification model, and inputting the deep learning sample library to train the preset fine tree species identification model to obtain the fine tree species identification model. The method is used for obtaining the accurate and newest tree species distribution map.

Description

Technical Field

[0001] The present invention relates to the field of forest management and ecological protection, and in particular to a method and system for constructing a fine tree species identification model, an identification method and a system thereof. Background Art

[0002] Forest ecosystems are the backbone of global terrestrial ecosystems. Accurately understanding the spatial distribution of forest types and tree species is crucial for forestry management, forest pest and fire monitoring, forest physiological parameter inversion, and the optimization of vegetation ecological models. Furthermore, a comprehensive understanding of existing forest vegetation types and their distribution is fundamental for global carbon monitoring, forest ecosystem service assessment, and biodiversity conservation. For ecologists and policymakers, obtaining accurate and up-to-date tree species distribution maps is a critical prerequisite for formulating relevant strategies and management measures.

[0003] Compared to time-consuming and expensive field sampling, remote sensing can efficiently and economically acquire high-temporal-resolution data over large areas, making it widely used in tree species classification. Currently, remote sensing-based tree species classification relies primarily on two main approaches: machine learning and deep learning. Traditional machine learning techniques excel at modeling nonlinear relationships in high-dimensional remote sensing data. However, these methods often rely on handcrafted feature sets and lack the ability to automatically extract hierarchical spatial and spectral representations. Furthermore, their generalizability and transferability to diverse data distributions and task settings are limited. To overcome these shortcomings, deep learning methods have been widely used in tree species identification in recent years. Compared to traditional methods, deep learning offers stronger capabilities for modeling spatial structure and automatically extracting multi-level features. Among these, convolutional neural networks (CNNs) are one of the most commonly used deep learning models. Classic architectures such as VGG (Visual Geometry Group), ResNet (Residual Network), DenseNet (Densely Connected Convolutional Neural Network), and EfficientNet (Efficient Neural Network) possess more complex, deep, nonlinear structures, achieving higher classification accuracy and enhanced robustness in diverse scenarios. However, these classical architectures still face challenges in accurately capturing fine-grained spatial features such as canopy boundaries and morphological details.

[0004] In the remote sensing tree species classification task, temporal features (such as multi-temporal spectral reflectance, vegetation index, etc.) can effectively reflect the changes in the spectral response of vegetation at different growth stages and are an important source of information for improving classification accuracy. However, directly extracting a large number of temporal features often brings two challenges: first, the number of features is huge and highly redundant, resulting in a significant increase in the computational overhead and complexity of the model; second, most existing feature selection methods are independently evaluated based on a single variable, ignoring the synergistic effects of the same type of features at different time points, making it difficult to reflect the overall contribution of the entire set of temporal features to the classification performance. This may not only mask the dynamic change patterns that are crucial to the discrimination results, but also reduce the stability and generalization ability of the model when dealing with seasonal changes and phenological differences. Therefore, there is an urgent need for a method that can evaluate the contribution of the feature group as a whole and effectively screen out high-value temporal features to balance model accuracy and computational efficiency.

[0005] In the prior art, Chinese patent document CN118228005A discloses a "forest tree species diversity monitoring method and device", which collects forest remote sensing images; obtains the spatiotemporal heterogeneity index of the forest based on the forest remote sensing images during the plant growing season as an image feature set; collects measured tree species diversity data; obtains the spatial matching relationship between the remote sensing image features and the diversity index based on the image feature set and the tree species diversity data; obtains input variables based on the spatial matching relationship between the remote sensing image features and the diversity index; and trains a preset deep learning model based on the input variables. According to the constructed model, the forest images of the test area are processed to obtain diversity results. However, this technical solution fails to comprehensively examine the overall contribution of the entire set of time series features to the model performance, and easily ignores the synergistic effect of features between different time points, thereby causing information loss in the feature selection process. In addition, due to the spatial resolution of Sentinel-2 images, fine structures in forest stands (such as crown morphology, branch and leaf texture, etc.) are difficult to distinguish accurately. In addition, the high heterogeneity in spatial distribution of forest tree species and random fluctuations in quantity make it difficult for deep learning models to effectively capture high-precision detail features, resulting in a decrease in the model's distinguishing ability in complex forest stands and significantly restricting the overall classification performance.

[0006] In summary, existing technologies fail to fully examine the overall contribution of the entire set of temporal features to model performance, resulting in information loss. Furthermore, due to limitations in the spatial resolution of Sentinel-2 images and the randomness of the spatial and quantitative distribution of forest tree species, deep learning models find it difficult to capture high-precision detail features, leading to a significant decline in the performance of remote sensing tree species classification models. Summary of the Invention

[0007] The present invention solves the technical problems that the existing technology fails to comprehensively examine the overall contribution of the entire set of temporal features to the model performance, and the performance of the remote sensing tree species classification model is significantly reduced because the deep learning model has difficulty in capturing the detailed features in the Sentinel-2 image.

[0008] The method for constructing a fine tree species identification model according to the present invention comprises the following steps: Step 1: Construct a high-dimensional multi-source feature dataset; Step 2: Feature selection is performed based on the high-dimensional multi-source feature dataset to obtain the optimal classification feature set; Step 3: Collect measured sample data, combine it with the optimal classification feature set, and build a deep learning sample library; Step 4: Preset a fine tree species recognition model, input a deep learning sample library to train the preset fine tree species recognition model, and obtain a fine tree species recognition model.

[0009] Furthermore, in an embodiment of the present invention, the high-dimensional multi-source feature dataset in step 1 includes principal component features, texture features, terrain features, spectral features and vegetation index.

[0010] Furthermore, in the embodiment of the present invention, in step 2, feature optimization is performed based on the high-dimensional multi-source feature data set to obtain the optimal classification feature set, specifically: The observation values ​​of each type of feature in the high-dimensional multi-source feature data set at different time phases are divided into a group to obtain multiple feature groups. The classification model is used to train the multiple feature groups to obtain the optimal classification feature set.

[0011] Furthermore, in the embodiment of the present invention, the classification model is used to train multiple feature groups to obtain the optimal classification feature set, specifically: The classification model is used to train multiple feature groups to obtain multiple single-channel feature importance indicators. All single-channel feature importance indicators in each feature group are aggregated to obtain the overall contribution of each feature group. The feature group with the largest overall contribution is taken as the optimal classification feature set.

[0012] Furthermore, in an embodiment of the present invention, the deep learning sample library in step 3 is divided into a training set and a validation set in a ratio of 8:2.

[0013] Furthermore, in the embodiment of the present invention, the preset fine tree species identification model in step 4 is specifically: Based on the data extraction module, multi-scale features are extracted from the input data, and the multi-scale features are aligned through the cross-scale deformable attention fusion module. The aligned multi-scale features are then fused to output the recognition results.

[0014] Furthermore, in the embodiment of the present invention, the cross-scale deformable attention fusion module is specifically: The multi-scale feature map is spatially offset, and the feature alignment of the spatially offset multi-scale feature map is performed using deformable convolution. The feature-aligned multi-scale feature map is feature spliced, and lightweight self-attention is used to obtain the self-attention result. The self-attention result is fused with the multi-scale feature map after feature splicing through dynamic channel gating to obtain the multi-scale feature map after feature fusion.

[0015] The method for fine tree species identification of the present invention is implemented based on any of the above methods for constructing a fine tree species identification model, specifically: Identify the forest images of the tested area and obtain detailed tree species identification results.

[0016] The system for constructing a fine tree species identification model of the present invention is implemented based on the above-mentioned method for constructing a fine tree species identification model and includes the following modules: Data collection module to build high-dimensional multi-source feature datasets; Feature optimization module, which performs feature optimization based on high-dimensional multi-source feature datasets to obtain the optimal classification feature set; The sample library construction module collects measured sample data and combines it with the optimal classification feature set to build a deep learning sample library; The training module presets a fine tree species recognition model, inputs a deep learning sample library to train the preset fine tree species recognition model, and obtains a fine tree species recognition model.

[0017] The fine tree species identification system of the present invention is implemented based on the fine tree species identification model construction system described above. The identification system includes: The recognition module recognizes the forest images of the test area and obtains fine tree species recognition results.

[0018] This invention solves the technical problems of existing technologies, which fail to fully examine the overall contribution of the entire set of temporal features to model performance, and the difficulty of deep learning models in capturing detailed features in Sentinel-2 imagery, which leads to a significant decline in the performance of remote sensing tree species classification models. Specific beneficial effects include: 1. A fine tree species identification method proposed by the present invention, such as Figure 1As shown in the figure, in the task of tree species identification in medium-resolution remote sensing images, a fusion-optimized ConvNeXt architecture was constructed around core issues such as improving the ability to extract spatial structural details and enhancing the ability to express spatiotemporal features. The classification accuracy of ConvNeXt was improved by fusing multi-scale feature maps, and the cross-scale high-resolution classification capability was improved while maintaining low computational overhead. In addition, the classification performance of remote sensing images under complex forest stand conditions was systematically optimized by combining multi-source feature fusion and feature optimization mechanisms and introducing temporal phenological information. This provides a lightweight, high-performance, and well-generalized fine tree species identification method for remote sensing tree species classification tasks, which is particularly suitable for fine-grained tree species identification and image segmentation tasks under complex forest stand structures. 2. This paper proposes a method for constructing a refined tree species identification model. It builds an end-to-end framework that integrates spatial, spectral, and temporal multidimensional features for tree species classification. By introducing joint training of temporal, spatial, and spectral features, it constructs a broadly adaptable refined tree species identification model, achieving a balance between classification accuracy and generalization capability. 3. The method for constructing a fine tree species identification model proposed in the present invention can overcome key technical bottlenecks such as the insufficient ability of traditional convolutional neural networks to express spatial details in forest type classification, and significantly improve the classification accuracy and reasoning efficiency of forest tree species identification in large-scale areas, enhance the deep perception ability of the spatial structure pattern of forest stands, and provide high-value parameter support for forest ecosystem monitoring and modeling, solving the problem of insufficient spatial detail information depiction of deep learning models in medium-resolution remote sensing images. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which: Figure 1 It is the tree species identification and mapping process based on the ConvNeXt architecture described in the invention content; Figure 2 It is the original framework of ConvNeXt described in the fifth embodiment; Figure 3 is a schematic diagram of the improved ConvNeXt described in the fifth embodiment; Figure 4 This is an example result of the forest tree species classification described in Implementation Method Seven. DETAILED DESCRIPTION

[0020] The following will clearly and completely describe various embodiments of the present invention in conjunction with the accompanying drawings. The embodiments described with reference to the accompanying drawings are exemplary and intended to be used to explain the present invention, but should not be understood as limiting the present invention.

[0021] Implementation method 1. The method for constructing a fine tree species identification model according to this implementation method comprises the following steps: Step 1: Construct a high-dimensional multi-source feature dataset; Step 2: Feature selection is performed based on the high-dimensional multi-source feature dataset to obtain the optimal classification feature set; Step 3: Collect measured sample data, combine it with the optimal classification feature set, and build a deep learning sample library; Step 4: Preset a fine tree species recognition model, input a deep learning sample library to train the preset fine tree species recognition model, and obtain a fine tree species recognition model.

[0022] Existing technologies have the technical problem of failing to comprehensively examine the overall contribution of the entire set of temporal features to model performance, and because deep learning models find it difficult to capture the detailed features in Sentinel-2 images, the performance of remote sensing tree species classification models has been significantly reduced.

[0023] In order to solve the above technical problems, this embodiment provides a method for constructing a fine tree species identification model, which specifically includes the following steps: Step 1: Construct a high-dimensional multi-source feature dataset; This implementation constructs a high-dimensional multi-source feature dataset based on Sentinel-2 multi-temporal images to improve the accuracy and stability of forest tree species classification.

[0024] Step 2: Feature selection is performed based on the high-dimensional multi-source feature dataset to obtain the optimal classification feature set; Step 3: Collect measured sample data, combine it with the optimal classification feature set, and build a deep learning sample library; In order to construct high-quality and representative pure forest training samples, this implementation method combines the 2018 forest category II survey small-scale vector data provided by the Tahe Forestry Bureau with high-resolution remote sensing images, and selects six major dominant tree species as classification targets. Given that the survey data does not record the spatial distribution of different tree species in mixed forests, it is difficult to achieve pixel-level accurate labeling, so only pure forests are selected as the source of training and validation samples. In each type of pure forest, a random sampling strategy is adopted to extract samples from the 15-meter buffer zone inside the polygon to avoid noise introduced by spatial positioning errors and inaccurate boundaries. The collected measured sample data is combined with the optimal feature set obtained by screening in step 2 to construct a deep learning sample library for model training. Considering that the deep learning sample library is limited in size and prone to model overfitting problems, the samples in the deep learning sample library are further cropped into 256×256 pixel image blocks, and the deep learning sample library is expanded to 100,000 through data enhancement methods such as vertical flipping, horizontal flipping, random rotation, and channel permutation.

[0025] Step 4: Preset a fine tree species recognition model, input a deep learning sample library to train the preset fine tree species recognition model, and obtain a fine tree species recognition model.

[0026] Therefore, this embodiment constructs a high-dimensional multi-source feature data set, performs feature optimization on it, combines it with measured sample data to form a deep learning sample, and uses the deep learning sample to train a preset fine tree species recognition model to obtain a fine tree species recognition model. This solves the technical problems that the existing technology fails to comprehensively examine the overall contribution of the entire set of time series features to the model performance, and due to the randomness of the distribution space and number of forest tree species, the performance of the remote sensing tree species classification model is significantly reduced.

[0027] Implementation method 2. This implementation method further limits the method for constructing a fine tree species identification model described in implementation method 1. The high-dimensional multi-source feature dataset in step 1 includes principal component features, texture features, terrain features, spectral features and vegetation index.

[0028] Because tree species identification is highly sensitive to seasonal spectral differences, introducing phenological information can help improve the performance of multispectral remote sensing data in tree species classification. In particular, Sentinel-2, with its high temporal resolution and multi-band imaging capabilities, can effectively capture the dynamic changes in vegetation reflectance characteristics during different phenological stages. To maximize the application potential of Sentinel-2 time series in tree species classification, this implementation method selects image data covering seven key time nodes throughout the growing season, namely April to October, and introduces observational images from the same period in the two years before and after to construct a time series dataset with low cloud interference and high data integrity. To systematically evaluate the contribution of different time series data to classification results, this implementation method proposes an importance aggregation method based on the gradient boosting tree model for importance assessment and screening of remote sensing multi-temporal feature groups, ultimately obtaining the most discriminative time series features. By integrating comprehensive features from the temporal, spatial, and spectral dimensions, an optimal classification feature set is constructed. This not only enhances the model's perception of phenological changes but also improves the accuracy of its depiction of fine-grained forest structural information. Ultimately, the spatially continuous, temporally intensive, and feature-fused tree species identification framework constructed based on Sentinel-2 time series data demonstrated good adaptability and application potential in addressing challenges such as complex species mixing and strong distribution heterogeneity in temperate mixed forests.

[0029] Furthermore, to fully utilize temporal information, the Sentinel-2 vegetation index time series from April 1 to October 31 was introduced to enhance the ability to identify phenological changes. To improve the classification accuracy of the fine tree species identification system in evaluating tree species with different characteristic types, the most discriminative feature combinations were obtained based on the feature screening of the RE-RFE (Random Forest Recursive Feature Elimination) algorithm. The four feature combination schemes are as follows: (1) Basic solution: fusion of original spectra, independent principal components and vegetation indices; (2) Texture enhancement scheme: Introducing texture features into the basic scheme; (3) Terrain expansion scheme: adding terrain variables based on the above; (4) Time series fusion solution: further integrate time series features.

[0030] The above four types of feature combination schemes help to systematically analyze the performance differences and complementarities of different feature dimensions in tree species identification, and provide a theoretical basis for constructing the optimal classification feature set.

[0031] First, independent principal component analysis (ICA) was used to reduce the dimensionality of imagery at seven key time points, yielding three independent non-Gaussian components. Forty-eight texture features were extracted from the ICA components, covering eight categories: mean, variance, contrast, entropy, correlation, homogeneity, angular second moment, and dissimilarity. Window sizes were set to 3×3 and 5×5, and the grayscale quantization level was 64. To enhance the ability to distinguish spectral differences and mitigate environmental variability, 41 vegetation indices were calculated, covering key bands such as the red edge, near-infrared, and shortwave infrared. Furthermore, three topographic factors—elevation, slope, and aspect—were calculated based on the SRTM 30m DEM (Global Digital Elevation Model) to characterize the ecological gradient effects of forest spatial distribution. Ultimately, principal component features, texture features, vegetation indices, spectral features, and topographic features were integrated into a unified multidimensional feature space, resulting in a high-dimensional, multi-source feature dataset. Topographic features include elevation, slope, and aspect. The following table shows the vegetation indices calculated based on the Sentinel-2 imagery.

[0032] Table 1

[0033] in, 、 、 、 、 、 They are the blue, green, red, red edge 1, red edge 2 and near-infrared bands in Sentinel-2 images.

[0034] Implementation method three. This implementation method is a further limitation of the method for constructing a fine tree species identification model described in Implementation method one. In step 2, feature optimization is performed based on a high-dimensional multi-source feature data set to obtain the optimal classification feature set, specifically: The observation values ​​of each type of feature in the high-dimensional multi-source feature dataset at different time phases are divided into a group to obtain multiple feature groups. The classification model is used to train the multiple feature groups to obtain multiple single-channel feature importance indicators. All single-channel feature importance indicators in each feature group are aggregated to obtain the overall contribution of each feature group. The feature group with the largest overall contribution is taken as the optimal classification feature set.

[0035] Implementation method 2 extracts multiple groups of time series features to enhance the model's expressive power, but large feature sets often have serious information redundancy, which not only increases the complexity of the model but may also weaken the model's classification performance. Traditional feature selection methods often construct a time series feature set first, and then perform feature selection on all time series features (each feature is an independent variable). This fails to comprehensively examine the overall contribution of the entire set of time series features to the model's performance, which can easily lead to the neglect of the synergistic effects between time series features and the inability to effectively capture key dynamic time series patterns, thereby affecting the model's ability to stably identify changes in different growth stages and seasons. The results are manifested in a decrease in the model's generalization ability and a blurred boundary between classes. In classification tasks of complex ecological environments and highly similar tree species, misclassification and unstable predictions are prone to occur. To address the above problems, this implementation method proposes an importance aggregation method based on a gradient boosting tree model for importance assessment and screening of remote sensing multi-temporal feature groups.

[0036] This implementation uses the physically meaningful feature categories of the vegetation index as units, grouping the observations of each feature category across multiple time phases. All groups are then fed into the XGBoost (gradient boosting classification model) for training. After model training, the output of the single-channel feature importance index is used to aggregate the channel importance indexes within each feature group to determine the overall contribution of each feature group. All feature groups are then sorted to identify the temporal feature groups with the greatest impact on classification performance. Based on this, an optimal classification feature set is constructed to improve model performance and simplify feature dimensionality. Table 2 shows the final selection of the 16 most discriminative features for subsequent modeling.

[0037] Table 2

[0038] in, is the Sentinel-2 red edge position index, is the excess greenness index, To convert the chlorophyll absorption ratio index, is the terrestrial chlorophyll index, and is the gray level index, representing the rows and columns in the gray level co-occurrence matrix, corresponding to the two gray values ​​of the pixels in the image, respectively. is the expected value or Mean, that is, the weighted average gray value of the gray-level co-occurrence matrix.

[0039] Implementation method 4. This implementation method further limits the method for constructing a fine tree species identification model described in implementation method 1. In step 3, the deep learning sample library is divided into a training set and a validation set in a ratio of 8:2.

[0040] Implementation 5. This implementation is a further limitation of the method for constructing a fine tree species identification model described in Implementation 1. The preset fine tree species identification model in step 4 is specifically: Based on the data extraction module, multi-scale features are extracted from the input data, and the multi-scale features are aligned through the cross-scale deformable attention fusion module. The aligned multi-scale features are then fused to output the recognition results.

[0041] Due to the limitations of medium-resolution remote sensing imagery in representing fine-grained information such as tree crown outlines, boundaries, and texture, CNNs face significant challenges in classification accuracy and class separability when used in temperate forests with a high species mix and complex species distribution. To address these issues, this implementation introduces a structurally optimized ConvNeXt as the model backbone. This structurally optimized ConvNeXt represents a systematic reconstruction of the traditional CNN.

[0042] Inspired by the Swin Transformer (layered vision using shifted windows), ConvNeXt systematically optimizes the ResNet-50 (Residual Network-50) infrastructure, combining powerful local detail capture capabilities with global context modeling capabilities, significantly enhancing the model's ability to capture local details and global context features. Figure 2As shown, a ConvNeXt (convolutional neural network) is used to extract multi-scale features from input data. Initial image encoding is performed using a 7×7 convolution kernel, effectively expanding the receptive field and enhancing global context perception. Parameter size is controlled by compressing the number of underlying convolution channels. During the feature extraction stage, depthwise separable convolution and pointwise convolution are introduced to improve computational efficiency while maintaining high expressiveness. Subsequently, a multi-scale fusion module, inspired by the multi-resolution parallel structure and multi-scale fusion concept of HRNet (High Resolution Network), is introduced to form a fused feature tensor. This fused feature tensor exhibits multi-scale and multi-level contextual feature expression. However, the fusion process cannot automatically align semantics and details across different scales and lacks a globally adaptive channel weight allocation mechanism. To address these technical issues, this implementation proposes a cross-scale deformable attention fusion module (CSDAF) based on deformable convolution and lightweight multi-head self-attention. This module aligns multi-scale features and fuses them together to output the recognition result. This implementation draws on the multi-scale fusion idea of ​​the multi-resolution parallel structure in HRNet, but no longer adopts a multi-branch stacking approach. Instead, it is based on the existing multi-stage features of ConvNeXt, aligning spatial positions through deformable convolution, achieving cross-scale interaction through lightweight multi-head attention, and dynamic fusion through adaptive channel gating. In this way, without introducing additional branches, it achieves feature enhancement effects similar to HRNet, while significantly reducing model parameters and computational costs.

[0043] In summary, this implementation uses the deep learning sample library constructed in step 3 as input, and proposes an improved model that integrates the ConvNeXt backbone network and the CSDAF module based on the Pytorch (deep learning framework) architecture in the Python (computer programming language) programming language to improve classification accuracy and spatial positioning capabilities. Figure 3As shown in the figure, the CSDAF module is integrated to achieve efficient interaction and dynamic alignment of multi-resolution features. It can adaptively capture key information flows between different scales and enhance the collaborative expression of spatial details and semantic information. Furthermore, the SE channel attention mechanism is introduced in the key feature extraction stage to adaptively assign weights to channel features, enhance discriminative feature expression, and suppress redundant information. This architecture achieves precise fusion of cross-scale information and fine-tuning of channel features while maintaining low computational overhead, effectively improving pixel-level classification accuracy and model robustness in complex scenarios. The improved ConvNeXt is applied to Sentinel-2 data, leveraging its unique remote sensing capabilities, including rich spectral bands, high spatial resolution, and short access period. The improved ConvNeXt enables high-precision forest tree species mapping and dynamic monitoring over long time series and large scales. This more effectively combines Sentinel-2 imagery with deep learning architectures to achieve long-term, large-scale forest tree species mapping and monitoring. This improved ConvNeXt significantly improves the spatial alignment and semantic interaction of multi-scale features without significantly increasing model size. Among them, S1, S2, S3 and S4 are the feature maps extracted by the ConvNeXt model from the first stage to the fourth stage, and X1, X2, X3 and X4 are the feature maps after upsampling of the feature maps in each stage.

[0044] Implementation 6. This implementation further limits the method for constructing a fine tree species recognition model described in Implementation 5. The cross-scale deformable attention fusion module is specifically: The multi-scale feature map is spatially offset, and the feature alignment of the spatially offset multi-scale feature map is performed using deformable convolution. The feature-aligned multi-scale feature map is feature spliced, and lightweight self-attention is used to obtain the self-attention result. The self-attention result is fused with the multi-scale feature map after feature splicing through dynamic channel gating to obtain the multi-scale feature map after feature fusion.

[0045] First, a small convolutional subnet offsets the convolution to spatially offset the multi-scale features output by each stage of ConvNeXt, and uses deformable convolution to align local details of different resolutions. The aligned features are then spliced ​​together, and lightweight self-attention with channel compression is used to capture cross-scale long-range dependencies. The self-attention results are then fused with the original spliced ​​features through dynamic channel gating.

[0046] Implementation VII. The fine tree species identification method described in this implementation is based on the fine tree species identification model construction method described in any one of Implementations 1 to 6, specifically: Identify the forest images of the tested area and obtain detailed tree species identification results.

[0047] like Figure 4 As shown, the fine tree species recognition results of the output area are identified by the fine tree species recognition model and further migrated to tree species recognition in different areas.

[0048] To evaluate the effectiveness of the proposed refined tree species identification model for temperate forest tree species classification, we calculated four metrics: F1 score, OA (overall accuracy), PA (producer accuracy), and UA (user accuracy). These four metrics collectively provide a systematic quantitative basis for evaluating the model's predictive performance and accuracy, ultimately resulting in a robust method for identifying forest tree species.

[0049] In summary, the present invention not only has important economic value in optimizing forest restoration projects and estimating ecological carbon sinks, but also has important application prospects in the fields of forest management and ecological protection.

[0050] Implementation 8. The fine tree species identification model construction system described in this implementation is implemented based on the fine tree species identification model construction method described in Implementation 1, and includes the following modules: Data collection module to build high-dimensional multi-source feature datasets; Feature optimization module, which performs feature optimization based on high-dimensional multi-source feature datasets to obtain the optimal classification feature set; The sample library construction module collects measured sample data and combines it with the optimal classification feature set to build a deep learning sample library; The training module presets a fine tree species recognition model, inputs a deep learning sample library to train the preset fine tree species recognition model, and obtains a fine tree species recognition model.

[0051] Implementation method 9. The fine tree species identification system described in this implementation method is based on the fine tree species identification model construction system of implementation method 8, and the identification system includes: The recognition module recognizes the forest images of the test area and obtains fine tree species recognition results.

[0052] The above is a detailed introduction to the method for constructing a fine tree species identification model and its system, the identification method and its system proposed in the present invention. Specific examples are used in this article to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea; at the same time, for general technical personnel in this field, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

Claims

1. A method for constructing a fine tree species identification model, characterized in that: The following steps are involved: Step 1: Construct a high-dimensional multi-source feature dataset; Step 2: Feature selection is performed based on the high-dimensional multi-source feature dataset to obtain the optimal classification feature set; Step 3: Collect measured sample data, combine it with the optimal classification feature set, and build a deep learning sample library; Step 4: Preset a fine tree species recognition model, input a deep learning sample library to train the preset fine tree species recognition model, and obtain a fine tree species recognition model.

2. The method for constructing a fine tree species identification model according to claim 1, characterized in that: The high-dimensional multi-source feature dataset in step 1 includes principal component features, texture features, terrain features, spectral features and vegetation index.

3. The method for constructing a fine tree species identification model according to claim 1, characterized in that: In step 2, feature selection is performed based on a high-dimensional multi-source feature dataset to obtain the optimal classification feature set, specifically: The observation values ​​of each type of feature in the high-dimensional multi-source feature data set at different time phases are divided into a group to obtain multiple feature groups. The classification model is used to train the multiple feature groups to obtain the optimal classification feature set.

4. The method for constructing a fine tree species identification model according to claim 3, characterized in that: The classification model is used to train multiple feature groups to obtain the optimal classification feature set, specifically: The classification model is used to train multiple feature groups to obtain multiple single-channel feature importance indicators. All single-channel feature importance indicators in each feature group are aggregated to obtain the overall contribution of each feature group. The feature group with the largest overall contribution is taken as the optimal classification feature set.

5. The method for constructing a fine tree species identification model according to claim 1, characterized in that: In step 3, the deep learning sample library is divided into a training set and a validation set in a ratio of 8:

2.

6. The method for constructing a fine tree species identification model according to claim 1, characterized in that: The preset fine tree species identification model in step 4 is specifically: Based on the data extraction module, multi-scale features are extracted from the input data, and the multi-scale features are aligned through the cross-scale deformable attention fusion module. The aligned multi-scale features are then fused to output the recognition results.

7. The method for constructing a fine tree species identification model according to claim 6, characterized in that: The cross-scale deformable attention fusion module is specifically: The multi-scale feature map is spatially offset, and the feature alignment of the spatially offset multi-scale feature map is performed using deformable convolution. The feature-aligned multi-scale feature map is feature spliced, and lightweight self-attention is used to obtain the self-attention result. The self-attention result is fused with the multi-scale feature map after feature splicing through dynamic channel gating to obtain the multi-scale feature map after feature fusion.

8. A method for fine tree species identification, wherein the method is implemented based on the method for constructing a fine tree species identification model according to any one of claims 1 to 7, and is characterized in that: Specifically: Identify the forest images of the tested area and obtain detailed tree species identification results.

9. A fine tree species identification model construction system, the system being implemented based on the fine tree species identification model construction method according to claim 1, characterized in that: Includes the following modules: Data collection module to build high-dimensional multi-source feature datasets; Feature optimization module, which performs feature optimization based on high-dimensional multi-source feature datasets to obtain the optimal classification feature set; The sample library construction module collects measured sample data and combines it with the optimal classification feature set to build a deep learning sample library; The training module presets a fine tree species recognition model, inputs a deep learning sample library to train the preset fine tree species recognition model, and obtains a fine tree species recognition model.

10. A fine tree species identification system, which is implemented based on the fine tree species identification model construction system according to claim 9, and is characterized in that: The identification system comprises: The recognition module recognizes the forest images of the test area and obtains fine tree species recognition results.

Images