A method and system for fine classification of tree crowns based on multi-scale texture fusion

Abstract

The application discloses a tree crown single-tree fine classification method and system based on multi-scale texture fusion, and the method comprises the following steps: acquiring tree crown remote sensing image data to be processed; based on the tree crown remote sensing image data to be processed, multi-scale features are synchronously captured by using parallel convolution layers with different expansion rates to obtain fused multi-scale texture features; according to the multi-scale texture features, noise features are suppressed by processing through a cascaded attention mechanism to obtain enhanced classification features; and according to the enhanced classification features, the image data is regularly segmented into a plurality of local area slices to output a tree crown single-tree fine classification result. Through multi-scale texture feature extraction and enhancement, combined with local slice fine-grained analysis, the application removes redundant noise and amplifies the slight differences between species, so that the precision and robustness of single-tree classification are improved.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing and smart forestry technology, and particularly relates to a method and system for fine classification of individual tree crowns based on multi-scale texture fusion. Background Technology

[0002] Accurate identification of tree species is a core research direction in the field of forestry remote sensing, and it is of great significance for forestry resource surveys, ecological environment construction, and the protection of famous and ancient trees. With the evolution of deep learning technology, tree species identification based on remote sensing imagery has evolved from general target classification to fine-grained image classification.

[0003] The research goal of fine-grained visual classification is to distinguish subclasses with subtle differences within the same broad category (such as different birds, tree species, car models, etc.). Its core challenge stems from the dual contradiction between inter-class subtlety and intra-class variability: subclass discriminative features are often limited to local fine structures (such as leaf texture or bird beak shape), while samples within the same class exhibit significant morphological differences due to pose, lighting, and environmental interference. Therefore, models must possess both the ability to capture subtle features and scene robustness. The development of fine-grained visual classification has gone through several important stages: from early strong supervision dependence to weak supervision feature enhancement, and then to the stage of deep architecture fusion and efficiency optimization. Current mainstream methods focus on lightweight architecture design, high-order relationship modeling, and cross-modal feature fusion.

[0004] Accurate identification of tree species is a key issue in the field of forestry remote sensing. Its goal is to distinguish between different tree species with highly similar visual features, and it is a typical problem in fine-grained image classification. In recent years, research progress in this field has been closely related to the evolution of deep learning methods.

[0005] In existing technologies, researchers initially attempted to transfer pre-trained convolutional neural networks (CNNs) to tree canopy remote sensing image classification tasks, and combined this with data augmentation strategies involving geometric and photometric transformations to improve the model's adaptability to real-world scenes. Subsequently, to address the need for fine-grained feature capture, specialized architectures such as Bilinear Squeeze Excitation Network (BiSENet) emerged, achieving information fusion through parallel subnetworks and bilinear pooling to enhance sensitivity to subtle differences. Furthermore, researchers have introduced the Transformer architecture into this field, embedding multi-scale feature fusion (MFF) mechanisms to improve the recognition performance of closely related tree species within the same family. Recently, multi-source data fusion (such as combining LiDAR point clouds with optical imagery) has also demonstrated the potential to improve classification accuracy.

[0006] Fine-grained classification of forest tree species faces the dual challenges of environmental complexity and biodiversity. In real forest scenarios, tree species identification needs to overcome dynamic factors such as changes in light intensity and shading interference, while also distinguishing closely related species with highly similar morphologies (such as pine and oak). This places higher demands on the discriminative ability of classification models.

[0007] Current research on tree species classification based on remote sensing technology mainly focuses on the inversion of vegetation parameters at the stand scale or the extraction of canopy outlines of individual trees, with less emphasis on fine-grained classification of high-resolution images at the canopy scale. At present, fine-grained classification in forestry mainly faces the following challenges: discriminative features exist at multiple scales, the field environment is complex and noisy, and there are large differences within classes and small differences between classes.

[0008] Therefore, although general fine-grained classification methods and existing deep learning-based forestry identification research have laid an important foundation, there is still a lack of specific improvement measures for canopy-scale images of the field environment under remote sensing, where there are large intra-class differences, small inter-class differences, and discriminative features existing at multiple scales. Many existing forestry classification methods are not specifically designed for fine-grained classification tasks and are often limited to coarse classification of tree species. Meanwhile, some attempts at fine-grained classification have shown limited discriminative ability when distinguishing visually similar and closely related species, mainly because they cannot comprehensively capture and model subtle but discriminative features across multiple scales. Summary of the Invention

[0009] To address the aforementioned technical problems, this invention proposes a method and system for fine classification of individual tree canopies based on multi-scale texture fusion. Addressing the issues of large intra-class differences, small inter-class differences, and difficulty in comprehensively capturing discriminative features in remote sensing tree canopy images of the field environment, this invention improves the accuracy and robustness of individual tree classification by extracting and enhancing multi-scale texture features, combined with fine-grained analysis of local slices, removing redundant noise, and amplifying subtle inter-species differences.

[0010] To achieve the above objectives, this invention provides a method for fine classification of individual tree crowns based on multi-scale texture fusion, comprising: Acquire the remote sensing image data of the tree canopy to be processed; Based on the tree canopy remote sensing image data to be processed, multi-scale features are captured synchronously using parallel convolutional layers with different dilation rates to obtain fused multi-scale texture features. Based on the multi-scale texture features, noise suppression features are processed through a cascaded attention mechanism to obtain enhanced classification features; Based on the enhanced classification features, the image data is regularly segmented into multiple local region slices, and the fine classification results of individual tree crowns are output.

[0011] Optionally, before simultaneously capturing multi-scale features using parallel convolutional layers with different dilation rates based on the tree canopy remote sensing image data to be processed, the following steps are included: preprocessing the tree canopy remote sensing image data to be processed to obtain the original feature map.

[0012] Optionally, based on the tree canopy remote sensing image data to be processed, multi-scale features are simultaneously captured using parallel convolutional layers with different dilation rates to obtain fused multi-scale texture features, including: The original feature maps are fed into three parallel dilated convolution paths, and spatial texture information under different receptive fields is captured by setting different dilation rates to obtain multi-scale feature maps. Based on the multi-scale feature map, the multi-scale feature map is concatenated by channels and then dimensionality is reduced using a 1×1 convolution kernel to obtain the preliminary fused features. Based on the preliminary features, gating weights are calculated using global average pooling and a fully connected network. After the gating weights are applied to the preliminary features, a weighted residual sum is performed with the original feature map to output texture enhancement features.

[0013] Optionally, based on the multi-scale texture features, the noise-suppressing features are processed through a cascaded attention mechanism to obtain enhanced classification features, including: Based on the multi-scale texture features, the input features are adaptively modulated using spatial and channel features through wood marker fusion to obtain wood marker fusion features; Based on the wood marker fusion features, global information is aggregated along the image height and width directions respectively through the canopy axial attention mechanism to obtain noise-suppressed classification features.

[0014] Optionally, based on the multi-scale texture features, adaptive modulation of spatial and channel features is performed on the input features through wood marker fusion to obtain wood marker fusion features including: Based on the multi-scale texture features, the features are projected into a triplet of query item, key item, and value item; Based on the query term, key term, and value term triple, feature modulation is performed using depthwise separable convolution to obtain the wood marker fusion feature. ; ; In the formula, Q , K , V These are the query item, key item, and value item, respectively. DWConv For depthwise separable convolution, For learnable scalar parameters, For the input feature map, For use in queries Q Spatial modulation function, For application to key K The channel modulation function, This is for element-wise multiplication.

[0015] Optionally, based on the wood marker fusion features, global information is aggregated along the image height and width directions respectively using a canopy axial attention mechanism to obtain noise-suppressed classification features, including: Based on the wood marker fusion features, global information is aggregated along the image height direction to generate a height attention map, and global information is aggregated along the image width direction to generate a width attention map; Based on the height attention map and the width attention map, the noise-suppressed classification features are obtained by multiplying the height attention map and the width attention map pixel by pixel.

[0016] Optionally, based on the enhanced classification features, the image data is regularly segmented into multiple local region slices, and the output of fine classification results for individual tree canopies includes: Based on the enhanced classification features, obtain the optimized feature map; The optimized feature map is resampled to a preset high-resolution space using an interpolation method to obtain a high-resolution feature map; Based on the high-resolution feature map, a sliding window operation with a convolution kernel is used to extract multiple overlapping local region labels from the high-resolution feature map to obtain a slice tensor; Based on the slice tensor, global labels obtained through global average pooling are embedded in the slice sequence to obtain a feature sequence that incorporates global context information; Based on the feature sequence that integrates global context information, a classifier outputs a fine classification result for individual trees in the tree canopy.

[0017] Optionally, a sliding window operation with convolutional kernels is used to extract multiple overlapping local region labels from the high-resolution feature map, resulting in a slice tensor including: Based on the high-resolution feature map, multiple overlapping local region labels are extracted using a convolution kernel sliding window operation; Based on the local region markers, the local region markers are uniformly scaled to a fixed size using bilinear interpolation to obtain standardized local markers. Based on the standardized local labels, the standardized local labels are reorganized into the slice tensor.

[0018] The present invention also provides a fine classification system for individual tree crowns based on multi-scale texture fusion, comprising: a data acquisition module, a texture analysis module, a feature enhancement module, and a slice classification module; The data acquisition module is used to acquire the tree canopy remote sensing image data to be processed; The texture analysis module is used to capture multi-scale features simultaneously using parallel convolutional layers with different dilation rates based on the tree canopy remote sensing image data to be processed, so as to obtain fused multi-scale texture features. The feature enhancement module is used to process the noise-suppressing features through a cascaded attention mechanism based on the multi-scale texture features to obtain enhanced classification features; The slice classification module is used to regularly segment the image data into multiple local region slices based on the enhanced classification features, and output the fine classification results of individual tree crowns.

[0019] Compared with the prior art, the present invention has the following advantages and technical effects: This invention effectively addresses the challenge of capturing subtle differences between tree species due to close phylogenetic relationships by employing parallel dilated convolution to capture multi-scale texture features. A cascaded attention mechanism is used to accurately identify and reduce interference features caused by shadows, background, and mutual occlusion, significantly reducing errors in remote sensing classification in field environments. Simultaneously, a canopy local slicing extraction module amplifies local discriminative features, while embedding global tokens to resolve intra-class variations caused by differences in growth environments in individual trees, ensuring the integrity of the classification logic. Attached Figure Description

[0020] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of a method for fine classification of individual tree crowns based on multi-scale texture fusion, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of the canopy texture analysis module according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the logic flow of the cascaded crown attention mechanism according to an embodiment of the present invention. Detailed Implementation

[0021] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0022] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0023] This embodiment proposes a fine-grained classification method for individual tree canopies based on multi-scale texture fusion, such as... Figure 1 As shown, the specific steps include: Acquire the remote sensing image data of the tree canopy to be processed; Based on the tree canopy remote sensing image data to be processed, multi-scale features are captured synchronously using parallel convolutional layers with different dilation rates to obtain fused multi-scale texture features. Based on the multi-scale texture features, noise suppression features are processed through a cascaded attention mechanism to obtain enhanced classification features; Based on the enhanced classification features, the image data is regularly segmented into multiple local region slices, and the fine classification results of individual tree crowns are output.

[0024] This embodiment specifically includes: S1. Parallel convolutional layers with different dilation rates are used to simultaneously capture multi-scale features, extract and fuse cross-scale canopy depth texture information; S2. The extracted multi-scale texture features are processed by a cascaded attention mechanism to suppress noise features representing background, shadow and occlusion, and obtain the enhanced classification feature space. S3. Regularly segment the image data into multiple local region slices, combine them with enhanced classification features, amplify the subtle local differences between species and reduce the differences within the same species caused by the growth environment, and output the fine classification results of individual tree crowns.

[0025] Furthermore, based on the tree canopy remote sensing image data to be processed, before simultaneously capturing multi-scale features using parallel convolutional layers with different dilation rates, the process includes: preprocessing the tree canopy remote sensing image data to be processed to obtain the original feature map.

[0026] Furthermore, based on the tree canopy remote sensing image data to be processed, multi-scale features are simultaneously captured using parallel convolutional layers with different dilation rates, resulting in fused multi-scale texture features including: The original feature maps are fed into three parallel dilated convolution paths, and spatial texture information under different receptive fields is captured by setting different dilation rates to obtain multi-scale feature maps. Based on the multi-scale feature map, the multi-scale feature map is concatenated by channels and then dimensionality is reduced using a 1×1 convolution kernel to obtain the preliminary fused features. Based on the preliminary features, gating weights are calculated using global average pooling and a fully connected network. After the gating weights are applied to the preliminary features, a weighted residual sum is performed with the original feature map to output texture enhancement features.

[0027] Specifically, S1 includes: S11. The input feature map is fed into three parallel dilated convolution paths. By setting different dilation rates, spatial texture information under different receptive fields is captured to obtain multi-scale feature maps. S12. Channel splicing is performed on the multi-scale feature map, and dimensionality reduction is performed using a 1×1 convolution kernel to obtain the preliminary fused features; S13. Calculate the gating weights using global average pooling and a fully connected network. Apply the gating weights to the preliminary features and sum the weighted residuals with the original input feature map to output texture enhancement features.

[0028] Furthermore, based on the multi-scale texture features, the noise-suppressing features are processed through a cascaded attention mechanism to obtain enhanced classification features, including: Based on the multi-scale texture features, the input features are adaptively modulated using spatial and channel features through wood marker fusion to obtain wood marker fusion features; Based on the wood marker fusion features, global information is aggregated along the image height and width directions respectively through the canopy axial attention mechanism to obtain noise-suppressed classification features.

[0029] Furthermore, based on the multi-scale texture features, adaptive modulation of spatial and channel features is performed on the input features through wood marker fusion to obtain wood marker fusion features including: Based on the multi-scale texture features, the features are projected into a triplet of query item, key item, and value item; Based on the query term, key term, and value term triple, feature modulation is performed using depthwise separable convolution to obtain the wood marker fusion feature. ; ; In the formula, Q , K , V These are the query item, key item, and value item, respectively. DWConv For depthwise separable convolution, For learnable scalar parameters, For the input feature map, For use in queries Q Spatial modulation function, For application to key K The channel modulation function, This is for element-wise multiplication.

[0030] Furthermore, based on the wood marker fusion features, global information is aggregated along the image height and width directions respectively using a canopy axial attention mechanism to obtain noise-suppressed classification features, including: Based on the wood marker fusion features, global information is aggregated along the image height direction to generate a height attention map, and global information is aggregated along the image width direction to generate a width attention map; Based on the height attention map and the width attention map, the noise-suppressed classification features are obtained by multiplying the height attention map and the width attention map pixel by pixel.

[0031] Specifically, canopy axial attention includes attention along the image height respectively. h and width w Directional aggregation of global information g=GAP (X) : ; The calculated high attention map a h (x) With width attention map a w (x) The final output is achieved through pixel-by-pixel multiplication: ; in, CAA(X) These are enhanced classification features after noise suppression.

[0032] Furthermore, based on the enhanced classification features, the image data is regularly segmented into multiple local region slices, and the output of fine classification results for individual tree canopies includes: Based on the enhanced classification features, obtain the optimized feature map; The optimized feature map is resampled to a preset high-resolution space using an interpolation method to obtain a high-resolution feature map; Based on the high-resolution feature map, a sliding window operation with a convolution kernel is used to extract multiple overlapping local region labels from the high-resolution feature map to obtain a slice tensor; Based on the slice tensor, global labels obtained through global average pooling are embedded in the slice sequence to obtain a feature sequence that incorporates global context information; Based on the feature sequence that integrates global context information, a classifier outputs a fine classification result for individual trees in the tree canopy.

[0033] Specifically, the optimized feature map is resampled to a preset high-resolution space using interpolation methods to preserve spatial details. Multiple overlapping local region labels are extracted from the high-resolution feature map using convolutional kernel sliding window operations and reorganized into slice tensors. Global tokens obtained through global average pooling are embedded in the slice sequence to maintain the association between local recognition and global context. Finally, a classifier is used to achieve fine recognition and classification results for individual tree canopies.

[0034] Furthermore, a sliding window operation with convolutional kernels is used to extract multiple overlapping local region labels from the high-resolution feature map, resulting in a slice tensor including: Based on the high-resolution feature map, multiple overlapping local region labels are extracted using a convolution kernel sliding window operation; Based on the local region markers, the local region markers are uniformly scaled to a fixed size using bilinear interpolation to obtain standardized local markers. Based on the standardized local labels, the standardized local labels are reorganized into the slice tensor.

[0035] Furthermore, the expansion rates of the parallel convolutional layers are preset to 1, 2, and 3 respectively to construct differentiated receptive fields; the cascaded canopy attention mechanism uses global average pooling combined with one-dimensional convolution to generate corresponding axial attention maps, which are used to accurately locate the canopy region in the spatial dimension and remove shadow interference; in the process of extracting local slices of the canopy, the local labels are uniformly scaled to a fixed size through bilinear interpolation, and then concatenated with global labels representing global context information to form the final feature vector used for fine classification.

[0036] The following is a detailed description of this embodiment with reference to the accompanying drawings: like Figure 1 As shown, this embodiment provides a method for fine classification of individual tree crowns based on multi-scale texture fusion, including the following steps: S1. Parallel convolutional layers with different dilation rates are used to simultaneously capture multi-scale features, extract and fuse cross-scale canopy depth texture information.

[0037] Specifically, S1 includes: S11. The input feature map is fed into three parallel dilated convolutional paths, and the dilation rates are preset to 1, 2 and 3 respectively to construct differential receptive fields and capture multi-scale feature maps under different receptive fields. F 1, F 2, F 3; Specifically, the calculation formula is as follows: ; ; ; in, W d1 , W d2 , W d3 These are the convolution weights.

[0038] S12. Perform channel concatenation on the multi-scale feature maps and use a 1×1 convolution kernel. Wf Dimensionality reduction and fusion are performed to obtain preliminary features containing global spatial distribution and local detailed textures. F fused ; ; S13. Calculate the gating weights using global average pooling and a fully connected network. g Apply the gating weights to F fused Then, a weighted residual sum is performed with the original input feature map to output the texture enhancement feature. X textured The specific process is as follows: ; ; Where σ is the Sigmoid activation function and α is a learnable scaling factor.

[0039] S2. The extracted multi-scale texture features are processed by a cascaded attention mechanism to suppress noise features representing background, shadow and occlusion, and obtain an enhanced classification feature space.

[0040] like Figure 3 As shown, in this embodiment, a cascaded canopy attention mechanism is used to process texture enhancement features. Specifically, the cascaded attention processing procedure is as follows: Adaptive modulation of spatial and channel features is performed on the input features through a wood marker fusion unit: ; In the formula, Q , K , V These are the query item, key item, and value item, respectively. DWConv For depthwise separable convolution, For learnable scalar parameters, For the input feature map, For use in queries Q Spatial modulation function, For application to key K The channel modulation function, For element-wise multiplication; Subsequently, along the image height h and width w The directions respectively aggregate global information g=GAP(X) Extract directional features and generate an axial attention map. a h (x) and a w (x) : ; The calculated high attention map a h (x) With width attention map a w (x) The final output is achieved through pixel-by-pixel multiplication: ; in, CAA(X) These are enhanced classification features after noise suppression.

[0041] S3. Regularly segment the image data into multiple local region slices, combine them with enhanced classification features, amplify the subtle local differences between species and reduce the differences within the same species caused by the growth environment, and output the fine classification results of individual tree crowns.

[0042] Specifically, the method for extracting the canopy slices includes: The optimized feature map is resampled to a preset high-resolution space using interpolation methods. R , R To preserve spatial details: ; A sliding window operation using convolutional kernels is used to extract multiple overlapping local region labels from a high-resolution feature map. U : ; in, k The kernel size is the convolution kernel size. c This is the output number of slices.

[0043] Then, bilinear interpolation is used to uniformly scale the local markers to a fixed size. p × p : ; Finally, a global token obtained through global average pooling is embedded in the slice sequence, and it is concatenated with the local region token sequence to maintain the association between local recognition and global context. Finally, a classifier is used to achieve fine recognition of individual trees in the canopy.

[0044] Example 2: This embodiment selects three representative canopy image datasets for verification: the NJFU dataset was collected from the campus of Nanjing Forestry University and contains a variety of typical southern street tree species; the SZU dataset was collected from the campus of Shenzhen University and covers common tree species in southern coastal cities; the ETH dataset comes from the public dataset of ETH Zurich and contains representative tree species of temperate forests.

[0045] This embodiment performs quantitative analysis on the three tree canopy image datasets (NJFU, SZU, ETH) mentioned above, quantitatively demonstrating their superiority in terms of accuracy, effectiveness, and robustness in fine classification of individual trees.

[0046] This embodiment quantitatively evaluates the recognition performance by testing the Top-1 Accuracy of different tree species. In Table 1, Methods represent the existing mainstream deep learning models used for comparison, including convolutional neural networks and Transformer architecture models; Top-1 Acc. (%) represents the fine-grained classification accuracy of each model in the corresponding scene. The classification method proposed in this invention achieves the best classification accuracy on all three datasets, indicating that this embodiment has a greater advantage than existing mainstream Transformer and CNN architectures when processing tree canopy images with subtle discriminative features.

[0047] Table 1 To verify the effectiveness of each key module, an ablation experiment was conducted in this embodiment. Table 2 shows the accuracy changes after progressively adding the canopy texture analysis module (CTA), the cascaded canopy attention mechanism (CCA), and the canopy slice extraction module (CPE) to the base model (BAP). This embodiment achieves optimal model performance when all three modules are integrated simultaneously, verifying the synergistic gain effect of multi-scale texture and attention mechanisms.

[0048] Table 2 Therefore, the multi-scale texture fusion method for fine classification of individual tree crowns proposed in this embodiment improves the accuracy, effectiveness, and completeness of identification and classification of closely related tree species in the field.

[0049] This embodiment also provides a fine classification system for individual tree crowns based on multi-scale texture fusion, including: a data acquisition module, a texture analysis module, a feature enhancement module, and a slice classification module; The data acquisition module is used to acquire the tree canopy remote sensing image data to be processed; The texture analysis module is used to capture multi-scale features simultaneously using parallel convolutional layers with different dilation rates based on the tree canopy remote sensing image data to be processed, so as to obtain fused multi-scale texture features. The feature enhancement module is used to process the noise-suppressing features through a cascaded attention mechanism based on the multi-scale texture features to obtain enhanced classification features; The slice classification module is used to regularly segment the image data into multiple local region slices based on the enhanced classification features, and output the fine classification results of individual tree crowns.

[0050] Specifically, such as Figure 2 As shown, the texture analysis module includes: a parallel path unit and a gating adjustment unit; The parallel path unit is used to capture multi-scale feature maps through convolutional paths with dilation rates of 1, 2, and 3. F 1, F 2, F 3.

[0051] The gating adjustment unit is used to achieve global average pooling ( GAP Generate modulation weights g The weighted residuals of the spliced ​​multi-scale features are summed based on the modulation weights to output texture enhancement features.

[0052] The feature enhancement module processes the extracted multi-scale texture features using a cascaded attention mechanism to suppress noise features representing background, shadows, and occlusions, thereby obtaining an enhanced classification feature space.

[0053] In this embodiment, a cascaded canopy attention mechanism is used to process the texture enhancement features, projecting the features as... Q , K , V Triples, and utilize depthwise separable convolution ( DWConv Perform feature modulation; extract directional features, and combine with 1 D Convolution and layer normalization generate attention maps, and the product of the two is used to filter out noise in non-canopy regions.

[0054] Specifically, the input features are adaptively modulated using spatial and channel features through a wood marker fusion unit: ; In the formula, Q , K , V These are the query item, key item, and value item, respectively. DWConv For depthwise separable convolution, For learnable scalar parameters, For the input feature map, For use in queries Q Spatial modulation function, For application to key K The channel modulation function, For element-wise multiplication; Subsequently, along the image height h and width w The directions respectively aggregate global information g=GAP(X) Extract directional features and generate an axial attention map. a h (x) anda w (x) : ; The calculated high attention map a h (x) With width attention map a w (x) The final output is achieved through pixel-by-pixel multiplication: ; in, CAA(X) These are enhanced classification features after noise suppression.

[0055] The slice classification module is used to regularly segment the image into multiple local region slices, and output the fine classification results of individual tree crowns based on the enhanced classification features.

[0056] The method for extracting the tree canopy slices includes: The optimized feature map is resampled to a preset high-resolution space using interpolation methods. R , R To preserve spatial details: ; A sliding window operation using convolutional kernels is used to extract multiple overlapping local region labels from a high-resolution feature map. U : ; in, k The kernel size is the convolution kernel size. c This is the output number of slices.

[0057] Then, bilinear interpolation is used to uniformly scale the local markers to a fixed size. p × p : ; Finally, global tokens obtained through global average pooling are embedded into the slice sequence. Specifically, feature tokens representing global context information are concatenated with local region token sequences to form a total of [number missing]. N The feature sequence with +1 is finally passed to the classifier to output a fine classification result.

[0058] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for fine classification of individual tree crowns based on multi-scale texture fusion, characterized in that, include: Acquire the remote sensing image data of the tree canopy to be processed; Based on the tree canopy remote sensing image data to be processed, multi-scale features are captured synchronously using parallel convolutional layers with different dilation rates to obtain fused multi-scale texture features. Based on the multi-scale texture features, noise suppression features are processed through a cascaded attention mechanism to obtain enhanced classification features; Based on the enhanced classification features, the image data is regularly segmented into multiple local region slices, and the fine classification results of individual tree crowns are output.

2. The method for fine classification of individual tree crowns based on multi-scale texture fusion according to claim 1, characterized in that, Based on the tree canopy remote sensing image data to be processed, the process of simultaneously capturing multi-scale features using parallel convolutional layers with different dilation rates includes: preprocessing the tree canopy remote sensing image data to be processed to obtain the original feature map.

3. The method for fine classification of individual tree crowns based on multi-scale texture fusion according to claim 2, characterized in that, Based on the tree canopy remote sensing image data to be processed, multi-scale features are simultaneously captured using parallel convolutional layers with different dilation rates, resulting in fused multi-scale texture features including: The original feature maps are fed into three parallel dilated convolution paths, and spatial texture information under different receptive fields is captured by setting different dilation rates to obtain multi-scale feature maps. Based on the multi-scale feature map, the multi-scale feature map is concatenated by channels and then dimensionality is reduced using a 1×1 convolution kernel to obtain the preliminary fused features. Based on the preliminary features, gating weights are calculated using global average pooling and a fully connected network. After the gating weights are applied to the preliminary features, a weighted residual sum is performed with the original feature map to output texture enhancement features.

4. The method for fine classification of individual tree crowns based on multi-scale texture fusion according to claim 2, characterized in that, Based on the multi-scale texture features, noise-suppressing features are processed through a cascaded attention mechanism to obtain enhanced classification features, including: Based on the multi-scale texture features, the input features are adaptively modulated using spatial and channel features through wood marker fusion to obtain wood marker fusion features; Based on the wood marker fusion features, global information is aggregated along the image height and width directions respectively through the canopy axial attention mechanism to obtain noise-suppressed classification features.

5. The method for fine classification of individual tree crowns based on multi-scale texture fusion according to claim 4, characterized in that, Based on the multi-scale texture features, adaptive modulation of spatial and channel features is performed on the input features through wood marker fusion to obtain wood marker fusion features including: Based on the multi-scale texture features, the features are projected into a triplet of query item, key item, and value item; Based on the query term, key term, and value term triple, feature modulation is performed using depthwise separable convolution to obtain the wood marker fusion feature. ; ； In the formula, Q , K , V These are the query item, key item, and value item, respectively. DWConv For depthwise separable convolution, For learnable scalar parameters, For the input feature map, For use in queries Q Spatial modulation function, For application to key K The channel modulation function, This is for element-wise multiplication.

6. The method for fine classification of individual tree crowns based on multi-scale texture fusion according to claim 4, characterized in that, Based on the wood marker fusion features, global information is aggregated along the image height and width directions using a canopy axial attention mechanism to obtain noise-suppressed classification features, including: Based on the wood marker fusion features, global information is aggregated along the image height direction to generate a height attention map, and global information is aggregated along the image width direction to generate a width attention map; Based on the height attention map and the width attention map, the noise-suppressed classification features are obtained by multiplying the height attention map and the width attention map pixel by pixel.

7. The method for fine classification of individual tree crowns based on multi-scale texture fusion according to claim 1, characterized in that, Based on the enhanced classification features, the image data is regularly segmented into multiple local region slices, and the fine classification results of individual tree canopies are output, including: Based on the enhanced classification features, obtain the optimized feature map; The optimized feature map is resampled to a preset high-resolution space using an interpolation method to obtain a high-resolution feature map; Based on the high-resolution feature map, a sliding window operation with a convolution kernel is used to extract multiple overlapping local region labels from the high-resolution feature map to obtain a slice tensor; Based on the slice tensor, global labels obtained through global average pooling are embedded in the slice sequence to obtain a feature sequence that incorporates global context information; Based on the feature sequence that integrates global context information, a classifier outputs a fine classification result for individual trees in the tree canopy.

8. The method for fine classification of individual tree crowns based on multi-scale texture fusion according to claim 7, characterized in that, A sliding window operation with convolutional kernels is used to extract multiple overlapping local region labels from the high-resolution feature map, resulting in a slice tensor including: Based on the high-resolution feature map, multiple overlapping local region labels are extracted using a convolution kernel sliding window operation; Based on the local region markers, the local region markers are uniformly scaled to a fixed size using bilinear interpolation to obtain standardized local markers. Based on the standardized local labels, the standardized local labels are reorganized into the slice tensor.

9. A fine-grained classification system for individual tree canopies based on multi-scale texture fusion, used to implement the method as described in any one of claims 1-8, characterized in that, include: Data acquisition module, texture analysis module, feature enhancement module, and slice classification module; The data acquisition module is used to acquire the tree canopy remote sensing image data to be processed; The texture analysis module is used to capture multi-scale features simultaneously using parallel convolutional layers with different dilation rates based on the tree canopy remote sensing image data to be processed, so as to obtain fused multi-scale texture features. The feature enhancement module is used to process the noise-suppressing features through a cascaded attention mechanism based on the multi-scale texture features to obtain enhanced classification features; The slice classification module is used to regularly segment the image data into multiple local region slices based on the enhanced classification features, and output the fine classification results of individual tree crowns.

Images