Forest tree species classification method and device, electronic equipment and storage medium

Abstract

The present application provides a kind of forest tree species classification method, device, electronic equipment and storage medium, belong to artificial intelligence technical field, method includes: forest remote sensing image is input to forest tree species classification model, obtains the output tree species classification result;Forest tree species classification model includes preliminary feature processing module, high-order feature extraction module and tree species classification result output module, high-order feature extraction module is sequentially connected first residual block stack layer, first down-sampling layer, second residual block stack layer, second down-sampling layer, third residual block stack layer, third down-sampling layer and fourth residual block stack layer, stacked residual block embedding dimension rearrangement layer, dimension recovery layer and convolution block attention module.The present application improves the dimension compatibility of CBAM and convolution architecture by embedding explicit dimension rearrangement layer and dimension recovery layer in the residual block embedded with CBAM, effectively improves the classification accuracy of forest remote sensing image for forest tree species classification task.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, electronic device, and storage medium for classifying forest tree species. Background Technology

[0002] Deep convolutional neural networks (CNNs) have become the mainstream tool for feature extraction from remote sensing images.

[0003] In the process of classifying tree species in tropical rainforests based on forest remote sensing images, existing deep learning backbone networks generally face the challenge of "insufficient feature discrimination ability".

[0004] Taking the ConvNeXt architecture in CNN as an example, after constructing a forest tree species classification model based on the ConvNeXt architecture, the accuracy of the tree species classification results obtained by inputting forest remote sensing images into the pre-trained forest tree species classification model is low. It has problems such as homogenized feature response and lack of adaptive calibration, weak ability to focus on spatial heterogeneity and low feature co-operation efficiency. Summary of the Invention

[0005] This invention provides a forest tree species classification method, apparatus, electronic device, and storage medium to address the shortcomings of existing technologies, such as lack of adaptive calibration of feature responses, weak focusing ability on spatial heterogeneity, and low feature coordination efficiency, thereby achieving a forest tree species classification scheme that improves the accuracy of tree species classification.

[0006] This invention provides a method for classifying forest tree species, comprising: The forest remote sensing image is input into the forest tree species classification model to obtain the tree species classification results output by the forest tree species classification model; The forest tree species classification model includes a preliminary feature processing module, a higher-order feature extraction module, and a tree species classification result output module. The preliminary feature processing module is used to extract an initial calibration feature map from the forest remote sensing image. The higher-order feature extraction module is used to extract a higher-order semantic feature map from the initial calibration feature map. The tree species classification result output module is used to determine the tree species classification result based on the higher-order semantic feature map. The high-order feature extraction module includes a first residual block stacked layer, a first downsampling layer, a second residual block stacked layer, a second downsampling layer, a third residual block stacked layer, a third downsampling layer, and a fourth residual block stacked layer connected in sequence. The residual blocks stacked in the first residual block stacking layer, the second residual block stacking layer, the third residual block stacking layer, and the fourth residual block stacking layer are embedded with a dimension rearrangement layer, a dimension restoration layer, and a convolutional block attention module; The dimension rearrangement layer is used to replace the tensor format of the input features of the dimension rearrangement layer from channel-first to channel-last; the dimension restoration layer is used to replace the tensor format of the input features of the dimension restoration layer from channel-last to channel-first.

[0007] According to a forest tree species classification method provided by the present invention, the residual block includes a depthwise separable convolutional layer, a first normalization layer, the dimension rearrangement layer, an inverted bottleneck structure layer, the dimension restoration layer, the convolutional block attention module, a residual connection layer, and a sub-output layer; The depthwise separable convolutional layer is used to extract a spatial context feature map with a large receptive field from the input feature map; The first normalization layer is used to perform a normalization operation on the pixel positions of the spatial context feature map to obtain a normalized spatial feature map; The dimension rearrangement layer is used to replace the tensor format of the normalized spatial feature map from channel-first to channel-last. The inverted bottleneck structure layer is used to perform depth enhancement processing on the normalized spatial feature map at the end of the channel in tensor format to obtain a depth spatial feature map. The dimension restoration layer is used to replace the tensor format of the depth spatial feature map from channel last to channel first. The convolutional block attention module is used to perform channel attention enhancement processing and spatial attention enhancement processing on the depth spatial feature map with channel-first tensor format to obtain an attention-enhanced feature map. The residual connection layer is used to perform residual connections between the attention-enhanced feature map and the input feature map of this residual block to obtain a refined residual feature map; The sub-output layer is used to output the fine residual feature map.

[0008] According to a forest tree species classification method provided by the present invention, the step of performing channel attention enhancement processing and spatial attention enhancement processing on the depth spatial feature map with channel-first tensor format includes: The depth spatial feature map is subjected to global pooling to obtain a global spatial aggregated feature map; The global spatial aggregated feature map is input into a multilayer perceptron to obtain a one-dimensional weight coefficient vector output by the multilayer perceptron. The one-dimensional weight coefficient vector is input into the Sigmoid activation function layer to obtain the key channel enhanced feature map output by the Sigmoid activation function layer. The key channel enhanced feature map is subjected to mean compression and maximum compression along the channel axis to obtain a two-dimensional feature map that reflects spatial saliency. Local context fusion is performed on the two-dimensional feature map to generate a spatial weight mask; Based on the spatial weight mask, the attention enhancement feature map is determined; The key channel enhancement feature map is a feature map that enhances the key spectral bands or specific phenological phase channels of the forest remote sensing image; the spatial weight mask has a higher response weight to tree species regions in the forest remote sensing image than to background noise.

[0009] According to a forest tree species classification method provided by the present invention, the kernel size of the depth-separable convolutional layer is determined based on the resolution of the forest remote sensing image.

[0010] According to a forest tree species classification method provided by the present invention, the first residual block stacking layer, the second residual block stacking layer and the fourth residual block stacking layer all stack a first number of the residual blocks; The third residual block stacking layer stacks a second number of the residual blocks; Wherein, the first quantity is less than the second quantity.

[0011] According to a forest tree species classification method provided by the present invention, the forest remote sensing images include multi-temporal time-series remote sensing images and / or sub-meter high-resolution remote sensing images.

[0012] According to a forest tree species classification method provided by the present invention, the preliminary feature processing module includes a preliminary feature extraction layer and a second normalization layer; The preliminary feature extraction layer is used to map the forest remote sensing image to a feature space of a preset depth to obtain an initial feature map; The second normalization layer is used to calibrate the channel feature distribution of the initial feature map to obtain the initial calibrated feature map.

[0013] The present invention also provides a forest tree species classification device, comprising: The classification model processing module is used to input forest remote sensing images into the forest tree species classification model and obtain the tree species classification results output by the forest tree species classification model. The forest tree species classification model includes a preliminary feature processing module, a higher-order feature extraction module, and a tree species classification result output module. The preliminary feature processing module is used to extract an initial calibration feature map from the forest remote sensing image. The higher-order feature extraction module is used to extract a higher-order semantic feature map from the initial calibration feature map. The tree species classification result output module is used to determine the tree species classification result based on the higher-order semantic feature map. The high-order feature extraction module includes a first residual block stacked layer, a first downsampling layer, a second residual block stacked layer, a second downsampling layer, a third residual block stacked layer, a third downsampling layer, and a fourth residual block stacked layer connected in sequence. The residual blocks stacked in the first residual block stacking layer, the second residual block stacking layer, the third residual block stacking layer, and the fourth residual block stacking layer are embedded with a dimension rearrangement layer, a dimension restoration layer, and a convolutional block attention module; The dimension rearrangement layer is used to replace the tensor format of the input features of the dimension rearrangement layer from channel-first to channel-last; the dimension restoration layer is used to replace the tensor format of the input features of the dimension restoration layer from channel-last to channel-first.

[0014] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the forest tree species classification method as described above.

[0015] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the forest tree species classification method as described above.

[0016] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the forest tree species classification method as described above.

[0017] The forest tree species classification method, apparatus, electronic device, and storage medium provided by this invention improve the dimensionality adaptability of CBAM to the convolutional architecture by embedding CBAM within all residual blocks stacked in the four hierarchical spatial evolution stages of the high-order feature extraction module of the forest tree species classification model, and simultaneously embedding explicit dimensional rearrangement layers and dimensional restoration layers within all residual blocks. This ensures the stability and effectiveness of gradient flow in the forest tree species classification model, effectively utilizing the channel and spatial dual-dimensional adaptive feature calibration capabilities of CBAM. On the one hand, it automatically identifies and dynamically increases the weight ratio of phenological feature bands that make key contributions in the high-dimensional phenological feature channels of multi-temporal satellite images, avoiding misclassification of forest tree species caused by the "different objects with the same spectrum" phenomenon. On the other hand, it automatically suppresses complex background noise interference such as shadows, bare land, and understory vegetation in high-resolution images, improving the focusing accuracy and anti-interference ability of high-resolution image details, and accurately capturing the edge contours and texture details of target tree crowns in remote sensing images, thereby improving the classification accuracy of forest tree species classification tasks based on forest remote sensing images. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the forest tree species classification method provided by the present invention.

[0020] Figure 2 This is a schematic diagram of the model structure of the forest tree species classification model provided by the present invention.

[0021] Figure 3 This is a schematic diagram of the residual block structure of the high-order feature extraction module provided by the present invention.

[0022] Figure 4 This is a schematic diagram of the downsampling layer of the high-order feature extraction module provided by the present invention.

[0023] Figure 5 This is a schematic diagram of the forest tree species classification device provided by the present invention.

[0024] Figure 6 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0026] The following is combined Figures 1 to 6 The present invention describes a method, apparatus, electronic device, and storage medium for classifying forest tree species.

[0027] The widespread application of remote sensing technology has led to a dramatic increase in the dimensionality and complexity of geospatial data. In practical classification tasks, the data sources relied upon are diverse, including single-temporal or multi-temporal multispectral data that focus on spectral temporal patterns (such as Sentinel-2), sub-meter high-resolution data that pursues fine spatial texture (such as WorldView), and multi-source heterogeneous fusion remote sensing image data constructed to balance spatial and spectral information.

[0028] In forest tree species classification algorithms, deep convolutional neural networks have become the mainstream tool for feature extraction. However, existing deep learning backbone networks generally face the challenge of "insufficient feature discrimination ability" when processing the aforementioned diverse types of remote sensing data.

[0029] Taking the ConvNeXt architecture in deep convolutional neural networks as an example, ConvNeXt is a modern pure convolutional network widely used in the field of general computer vision. It draws on the macroscopic design ideas of Vision Transformer (ViT). While retaining the inherent advantage of convolutional networks in efficiently extracting local features, it has obtained global context modeling capabilities similar to Transformer. It is one of the cutting-edge backbone networks for remote sensing image feature extraction.

[0030] In actual research and development and comparative experiments, the inventors found that the standard ConvNeXt architecture has limitations when dealing with the task of classifying tree species in tropical rainforests based on forest remote sensing images. These limitations include homogenized feature responses, lack of adaptive calibration, weak ability to focus on spatial heterogeneity, and low feature co-operation efficiency.

[0031] Specifically, the homogenization of feature responses and lack of adaptive calibration mean that the convolutional blocks of the ConvNeXt architecture perform undifferentiated weighted calculations on all input feature channels. Since forest remote sensing images often have highly overlapping spectra or redundant channels, the model cannot automatically identify and enhance the most discriminative tree species features, resulting in key information (such as specific phenological characteristics) being diluted by background noise. The weak ability to focus on spatial heterogeneity means that, due to the lack of spatial attention constraints, the ConvNeXt architecture has a relatively scattered response distribution in the receptive field when processing forest remote sensing images containing complex backgrounds (such as shadows, bare ground, and mixed pixels), making it difficult to adaptively lock onto the core region of the target ground feature (such as the tree canopy), thus limiting further improvement in classification accuracy. The low efficiency of feature collaboration means that, due to the lack of an adaptive calibration mechanism, the ConvNeXt architecture has difficulty effectively handling the dimensional differences and correlations between heterogeneous features, resulting in low classification accuracy.

[0032] In view of this, the present invention provides a method for classifying forest tree species to solve at least one of the above problems.

[0033] Figure 1 This is a flowchart illustrating the forest tree species classification method provided by the present invention, as shown below. Figure 1 As shown, the forest tree species classification method includes, but is not limited to, step 101.

[0034] It should be noted that the subject executing the forest tree species classification method provided by this invention can be a server, computer equipment, such as mobile phone, tablet computer, laptop computer, handheld computer, vehicle electronic equipment, wearable device, ultra-mobile personal computer (UMPC), netbook, or personal digital assistant (PDA), etc.

[0035] Step 101: Input the forest remote sensing image into the forest tree species classification model to obtain the tree species classification results output by the forest tree species classification model.

[0036] Figure 2 This is a schematic diagram of the model structure of the forest tree species classification model provided by the present invention, as shown below. Figure 2 As shown, the forest tree species classification model includes a preliminary feature processing module, a high-order feature extraction module, and a tree species classification result output module.

[0037] The preliminary feature processing module is used to extract initial calibration feature maps from forest remote sensing images.

[0038] The higher-order feature extraction module is used to extract higher-order semantic feature maps from the initial calibration feature map.

[0039] The tree species classification result output module is used to determine the tree species classification result based on the high-order semantic feature map.

[0040] The high-order feature extraction module includes a first residual block stacked layer, a first downsampling layer, a second residual block stacked layer, a second downsampling layer, a third residual block stacked layer, a third downsampling layer, and a fourth residual block stacked layer connected in sequence. The stacked residual blocks in the first, second, third, and fourth residual block stacked layers are embedded with a dimension rearrangement layer, a dimension restoration layer, and a convolutional block attention module (CBAM). The dimension rearrangement layer is used to replace the tensor format of the input features of the dimension rearrangement layer from channel-first to channel-last. The dimension restoration layer is used to replace the tensor format of the input features of the dimension restoration layer from channel-last to channel-first.

[0041] Forest remote sensing images are multi-source heterogeneous data with highly overlapping spectra or redundant channels.

[0042] Optionally, forest remote sensing imagery includes multi-temporal time-series remote sensing imagery (such as Sentinel-2) and / or sub-meter high-resolution remote sensing imagery (such as WorldView).

[0043] The first, second, third, and fourth residual block stacks are all residual block stacks with at least two residual blocks stacked on top of each other. The number of residual blocks stacked in each residual block stack can be set according to specific circumstances.

[0044] Specifically, in combination Figure 2 As shown, before processing the forest tree species classification task based on forest remote sensing images, a forest tree species classification model is pre-constructed. The pre-training of the forest tree species classification model is completed using multiple forest tree species classification training samples. Each forest tree species classification training sample includes a forest remote sensing image and its corresponding tree species category label.

[0045] When processing forest tree species classification tasks based on forest remote sensing images, the forest remote sensing images are input into a pre-trained forest tree species classification model. First, the preliminary feature processing module of the forest tree species classification model performs feature mapping and feature distribution calibration on the forest remote sensing images, extracting an initial calibration feature map from the forest remote sensing images. Then, the high-order feature extraction module of the forest tree species classification model extracts a high-order semantic feature map from the initial calibration feature map. Finally, the tree species classification result output module determines the tree species classification result based on the high-order semantic feature map and outputs it.

[0046] Furthermore, during the process of extracting high-order semantic feature maps from the initial calibration feature map by the high-order feature extraction module, four hierarchical spatial evolution stages of feature extraction are performed. Each of the four stages performs feature extraction processing by stacking a specific number of residual blocks. During the feature transition, the initial calibration feature map sequentially passes through the first residual block stacking layer, the first downsampling layer, the second residual block stacking layer, the second downsampling layer, the third residual block stacking layer, the third downsampling layer, and the fourth residual block stacking layer. The downsampling layer at each stage compresses the spatial resolution of the feature map layer by layer, while simultaneously increasing the channel dimensions of the feature map by gradients of 96, 192, 384, and 768. Thus, through the progressive progression of network layers, the hierarchical modeling of ground cover information in forest remote sensing images is achieved from the underlying geometric texture to the deep spectral semantics, extracting a 7×7×768 high-order semantic feature map from the 56×56×96 initial calibration feature map. The output dimension of the first residual block stack layer is 96, the output dimension of the second residual block stack layer is 192, the output dimension of the third residual block stack layer is 384, and the output dimension of the fourth residual block stack layer is 768.

[0047] Furthermore, considering that in the process of extracting high-order semantic feature maps from the initial calibration feature maps, the single data adaptation dimension of conventional high-order feature extraction modules leads to limitations in the collaborative processing of multi-source data of forest remote sensing images, lacks channel calibration capabilities, and is difficult to adaptively select the features that contribute most to classification from more than ten phenological phase channels of forest remote sensing images. There is a problem that the discrimination performance of multi-temporal spectral information or medium- and low-resolution heterogeneous data in forest remote sensing images is significantly reduced, and feature overfitting or dilution of spectral information by spatial details is likely to occur.

[0048] Therefore, convolutional block attention modules (CBAMs) are embedded inside all residual blocks stacked in each residual block stacking layer of the higher-order feature extraction module.

[0049] CBAM is an adaptive feature calibration module that combines channel and spatial dimensions. In remote sensing data processing, channel attention can be used to identify and enhance the spectral bands or phenological phases that contribute the most to tree species classification, while spatial attention helps to focus on target areas (such as tree canopies) in forest remote sensing images and suppress the interference of background noise such as shadows and bare ground.

[0050] By embedding CBAM into all residual blocks stacked in the stacked residual block layers of the high-order feature extraction module, and utilizing the adaptive feature calibration capability of CBAM in both channel and spatial dimensions, the strong correlation between forest remote sensing images in channel and spatial dimensions is fully considered during the feature extraction process. Even when facing complex land cover with high texture heterogeneity or severe spectral overlap, a robust discriminative feature space can be constructed for forest remote sensing images. This allows for the selection of the most contributing features for classification from more than ten phenological phase channels of forest remote sensing images, thereby improving discriminative performance and the ability to generalize to heterogeneous features.

[0051] Furthermore, when dealing with forest tree species classification tasks based on forest remote sensing images, the embedding of CBAM within the residual blocks of each residual block stacking layer in the high-order feature extraction module leads to unstable attention mechanisms in the model gradient flow and dimension adaptation issues in the convolutional architecture. Therefore, in addition to embedding CBAM, an adaptive tensor dimension permutation and recombination time series is designed within all residual blocks of each residual block stacking layer in the high-order feature extraction module. At the same time, a dimension reordering layer is embedded to replace the tensor format of the input features of the dimension rearrangement layer from channel-first to channel-last, and a dimension restoration layer is embedded to replace the tensor format of the input features of the dimension restoration layer from channel-last to channel-first.

[0052] By performing two explicit permutation operations, the format adaptation conflict problem can be resolved, thereby achieving deep decoupling and functional synergy between the convolutional block attention mechanism and normalization in the convolutional network architecture. This ensures the stability and effectiveness of gradient flow during model training, improves the training accuracy of the forest tree species classification model, and ultimately enhances the accuracy of forest tree species classification.

[0053] Optionally, the forest tree species classification model is implemented based on the ConvNeXt architecture.

[0054] Optionally, the first residual block stacking layer, the second residual block stacking layer, and the fourth residual block stacking layer all stack a first number of residual blocks; the third residual block stacking layer stacks a second number of residual blocks; wherein the first number is less than the second number.

[0055] Optionally, the number of residual blocks stacked in the first residual block stack layer, the second residual block stack layer, the third residual block stack layer, and the fourth residual block stack layer are 3, 3, 9, and 3, respectively.

[0056] The forest tree species classification method provided by this invention embeds CBAM within all residual blocks stacked in the four hierarchical spatial evolution stages of the high-order feature extraction module of the forest tree species classification model. Simultaneously, explicit dimensional rearrangement layers and dimensional restoration layers are embedded within all residual blocks to improve the dimensionality adaptability of CBAM to the convolutional architecture, ensuring the stability and effectiveness of gradient flow in the forest tree species classification model. This effectively utilizes the channel and spatial dual-dimensional adaptive feature calibration capabilities of CBAM. On one hand, it automatically identifies and dynamically increases the weight of key phenological feature bands in the high-dimensional phenological feature channels of multi-temporal satellite images, avoiding misclassification of forest tree species caused by the "different species with the same spectrum" phenomenon. On the other hand, it automatically suppresses complex background noise interference such as shadows, bare land, and understory vegetation in high-resolution images, improving the focusing accuracy and anti-interference capability of high-resolution image details. It accurately captures the edge contours and texture details of target tree crowns in remote sensing images, thereby improving the classification accuracy of forest tree species classification tasks based on forest remote sensing images.

[0057] In one embodiment, combined with Figure 2As shown, the tree species classification output module includes a globally average pooling layer, a normalization layer, a linear mapping layer, and an output layer connected in sequence. The globally average pooling layer performs global average pooling on the high-order semantic feature map, transforming it into a one-dimensional feature vector. The normalization layer normalizes the one-dimensional feature vector. The linear mapping layer, a fully connected layer adapted to the number of tree species categories in the classification results, performs linear mapping on the normalized one-dimensional feature vector. The output layer uses the Softmax activation function to output the predicted probability that each object to be classified in the forest remote sensing image belongs to a different specific tree species, thus determining the tree species classification result. The tree species classification result includes sub-classification results corresponding to each object to be classified in the forest remote sensing image, and each sub-classification result includes the predicted probability that an object to be classified belongs to a different specific tree species.

[0058] For example, the 768-dimensional high-order feature map output by the tree species classification result output module to the high-order feature extraction module is converted into a one-dimensional feature vector, and the tree species classification result is output, which is the predicted probability that each object to be classified belongs to a specific tree species.

[0059] Based on the above embodiments, as an optional embodiment, the residual block includes a depthwise separable convolutional layer, a first normalization layer, the dimension rearrangement layer, an inverted bottleneck structure layer, the dimension restoration layer, the convolutional block attention module, a residual connection layer, and a sub-output layer; The depthwise separable convolutional layer is used to extract a spatial context feature map with a large receptive field from the input feature map; The first normalization layer is used to perform a normalization operation on the pixel positions of the spatial context feature map to obtain a normalized spatial feature map; The dimension rearrangement layer is used to replace the tensor format of the normalized spatial feature map from channel-first to channel-last. The inverted bottleneck structure layer is used to perform depth enhancement processing on the normalized spatial feature map at the end of the channel in tensor format to obtain a depth spatial feature map. The dimension restoration layer is used to replace the tensor format of the depth spatial feature map from channel last to channel first. The convolutional block attention module is used to perform channel attention enhancement processing and spatial attention enhancement processing on the depth spatial feature map with channel-first tensor format to obtain an attention-enhanced feature map. The residual connection layer is used to perform residual connections between the attention-enhanced feature map and the input feature map of the current residual block to obtain a refined residual feature map; The sub-output layer is used to output the fine residual feature map.

[0060] Optionally, the normalization operator of the first normalization layer is any one of LayerNorm, GroupNorm, InstanceNorm, and adaptive normalization operator to adapt to the optimization needs of different computing platforms.

[0061] Specifically, Figure 3 This is a schematic diagram of the residual block structure of the high-order feature extraction module provided by the present invention, as shown below. Figure 3 As shown, when inputting the input feature map into the residual block, which serves as the basic building block, to achieve feature calibration, the input feature map is first input into the depthwise separable convolutional layer of the residual block. The depthwise separable convolutional layer extracts and outputs a spatial context feature map with a large receptive field from the input feature map, thereby realizing spatial statistical feature extraction.

[0062] The spatial context feature map output by the depthwise separable convolutional layer is input to the first normalization layer. The first normalization layer performs a normalization operation on the pixel positions of the spatial context feature map to obtain and output a normalized spatial feature map.

[0063] The normalized spatial feature map output from the first normalization layer is input to the dimension rearrangement layer. The dimension rearrangement layer replaces the tensor format of the normalized spatial feature map from channel-first to channel-last, achieving explicit alignment of tensor dimensions for the first time.

[0064] For example, the dimension rearrangement layer performs a permute(0,2,3,1) operation, replacing the tensor format from channel-first (NCHW) to channel-last (NHWC) to adapt to the precise normalization performed by the first normalization layer for each pixel position.

[0065] The tensor format is used to input the final normalized spatial feature map of the channel into the inverted bottleneck structure layer. The inverted bottleneck structure layer performs depth enhancement processing on the final normalized spatial feature map of the channel to obtain and output the depth spatial feature map.

[0066] The depth space feature map output by the inverted bottleneck structure layer is input to the dimension restoration layer. The dimension restoration layer replaces the tensor format of the depth space feature map from channel last to channel first, realizing the second explicit alignment of the tensor dimensions.

[0067] For example, the dimension restoration layer performs the permute(0,3,1,2) operation to restore the tensor format from the last channel format to the channel-first format.

[0068] The depth spatial feature map output by the dimension restoration layer, which is in channel-first format, is input to CBAM. CBAM performs adaptive feature calibration operations on the channel-first format depth spatial feature map, which is then processed by channel attention enhancement and spatial attention enhancement, to obtain and output the attention-enhanced feature map.

[0069] The attention-enhanced feature map output by CBAM is input to the residual connection layer. The residual connection layer performs element-wise addition fusion on the attention-enhanced feature map and the input feature map, and outputs the adaptively calibrated fine residual feature map.

[0070] Finally, the sub-output layers of the residual blocks are used to output fine residual feature maps.

[0071] Understandably, when the residual block is the first residual block in the stacked layers of the first residual block in the higher-order feature extraction module, the input feature map of the depthwise separable convolutional layer of this residual block is the initial calibration feature map. When the residual block is the last residual block in the stacked layers of the fourth residual block in the higher-order feature extraction module, the fine residual feature map of the sub-output layer of this residual block is the higher-order semantic feature map.

[0072] Optionally, performing channel attention enhancement processing and spatial attention enhancement processing on the depth spatial feature map with channel-first tensor format includes: performing channel attention enhancement processing on the depth spatial feature map to obtain a key channel enhanced feature map; and performing spatial attention enhancement processing on the key channel enhanced feature map to obtain the attention-enhanced feature map.

[0073] Optionally, performing channel attention enhancement processing and spatial attention enhancement processing on the depth spatial feature map with channel-first tensor format includes: performing spatial attention enhancement processing on the depth spatial feature map to obtain a spatially enhanced feature map; and performing channel attention enhancement processing on the spatially enhanced feature map to obtain the attention-enhanced feature map.

[0074] In one embodiment, combined with Figure 3 As shown, before inputting the deep spatial feature map in channel-first tensor format into CBAM, the deep spatial feature map in channel-first tensor format is input into a scaling layer connected to the dimension restoration layer. The scaling layer scales the deep spatial feature map to obtain a channel-first tensor format deep spatial feature map after layer scaling. This deep spatial feature map is further input into a droppath layer connected to the scaling layer. The droppath layer performs random depth regularization on the deep spatial feature map through the droppath strategy to obtain a channel-first tensor format deep spatial feature map after layer scaling and regularization, which is then input into CBAM.

[0075] In one embodiment, the inverted bottleneck structure layer includes a first fully connected layer, a GELU activation function layer, and a second fully connected layer connected in sequence; wherein, the first fully connected layer and the second fully connected layer are used to expand the channel dimension of the channel-based normalized spatial feature map in the tensor format by 4 times, and after being processed by the GELU activation function of the GELU activation function layer, it is mapped back to the original dimension, thereby enhancing the representation depth.

[0076] Optionally, the kernel size of both the first fully connected layer and the second fully connected layer is 1×1, and the stride is 1.

[0077] By extracting spatial features through depthwise separable convolution, the channel dimension is expanded to four times the original dimension through the inverted bottleneck structure of point convolution, and then mapped back to the original dimension after nonlinear activation, which helps to reduce information loss during feature transfer.

[0078] Furthermore, by employing a normalization operator other than the batch normalization operator and using the GELU activation function instead of the ReLU activation function, the normalization and activation mechanisms of the basic building blocks are optimized, thereby improving training stability and the nonlinearity of feature representation.

[0079] Compared to embedding CBAM in other locations within a forest tree species classification model, the forest tree species classification method provided in this invention, through internal architecture-level reshaping of residual blocks, sets each residual block to include a depthwise separable convolutional layer, a first normalization layer, a dimension rearrangement layer, an inverted bottleneck structure layer, a dimension restoration layer, a convolutional block attention module, a residual connection layer, and a sub-output layer. By placing the CBAM module after dimension restoration and before residual summation within the residual blocks of the convolutional network—that is, embedding CBAM after the feature extraction process of the residual blocks—the attention mechanism of CBAM can be focused on the fully extracted, complete, high-order features, thereby improving... The accuracy of spectral channel weight calibration and spatial target region focusing is improved. Furthermore, CBAM embedding before the residual connection can enhance the features of the residual branches without changing the main identity mapping path. Combined with the dimension rearrangement layer and dimension restoration layer, it can ensure smooth gradient backpropagation and stable training process, avoiding model convergence difficulties and accuracy decline caused by the attention mechanism interfering with the normalized statistical distribution or nonlinear mapping process. It achieves adaptive feature calibration and maximizes the feature calibration efficiency without destroying the original convolution structure features and computational logic, thereby improving the model's feature discrimination ability and tree species classification accuracy for complex forest remote sensing images.

[0080] Furthermore, by adopting an architecture-level reshaping scheme that combines layer normalization with dimension rearrangement, robust feature alignment can be achieved at the pixel scale, avoiding the introduction of statistical noise in the gradient backpropagation path and improving the convergence accuracy of the model for refined tree features.

[0081] As an optional embodiment, the kernel size of the depth-separable convolutional layer is determined based on the resolution of the forest remote sensing image, and the kernel size increases with increasing resolution.

[0082] For example, the kernel size can be 5×5, 7×7, or 9×9.

[0083] By determining the kernel size of depth-separable convolutional layers based on the resolution of forest remote sensing images, the local context capture range can be adaptively changed, thereby achieving adaptive adjustment of the receptive field range.

[0084] Optionally, the kernel size of the depth-separable convolutional layer is 7×7.

[0085] Optionally, the kernel size of the depth-separable convolutional layer is 7×7, the stride is 1, and the padding is 3.

[0086] For residual blocks using 3×3 small kernel convolutions, the receptive field needs to be expanded by repeatedly stacking 3×3 small kernel convolutions. This means that the model must go through extremely deep network layers to obtain sufficient contextual information to identify the macroscopic forest stand structure. The progressive evolution pattern causes the spatial attention weights to only act on local textures when dealing with complex tree species classification, resulting in a lag in the extraction of global discriminative features. Consequently, the forest tree species classification model has insufficient overall focusing ability on the target ground features, reducing the automation and robustness of feature interpretation.

[0087] By employing 7×7 large-kernel depthwise separable convolution at the input of the residual blocks, which serve as the basic building blocks, the effective receptive field can be expanded while significantly reducing the number of parameters. This allows for the capture of long-range spatial dependencies in images, enabling the establishment of global feature associations in the early stages of feature extraction. Spatial context information can be identified without the need for extremely deep network layers, thus improving the overall focusing ability of the forest tree species classification model on target ground features.

[0088] Based on the above embodiments, as an optional embodiment, the process of performing channel attention enhancement processing and spatial attention enhancement processing on the depth spatial feature map with a channel-first tensor format includes: The depth spatial feature map is subjected to global pooling to obtain a global spatial aggregated feature map; The global spatial aggregated feature map is input into a multilayer perceptron to obtain a one-dimensional weight coefficient vector output by the multilayer perceptron. The one-dimensional weight coefficient vector is input into the Sigmoid activation function layer to obtain the key channel enhanced feature map output by the Sigmoid activation function layer. The key channel enhanced feature map is subjected to mean compression and maximum compression along the channel axis to obtain a two-dimensional feature map that reflects spatial saliency. Local context fusion is performed on the two-dimensional feature map to generate a spatial weight mask; Based on the spatial weight mask, the attention enhancement feature map is determined; The key channel enhancement feature map is a feature map that enhances the key spectral bands or specific phenological phase channels of the forest remote sensing image; the spatial weight mask has a higher response weight to tree species regions in the forest remote sensing image than to background noise.

[0089] Specifically, in the process of CBAM performing channel attention enhancement and spatial attention enhancement on the channel-priority depth spatial feature map in tensor format, the channel attention subroutine is first used to perform channel attention enhancement on the depth spatial feature map to obtain the key channel enhanced feature map; then the spatial attention subroutine is used to perform spatial attention enhancement on the key channel enhanced feature map to obtain the attention-enhanced feature map.

[0090] In the channel attention subroutine, the channel-priority depth spatial feature map in tensor format undergoes global pooling, including global average pooling and global max pooling branches to aggregate global spatial information, resulting in a compressed global spatial aggregated feature map. This global spatial aggregated feature map is then input to a multi-layer perceptron (MLP) with shared weights for transformation. The MLP generates a one-dimensional weight coefficient vector by learning the nonlinear interdependencies between channels and outputs it. The one-dimensional weight coefficient vector output by the MLP is then input to a sigmoid activation function layer, which processes this vector and maps it back to the original feature stream, thereby enhancing key spectral bands or specific phenological phase channels. This results in a key channel enhanced feature map that suppresses redundant features. Therefore, the key channel enhanced feature map is a feature map that enhances key spectral bands or specific phenological phase channels in forest remote sensing images.

[0091] In the spatial attention subroutine, mean compression and maximum compression operations are performed along the channel axis on the enhanced feature maps of key channels that have completed channel calibration, generating two-dimensional feature maps that reflect spatial saliency. Local context fusion is then performed on the compressed two-dimensional feature maps using 7×7 convolutional kernels to generate a pixel-level directional spatial weight mask. This spatial weight mask assigns higher response weights to target tree species regions in the forest remote sensing image, resulting in an attention-enhanced feature map that effectively reduces the interference of background noise such as understory shadows and bare ground on feature discriminativeness. It is evident that the spatial weight mask assigns higher response weights to tree species regions in the forest remote sensing image than to background noise.

[0092] Optionally, global pooling processing includes: global average pooling, global max pooling, and / or global second-order pooling.

[0093] Global second-order pooling can capture higher-level channel correlations.

[0094] The forest tree species classification method provided by this invention embeds CBAM within all residual blocks stacked in the four hierarchical spatial evolution stages of the high-order feature extraction module of the forest tree species classification model. At the same time, explicit dimension rearrangement layers and dimension restoration layers are also embedded within all residual blocks to improve the dimensionality adaptability of CBAM to the convolutional architecture, ensuring the stability and effectiveness of gradient flow in the forest tree species classification model. This effectively utilizes the channel and spatial dual-dimensional adaptive feature calibration capability of CBAM to improve the classification accuracy of forest tree species classification tasks based on forest remote sensing images.

[0095] Based on the above embodiments, as an optional embodiment, the preliminary feature processing module includes a preliminary feature extraction layer and a second normalization layer; The preliminary feature extraction layer is used to map the forest remote sensing image to a feature space of a preset depth to obtain an initial feature map; The second normalization layer is used to calibrate the channel feature distribution of the initial feature map to obtain the initial calibrated feature map.

[0096] Optionally, the initial feature extraction layer uses a convolutional layer with a kernel size of 4×4 and a stride of 4.

[0097] Specifically, in combination Figure 2 As shown, the preliminary feature processing module includes a preliminary feature extraction layer and a second normalization layer connected in sequence. It has a channel adaptive mechanism and can receive heterogeneous or homogeneous data streams, thereby ensuring the model's compatibility with data from different sensor sources.

[0098] After inputting forest remote sensing images, which contain multi-source heterogeneous data including multi-temporal time-series remote sensing images and / or sub-meter high-resolution remote sensing images, into a pre-trained forest tree species classification model, the forest remote sensing images are first input into the preliminary feature extraction layer of the preliminary feature processing module. The preliminary feature extraction layer uses a convolution operator with a stride of 4 and a kernel size of 4×4 to perform spatial downsampling on the forest remote sensing images, which are used as the original tensor, thereby mapping the forest remote sensing images to a feature space of a preset depth, obtaining an initial feature map, and outputting it.

[0099] After the convolutional mapping is completed, the initial feature map output by the preliminary feature extraction layer is sequentially input into the second normalization layer (LayerNorm). The second normalization layer calibrates the channel feature distribution of each channel of the initial feature map to obtain the initial calibrated feature map and output it.

[0100] The forest tree species classification method provided by this invention, by constructing a trunk layer input architecture with adaptive channel dimensions and utilizing spatial resolution compression and feature dimension expansion, completes the alignment of forest remote sensing image data from different sources in terms of statistical dimensions in the initial feature extraction stage, providing an initial calibration feature map with consistent data representation for the subsequent high-order semantic feature extraction stage.

[0101] Optionally, Figure 4 This is a schematic diagram of the downsampling layer structure of the high-order feature extraction module provided by the present invention, combined with... Figure 2 and Figure 4 As shown, the initial feature extraction layer uses a convolutional layer with a kernel size of 4×4 and a stride of 4. The first downsampling layer, the second downsampling layer, and the third downsampling layer all include a normalization layer, a convolutional layer, and a downsampling output layer connected in sequence. The convolutional layer has a kernel size of 2×2 and a stride of 2.

[0102] By designing a 4×4 initial feature extraction layer with a step size of 4 and a downsampling layer with a step size of 2, the spatial resolution can be gradually reduced and the number of channels increased.

[0103] To better illustrate the effectiveness of the forest tree species classification method provided by this invention, an embodiment is provided below. A CBAM mechanism is integrated within the residual block structure of ConvNeXt, and a specific temporal dimension rearrangement operation is applied to achieve deep coupling between the normalization layer and the CBAM attention module. Table 1 shows the classification experimental results of the improved ConvNeXt building block architecture with integrated adaptive calibration mechanism based on multi-temporal Sentinel-2 data and sub-meter-level remote sensing images (WorldView). As shown in Table 1, this improved ConvNeXt building block architecture with integrated adaptive calibration mechanism has the technical effects of improving the accuracy of multi-temporal phenological feature mining, enhancing high-resolution spatial detail discrimination and noise suppression, and achieving higher accuracy in multi-source heterogeneous data collaboration in tropical rainforest tree species classification tasks.

[0104] Table 1

[0105] Specifically, to improve the accuracy of multi-temporal phenological feature mining, this invention performs weight calibration on multi-temporal channels through a channel attention mechanism. The overall accuracy on the multi-temporal Sentinel-2 dataset is 89.06%, which is higher than VGG-16's 76.95% (difference of 12.11%) and ResNet18's 85.55% (difference of 3.51%). Furthermore, this result outperforms GoogLeNet and DenseNet121's 85.16%. Compared to the original ConvNeXt (86.33%), the overall accuracy is improved by 2.73%. Experimental data demonstrate that the ConvNeXt Block With CBAM structure can extract the dynamic phenological features of forest stands evolving over time.

[0106] In terms of enhancing high-resolution spatial detail discrimination and noise suppression, the overall accuracy of this invention reaches 93.75% when processing WorldView sub-meter level imagery. Comparative experimental results show that this accuracy is higher than VGG-16's 89.06% and ResNet18's 91.41%, and also surpasses GoogLeNet's 91.80% and DenseNet121's 92.58%. The spatial attention submodule generates a saliency mask to adaptively suppress the response weights of understory shadows and non-target backgrounds, thereby accurately representing the spatial feature distribution of forest stand objects.

[0107] Regarding higher accuracy in noise suppression and multi-source heterogeneous data fusion, in multi-source fusion experiments using WorldView sub-meter level imagery and multi-temporal Sentinel-2 imagery, the overall accuracy of this invention reached 97.27%, with a Kappa coefficient of 0.967. Compared to the classification results of single-temporal Sentinel-2 data, the overall accuracy of this invention was improved by 46.48%. Under the same data fusion process, the accuracy improvements of the control models VGG-16, ResNet18, and the original ConvNeXt were 30.43%, 36.52%, and 39.20%, respectively. Experimental data show that the overall accuracy of this invention is higher than that of GoogLeNet (96.88%).

[0108] Figure 5 This is a schematic diagram of the forest tree species classification device provided by the present invention, as shown below. Figure 5 As shown, the forest tree species classification device includes, but is not limited to, the classification model processing module 501.

[0109] The classification model processing module 501 is used to input forest remote sensing images into the forest tree species classification model and obtain the tree species classification results output by the forest tree species classification model.

[0110] The forest tree species classification model includes a preliminary feature processing module, a higher-order feature extraction module, and a tree species classification result output module. The preliminary feature processing module is used to extract an initial calibration feature map from the forest remote sensing image. The higher-order feature extraction module is used to extract a higher-order semantic feature map from the initial calibration feature map. The tree species classification result output module is used to determine the tree species classification result based on the higher-order semantic feature map. The high-order feature extraction module includes a first residual block stacked layer, a first downsampling layer, a second residual block stacked layer, a second downsampling layer, a third residual block stacked layer, a third downsampling layer, and a fourth residual block stacked layer connected in sequence. The residual blocks stacked in the first residual block stacking layer, the second residual block stacking layer, the third residual block stacking layer, and the fourth residual block stacking layer are embedded with a dimension rearrangement layer, a dimension restoration layer, and a convolutional block attention module; The dimension rearrangement layer is used to replace the tensor format of the input features of the dimension rearrangement layer from channel-first to channel-last; the dimension restoration layer is used to replace the tensor format of the input features of the dimension restoration layer from channel-last to channel-first.

[0111] It should be noted that the forest tree species classification device provided by the present invention can execute the forest tree species classification method described in any of the above embodiments during specific operation, and this embodiment will not elaborate on this.

[0112] The forest tree species classification device provided by this invention embeds CBAM within all residual blocks stacked in the four hierarchical spatial evolution stages of the high-order feature extraction module of the forest tree species classification model. Simultaneously, it embeds explicit dimensional rearrangement layers and dimensional restoration layers within all residual blocks to improve the dimensionality adaptability of CBAM to the convolutional architecture, ensuring the stability and effectiveness of gradient flow in the forest tree species classification model. This effectively utilizes the channel and spatial dual-dimensional adaptive feature calibration capabilities of CBAM. On one hand, it automatically identifies and dynamically increases the weight of key phenological feature bands in the high-dimensional phenological feature channels of multi-temporal satellite images, avoiding misclassification of forest tree species caused by the "different objects with the same spectrum" phenomenon. On the other hand, it automatically suppresses complex background noise interference such as shadows, bare land, and understory vegetation in high-resolution images, improving the focusing accuracy and anti-interference capability of high-resolution image details. It accurately captures the edge contours and texture details of target tree crowns in remote sensing images, thereby improving the classification accuracy of forest tree species classification tasks based on forest remote sensing images.

[0113] Figure 6 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 6As shown, the electronic device may include: a processor 610, a communication interface 620, a memory 630, and a communication bus 640, wherein the processor 610, the communication interface 620, and the memory 630 communicate with each other through the communication bus 640. The processor 610 can call logical instructions in the memory 630 to execute the forest tree species classification method provided in any of the above embodiments. The forest tree species classification method includes, but is not limited to, the following steps: inputting forest remote sensing images into a forest tree species classification model to obtain tree species classification results output by the forest tree species classification model; wherein the forest tree species classification model includes a preliminary feature processing module, a high-order feature extraction module, and a tree species classification result output module; the preliminary feature processing module is used to extract an initial calibration feature map from the forest remote sensing image; the high-order feature extraction module is used to extract a high-order semantic feature map from the initial calibration feature map; and the tree species classification result output module is used to extract a high-order semantic feature map based on the high-order semantic feature map. The tree species classification result is determined by the image extraction; the high-order feature extraction module includes a first residual block stacked layer, a first downsampling layer, a second residual block stacked layer, a second downsampling layer, a third residual block stacked layer, a third downsampling layer, and a fourth residual block stacked layer connected in sequence; the residual blocks stacked in the first residual block stacked layer, the second residual block stacked layer, the third residual block stacked layer, and the fourth residual block stacked layer are embedded with a dimension rearrangement layer, a dimension restoration layer, and a convolutional block attention module; the dimension rearrangement layer is used to replace the tensor format of the input features of the dimension rearrangement layer from channel-first to channel-last; the dimension restoration layer is used to replace the tensor format of the input features of the dimension restoration layer from channel-last to channel-first.

[0114] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0115] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the forest tree species classification method provided in any of the above embodiments. The forest tree species classification method includes, but is not limited to, the following steps: inputting forest remote sensing images into a forest tree species classification model to obtain tree species classification results output by the forest tree species classification model; wherein, the forest tree species classification model includes a preliminary feature processing module, a high-order feature extraction module, and a tree species classification result output module; the preliminary feature processing module is used to extract an initial calibration feature map from the forest remote sensing images; the high-order feature extraction module is used to extract high-order features from the initial calibration feature map. The semantic feature map; the tree species classification result output module is used to determine the tree species classification result based on the high-order semantic feature map; the high-order feature extraction module includes a first residual block stacked layer, a first downsampling layer, a second residual block stacked layer, a second downsampling layer, a third residual block stacked layer, a third downsampling layer, and a fourth residual block stacked layer connected in sequence; the residual blocks stacked in the first residual block stacked layer, the second residual block stacked layer, the third residual block stacked layer, and the fourth residual block stacked layer are embedded with a dimension rearrangement layer, a dimension restoration layer, and a convolutional block attention module; the dimension rearrangement layer is used to replace the tensor format of the input features of the dimension rearrangement layer from channel-first to channel-last; the dimension restoration layer is used to replace the tensor format of the input features of the dimension restoration layer from channel-last to channel-first.

[0116] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the forest tree species classification method provided in any of the above embodiments. The forest tree species classification method includes, but is not limited to, the following steps: inputting forest remote sensing images into a forest tree species classification model to obtain tree species classification results output by the forest tree species classification model; wherein, the forest tree species classification model includes a preliminary feature processing module, a higher-order feature extraction module, and a tree species classification result output module; the preliminary feature processing module is used to extract an initial calibration feature map from the forest remote sensing images; the higher-order feature extraction module is used to extract a higher-order semantic feature map from the initial calibration feature map; and the tree species classification result is output. The module is used to determine the tree species classification result based on the high-order semantic feature map; the high-order feature extraction module includes a first residual block stacked layer, a first downsampling layer, a second residual block stacked layer, a second downsampling layer, a third residual block stacked layer, a third downsampling layer, and a fourth residual block stacked layer connected in sequence; the residual blocks stacked in the first residual block stacked layer, the second residual block stacked layer, the third residual block stacked layer, and the fourth residual block stacked layer are embedded with a dimension rearrangement layer, a dimension restoration layer, and a convolutional block attention module; the dimension rearrangement layer is used to replace the tensor format of the input features of the dimension rearrangement layer from channel-first to channel-last; the dimension restoration layer is used to replace the tensor format of the input features of the dimension restoration layer from channel-last to channel-first.

[0117] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0118] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0119] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method of classifying forest tree species, characterized by, include: Forest remote sensing images are input into a forest tree species classification model to obtain the tree species classification results output by the forest tree species classification model; the forest remote sensing images include multi-temporal time-series remote sensing images and / or sub-meter high-resolution remote sensing images; the tree species classification results include the sub-classification results corresponding to each object to be classified in the forest remote sensing images, and each sub-classification result includes the predicted probability that an object to be classified belongs to a different specific tree species; The forest tree species classification model includes a preliminary feature processing module, a higher-order feature extraction module, and a tree species classification result output module. The preliminary feature processing module is used to extract an initial calibration feature map from the forest remote sensing image. The higher-order feature extraction module is used to extract a higher-order semantic feature map from the initial calibration feature map. The tree species classification result output module is used to determine the tree species classification result based on the higher-order semantic feature map. The high-order feature extraction module includes a first residual block stacked layer, a first downsampling layer, a second residual block stacked layer, a second downsampling layer, a third residual block stacked layer, a third downsampling layer, and a fourth residual block stacked layer connected in sequence. The residual blocks stacked in the first residual block stacking layer, the second residual block stacking layer, the third residual block stacking layer, and the fourth residual block stacking layer are embedded with a dimension rearrangement layer, a dimension restoration layer, and a convolutional block attention module; The dimension rearrangement layer is used to replace the tensor format of the input features of the dimension rearrangement layer from channel-first to channel-last; the dimension restoration layer is used to replace the tensor format of the input features of the dimension restoration layer from channel-last to channel-first. The residual block includes a depthwise separable convolutional layer, a first normalization layer, the dimension rearrangement layer, an inverted bottleneck structure layer, the dimension restoration layer, the convolutional block attention module, a residual connection layer, and a sub-output layer; The depthwise separable convolutional layer is used to extract a spatial context feature map with a large receptive field from the input feature map; The first normalization layer is used to perform a normalization operation on the pixel positions of the spatial context feature map to obtain a normalized spatial feature map; The dimension rearrangement layer is used to replace the tensor format of the normalized spatial feature map from channel-first to channel-last. The inverted bottleneck structure layer is used to perform depth enhancement processing on the normalized spatial feature map at the end of the channel in tensor format to obtain a depth spatial feature map. The dimension restoration layer is used to replace the tensor format of the depth spatial feature map from channel last to channel first. The convolutional block attention module is used to perform channel attention enhancement processing and spatial attention enhancement processing on the depth spatial feature map with channel-first tensor format to obtain an attention-enhanced feature map. The residual connection layer is used to perform residual connections between the attention-enhanced feature map and the input feature map of this residual block to obtain a refined residual feature map; The sub-output layer is used to output the fine residual feature map.

2. The forest tree species classification method according to claim 1, characterized in that, The process of performing channel attention enhancement and spatial attention enhancement on the depth spatial feature map with channel-first tensor format includes: The depth spatial feature map is subjected to global pooling to obtain a global spatial aggregated feature map; The global spatial aggregated feature map is input into a multilayer perceptron to obtain a one-dimensional weight coefficient vector output by the multilayer perceptron. The one-dimensional weight coefficient vector is input into the Sigmoid activation function layer to obtain the key channel enhanced feature map output by the Sigmoid activation function layer. The key channel enhanced feature map is subjected to mean compression and maximum compression along the channel axis to obtain a two-dimensional feature map that reflects spatial saliency. Local context fusion is performed on the two-dimensional feature map to generate a spatial weight mask; Based on the spatial weight mask, the attention enhancement feature map is determined; The key channel enhancement feature map is a feature map that enhances the key spectral bands or specific phenological phase channels of the forest remote sensing image; the spatial weight mask has a higher response weight to tree species regions in the forest remote sensing image than to background noise.

3. The forest tree species classification method according to claim 1, characterized in that, The kernel size of the depth-separable convolutional layer is determined based on the resolution of the forest remote sensing image.

4. The forest tree species classification method according to any one of claims 1-3, characterized in that, The first residual block stacking layer, the second residual block stacking layer, and the fourth residual block stacking layer all stack a first number of the residual blocks; The third residual block stacking layer stacks a second number of the residual blocks; Wherein, the first quantity is less than the second quantity.

5. The forest tree species classification method according to claim 1, characterized in that, The preliminary feature processing module includes a preliminary feature extraction layer and a second normalization layer; The preliminary feature extraction layer is used to map the forest remote sensing image to a feature space of a preset depth to obtain an initial feature map; The second normalization layer is used to calibrate the channel feature distribution of the initial feature map to obtain the initial calibrated feature map.

6. A forest tree species classification device, characterized in that: The classification model processing module is used to input forest remote sensing images into a forest tree species classification model to obtain the tree species classification results output by the forest tree species classification model; the forest remote sensing images include multi-temporal time-series remote sensing images and / or sub-meter high-resolution remote sensing images; the tree species classification results include sub-classification results corresponding to each object to be classified in the forest remote sensing images, and each sub-classification result includes the predicted probability that an object to be classified belongs to a different specific tree species; The forest tree species classification model includes a preliminary feature processing module, a higher-order feature extraction module, and a tree species classification result output module. The preliminary feature processing module is used to extract an initial calibration feature map from the forest remote sensing image. The higher-order feature extraction module is used to extract a higher-order semantic feature map from the initial calibration feature map. The tree species classification result output module is used to determine the tree species classification result based on the higher-order semantic feature map. The high-order feature extraction module includes a first residual block stacked layer, a first downsampling layer, a second residual block stacked layer, a second downsampling layer, a third residual block stacked layer, a third downsampling layer, and a fourth residual block stacked layer connected in sequence. The residual blocks stacked in the first residual block stacking layer, the second residual block stacking layer, the third residual block stacking layer, and the fourth residual block stacking layer are embedded with a dimension rearrangement layer, a dimension restoration layer, and a convolutional block attention module; The dimension rearrangement layer is used to replace the tensor format of the input features of the dimension rearrangement layer from channel-first to channel-last; the dimension restoration layer is used to replace the tensor format of the input features of the dimension restoration layer from channel-last to channel-first. The residual block includes a depthwise separable convolutional layer, a first normalization layer, the dimension rearrangement layer, an inverted bottleneck structure layer, the dimension restoration layer, the convolutional block attention module, a residual connection layer, and a sub-output layer; The depthwise separable convolutional layer is used to extract a spatial context feature map with a large receptive field from the input feature map; The first normalization layer is used to perform a normalization operation on the pixel positions of the spatial context feature map to obtain a normalized spatial feature map; The dimension rearrangement layer is used to replace the tensor format of the normalized spatial feature map from channel-first to channel-last. The inverted bottleneck structure layer is used to perform depth enhancement processing on the normalized spatial feature map at the end of the channel in tensor format to obtain a depth spatial feature map. The dimension restoration layer is used to replace the tensor format of the depth spatial feature map from channel last to channel first. The convolutional block attention module is used to perform channel attention enhancement processing and spatial attention enhancement processing on the depth spatial feature map with channel-first tensor format to obtain an attention-enhanced feature map. The residual connection layer is used to perform residual connections between the attention-enhanced feature map and the input feature map of this residual block to obtain a refined residual feature map; The sub-output layer is used to output the fine residual feature map.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the forest tree species classification method as described in any one of claims 1 to 5.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the forest tree species classification method as described in any one of claims 1 to 5.

Images