Muti-spectral tree species classification method based on separable atrous spatial pyramid pooling

Abstract

The application discloses a multispectral tree species classification method based on a separable cavity ASPP, and relates to the technical field of semantic segmentation; in the application, a PCA process is combined with ANR, high-frequency noise is removed on the basis of principal components of data through an adaptive technique, utilization of near-infrared band information is enhanced, the quality of data is improved, and the feature expression of tree species to be classified is further enhanced; in the application, a basic DeepLabV3+ model is improved for semantic segmentation, a plurality of modules including an EnhancedMobileNetV2 module, an SEBlock module, a SeparableASPP module and an FPN Fusion feature fusion module are introduced, and accurate classification of different tree species is realized.

Description

Technical Field

[0001] This invention relates to the field of semantic segmentation technology, specifically to a multispectral tree species classification method based on separable void ASPP. Background Technology

[0002] In the field of forest resource monitoring, accurate tree species classification is of great significance for ecological research, rational planning of forest resources, and sustainable development. Different tree species play different functions in the ecosystem. Accurate identification and classification of tree species can provide key scientific basis for quantifying the ecological contribution of tree species and optimizing forest community structure, thereby promoting the scientific management and protection of forest ecosystems.

[0003] Traditional forest surveys rely on manual fieldwork, which suffers from long cycles, difficulty in accurate statistics in densely wooded areas, and inability to monitor large areas of forest resources in a timely and comprehensive manner. The application of deep learning algorithms has brought new opportunities, with the widespread use of DeepLabv3+ networks and the provision of abundant data for tree species classification by UAV multispectral imagery. However, existing technologies still face several challenges: traditional tree species identification methods are inefficient and inaccurate; the small sample datasets of deep learning models result in insufficient generalization ability; and the DeepLabv3+ network models suffer from high computational costs, insufficient recognition accuracy, and poor feature fusion performance. Therefore, a new method is urgently needed to address these issues and improve the accuracy and efficiency of tree species classification. Summary of the Invention

[0004] The purpose of this invention is to provide a multispectral tree species classification method based on separable void ASPP, in order to solve the problems raised in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a multispectral tree species classification method based on separable void ASPP, the classification method comprising the following steps:

[0006] Step S1: Acquire tree species images using a drone; perform data preprocessing on the tree species images to obtain label images;

[0007] Step S2: Perform data augmentation on the labeled images to obtain a sample dataset;

[0008] Step S3: Introduce the EnhancedMobileNetV2 module, SEBlock module, SeparableASPP module, and FPNFSion feature fusion module into the basic DeepLabV3+ model to obtain the optimized DeepLabV3+ model; input the sample dataset into the optimized DeepLabV3+ model to obtain the semantic segmentation result.

[0009] Furthermore, step S1 includes:

[0010] Step S1-1: Collect tree species images using a drone, and stitch the tree species images collected by the drone together to obtain a complete tree species image in TIFF format;

[0011] Step S1-2: The complete tree species image is preprocessed using a combination of principal component analysis and adaptive noise reduction techniques;

[0012] Step S1-3: Crop the preprocessed complete tree species image according to the preset first size;

[0013] Steps S1-4: Label the cropped complete tree species image with data, and convert the labeled complete tree species image from TIFF format to JSON format to obtain the label image;

[0014] Because multispectral images consist of multiple spectral bands, each corresponding to a specific spectral range, including visible light (red, green, and blue), near-infrared, and other bands; in multispectral images, the near-infrared band contains a large amount of useful information, but the near-infrared band is also easily affected by noise. In order to effectively utilize the information in the near-infrared band, this invention uses principal component analysis and adaptive noise removal (PCA-ANR) to process the band information of multispectral images.

[0015] Furthermore, the preprocessing of the complete tree species image includes:

[0016] The data of the complete tree species image are standardized; the covariance between every two bands is calculated to obtain the covariance matrix, and the covariance matrix is ​​decomposed into eigenvalues ​​to obtain eigenvalues;

[0017] Principal components with values ​​greater than the eigenvalues ​​are selected for data reconstruction to remove noise, so that the complete tree species image is displayed in near-infrared mode.

[0018] The above standardization process for the complete tree species image data aims to make the mean of each band 0 and the variance 1. Principal components with eigenvalues ​​greater than the eigenvalues ​​are selected, while those with smaller eigenvalues ​​are discarded, because the principal components with smaller eigenvalues ​​correspond to noise, while the principal components with larger eigenvalues ​​correspond to meaningful information. This process results in a greater weight for the near-infrared band, allowing the image to be displayed in a near-infrared manner.

[0019] Furthermore, step S2 includes:

[0020] Step S2-1: Using bilinear interpolation, the complete tree species image and label image cropped according to the first size are further cropped according to the preset second size. The complete tree species image and label image cropped according to the second size are simultaneously converted to JPG and PNG formats to obtain scaled image data and label data.

[0021] Step S2-2: Perform linear weighted synthesis on the image data and the label data using Mixup technology to obtain a synthesized image; perform geometric transformation on the synthesized image to obtain a sample dataset;

[0022] The above steps mainly expand the dataset sample size for the labeled images preprocessed using the PCA-ANR method. The Mixup technique selects two images, randomly generates a scaling factor (the scaling factor ranges between 0 and 1), mixes the two images proportionally through linear interpolation, and then performs linear weighted synthesis of the two images, with the target label also being weighted. These steps are used to enhance the diversity of the dataset and avoid overfitting.

[0023] Furthermore, step S3 includes:

[0024] The sample dataset is processed by the EnhancedMobileNetV2 module to extract low-level and high-level features; the SEBlock module processes the extracted features by adjusting the feature responses of each channel; the features processed by the SEBlock module are passed to the SeparableASPP module, which uses a preset depthwise separable dilated convolution to extract spatial context information; the FPNFusion feature fusion module generates a feature map by upsampling and fusing shallow and deep features layer by layer; the feature map is then processed by the DeepLabV3+ decoder to obtain the semantic segmentation result.

[0025] Furthermore, the EnhancedMobileNetV2 module includes

[0026] Low-level and high-level features are extracted using an inverse residual structure. The inverse residual structure expands the input feature map using a 1×1 convolutional kernel through an expansion layer, then performs a depthwise separable convolution operation on the expanded feature map using a 3×3 convolutional kernel. A dimensionality reduction layer performs a linear projection operation on the feature map after the depthwise separable convolution operation using a 1×1 convolutional kernel. The 1×1 convolutional kernel of the dimensionality reduction layer does not contain an activation function. The depthwise separable convolutional layer performs independent convolution operations on each input channel using depthwise convolution, and maps the output of the depthwise convolution to a new channel space using a 1×1 pointwise convolution.

[0027] The EnhancedMobileNetV2 module mentioned above is an enhanced version based on the MobileNetV2 module. It reduces the computational cost of the network by using inverted residual structures and depthwise separable convolutions. SEBlock is added to both shallow and deep features to enhance the channel weights and improve the model's attention to important features.

[0028] Furthermore, the SEBlock module is integrated into the low-level and high-level features of the EnhancedMobileNetV2 module backbone network; the SEBlock module processes the extracted features through the following steps:

[0029] Step S7-1: Compress the input feature map in the spatial dimension by global average pooling to generate global statistics for each channel;

[0030] Step S7-2: Learn the relationship between channels through two convolutional layers; wherein, the first convolutional layer uses the reduction parameter to control the compression ratio to reduce the number of channels of the global statistics and uses the ReLU activation function to strengthen the channels; the second convolutional layer restores the number of channels of the global statistics to the initial size and uses the activation function to limit the channel weights to [0,1] to obtain the channel weights of each channel;

[0031] Step S7-3: Use the obtained channel weights to perform weighted processing on the input feature map;

[0032] The aforementioned SEBlock module is integrated into the low-level and high-level features of the EnhancedMobileNetV2 module backbone network, which can enhance the expressive power of the feature maps.

[0033] Furthermore, the SeparableASPP module includes:

[0034] Step S8-1: Extract multi-scale features using depthwise separable convolution; wherein, depthwise separable convolution consists of depthwise convolution and pointwise convolution; wherein, depthwise convolution is performed on each input channel; pointwise convolution mixes information between channels using a 1×1 convolution kernel;

[0035] Step S8-2: Obtain global context information through global average pooling, and perform convolution and upsampling processing; wherein, by calculating the average value of all pixels in each channel, the spatial dimension information is compressed into a single value, thereby extracting global context information; after global average pooling, the input feature map retains only one average value for each output channel, and each average value represents the global information of the corresponding channel; after global average pooling, the features are processed by a 1×1 convolution kernel and upsampled to the initial size;

[0036] Step S8-3: Concatenate the multi-scale features extracted by depthwise separable convolution with the global context information obtained by global average pooling along the channel dimension to obtain the concatenated features; then fuse the concatenated features using a 1×1 convolution kernel.

[0037] The above steps use depthwise separable convolution to extract multi-scale features. This mainly involves using depthwise separable convolution with different dilation rates (such as 6, 12, and 18) to extract multi-scale features, thereby capturing features from spatial regions of different scales.

[0038] Furthermore, the FPNFusion feature fusion module is used to fuse the features output by the SeparableASPP module and the EnhancedMobileNetV2 module, including the following steps:

[0039] Step S9-1: Process the low-level features using a 1×1 convolution kernel to convert the number of channels of the low-level features from 24 to 48, so that the number of channels matches the number of channels of the high-level features;

[0040] Step S9-2: Upsample using bilinear interpolation to upsample high-level features to the same spatial resolution as low-level features; use a concatenation function to concatenate the high-level and low-level features to obtain the fused feature map;

[0041] Step S9-3: The fused feature maps are fused using a 3×3 convolution kernel to obtain a feature map with spatial details and semantic information;

[0042] Since the semantic features at higher levels have smaller spatial resolution, while the spatial detail features at lower levels have larger spatial resolution, the bilinear interpolation method is used for upsampling in the above steps.

[0043] Compared with the prior art, the beneficial effects of the present invention are:

[0044] 1. The image preprocessing part adopts principal component analysis and adaptive noise removal (PCA-ANR). Compared with traditional methods such as median filtering and wavelet transform, it effectively removes interference caused by sensor noise and outliers, and dynamically adjusts parameters adaptively to match the spatial heterogeneity of remote sensing images. This method can more effectively remove noise interference in the near-infrared band, making the information in the near-infrared band more significant, thereby better preserving band information and improving computational efficiency.

[0045] 2. This invention uses bilinear interpolation, the Mixup method, and geometric transformations to augment the dataset. Bilinear interpolation not only reduces image size but also smooths pixel values, avoiding coarse distortion during image scaling. This ensures that even after image reduction, the image maintains high quality, facilitating subsequent processing. Mixup is used to blend two different images, generating multiple new samples with different combinations of image features and labels. This method increases the diversity of training data and enhances the robustness of the model by generating synthetic samples. This invention fully leverages the advantages of bilinear interpolation and the Mixup method, providing stronger generalization ability and robustness during data augmentation.

[0046] 3. This invention introduces the EnhancedMobileNetV2 module, SEBlock module, SeparableASPP module, and FPNFusion feature fusion module into the basic DeepLabV3+ model. The EnhancedMobileNetV2 module is used as the backbone neural network. Compared to the traditional MobileNetV2 module, the enhanced module helps the network focus on channels with important information and ignore less useful channels, thereby improving model performance. The SEBlock module is introduced to enhance the model's expressive power by adaptively readjusting channel feature responses, allowing the model to focus more on useful features and suppress less useful ones. The SeparableASPP module replaces the traditional ASPP module. Compared to the traditional ASPP module, depthwise separable convolution not only reduces computation and improves classification performance but also effectively extracts features with different scales. The FPNFusion feature fusion module outputs feature maps containing rich information to enhance the model's segmentation performance. Attached Figure Description

[0047] Figure 1 This is a structural diagram of the improved DeepLabV3+ model based on the multispectral tree species classification method of separable void ASPP according to the present invention.

[0048] Figure 2 The image is taken by a drone before data processing for the multispectral tree species classification method based on separable void ASPP of this invention.

[0049] Figure 3 This is an image after data processing of the multispectral tree species classification method based on separable void ASPP of this invention;

[0050] Figure 4 The original label image is for the multispectral tree species classification method based on separable void ASPP of this invention;

[0051] Figure 5This is a graph showing the prediction results of the DeepLabV3+ model, which is an improved multispectral tree species classification method based on separable cavity ASPP according to the present invention.

[0052] Figure 6 This is a prediction result diagram of the DeepLabV3+ model based on the multispectral tree species classification method of separable void ASPP of this invention.

[0053] Figure 7 This is a parameter diagram for verifying the accuracy of the model after data augmentation of the multispectral tree species classification method based on separable void ASPP of the present invention;

[0054] Figure 8 This is a parameter diagram for verifying the model accuracy before data augmentation of the multispectral tree species classification method based on separable void ASPP of this invention.

[0055] Figure 1 Image represents the input image, SEBlock represents compression and excitation, SeparableASPP represents separable hollow pyramid pooling, FPNFusion concat represents feature pyramid network fusion and concatenation operation, Encoder represents the encoder, Decoder represents the decoder, 1×1Conv represents a 1×1 convolutional kernel, Upsample by4 represents upsampling by 4, 3×3Conv represents a 3×3 convolutional kernel, and Prediction represents the predicted image.

[0056] Figure 2-3 MPA represents average pixel precision, MioU represents average intersection-to-union ratio, TotalLoss represents training loss, and ValLoss represents validation loss. Detailed Implementation

[0057] Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] Example: Figures 1-8 As shown, this invention provides a technical solution: a multispectral tree species classification method based on separable void ASPP, the classification method comprising the following steps:

[0059] Step S1: Acquire tree species images using a drone; perform data preprocessing on the tree species images to obtain label images;

[0060] Step S1 includes:

[0061] Step S1-1: Collect tree species images using a drone, and stitch the tree species images collected by the drone together to obtain a complete tree species image in TIFF format;

[0062] Step S1-2: The complete tree species image is preprocessed using a combination of principal component analysis and adaptive noise reduction techniques;

[0063] Step S1-3: Crop the preprocessed complete tree species image according to the preset first size;

[0064] Steps S1-4: Label the cropped complete tree species image with data, and convert the labeled complete tree species image from TIFF format to JSON format to obtain the label image;

[0065] The preprocessing of the complete tree species image includes: first, standardizing the data of the complete tree species image; calculating the covariance between every two bands to obtain the covariance matrix; performing eigenvalue decomposition on the covariance matrix to obtain eigenvalues; and then selecting principal components with values ​​greater than the eigenvalues ​​for data reconstruction to remove noise and display the complete tree species image in near-infrared mode.

[0066] In this embodiment of the invention, broadleaf forests are selected as the research object. Multispectral image datasets are collected using drones. PCA is first performed on the collected multispectral image datasets. Then, based on the principal components of the data (i.e., the components retaining the largest variance), a dynamic denoising strategy is used to optimize the data for local features and noise distribution, achieving adaptive denoising in the drone multispectral data. The domain and weights are dynamically adjusted for different land cover types (such as trees, water bodies, buildings, etc.) and noise sources (sensors, shadows) to eliminate noise in the near-infrared band of the multispectral images, preserving as much near-infrared information as possible. Appropriate parameters are selected based on the characteristics of the drone data (such as the number of bands and resolution), and the effect is verified using ground truth data to ensure that the data after noise removal meets the requirements for subsequent classification of broadleaf forest tree species. Specific effects are as follows: Figure 2 , Figure 3 As shown;

[0067] Step S2: Perform data augmentation on the labeled images to obtain a sample dataset;

[0068] Step S2 includes:

[0069] Step S2-1: Using bilinear interpolation, the complete tree species image and label image cropped according to the first size are further cropped according to the preset second size. The complete tree species image and label image cropped according to the second size are simultaneously converted to JPG and PNG formats to obtain scaled image data and label data.

[0070] Step S2-2: Perform linear weighted synthesis on the image data and the label data using Mixup technology to obtain a synthesized image; perform geometric transformation on the synthesized image to obtain a sample dataset;

[0071] In an embodiment of the present invention, data augmentation operations are performed on the preprocessed tree species dataset. By scaling, horizontal flipping, rotating, and other transformations on the original images, the dataset, which originally contained only 150 images, is expanded to 4160 images.

[0072] Step S3: Introduce the EnhancedMobileNetV2 module, SEBlock module, SeparableASPP module, and FPNFusion feature fusion module into the basic DeepLabV3+ model to obtain the optimized DeepLabV3+ model; input the sample dataset into the optimized DeepLabV3+ model to obtain the semantic segmentation result;

[0073] Step S3 includes:

[0074] The sample dataset is processed by the EnhancedMobileNetV2 module to extract low-level and high-level features; the SEBlock module processes the extracted features by adjusting the feature responses of each channel; the features processed by the SEBlock module are passed to the SeparableASPP module, which uses a preset depthwise separable dilated convolution to extract spatial context information; the FPNFusion feature fusion module generates a feature map by upsampling and fusing shallow and deep features layer by layer; the feature map is then processed by the DeepLabV3+ decoder to obtain the semantic segmentation result;

[0075] In the embodiments of this invention, the improved DeepLabV3+ model enhances its feature extraction and fusion capabilities by introducing modules such as SEBlock, SeparableASPP, and FPNbusion under different computational loads, thereby improving both performance and segmentation accuracy. Specific segmentation results and accuracy verification parameters are as follows: Figure 4 , Figure 5 , Figure 6 , Figure 7 , Figure 8 As shown; among them, by Figure 7 and Figure 8It can be seen that the improved DeepLabV3+ model has a significant overall performance improvement on the data-augmented tree dataset. Specifically, the improved DeepLabV3+ model achieved an overall accuracy of 93.73% when trained on the data-augmented dataset, which is about 2.84% higher than the overall accuracy of the basic DeepLabV3+ model (DeepLabV3+(mobileV2)) trained on the data-unaugmented tree dataset, and about 0.51% higher than the overall accuracy of the basic DeepLabV3+ model trained on the data-augmented tree dataset. Figure 7 and Figure 8 DeepLabV3+ (mobileV2), DeepLabV3+ (Xception), Hrnet, and PSPnet are all existing models.

[0076] The EnhancedMobileNetV2 module includes:

[0077] Low-level and high-level features are extracted using an inverse residual structure. The inverse residual structure expands the input feature map using a 1×1 convolutional kernel through an expansion layer, then performs a depthwise separable convolution operation on the expanded feature map using a 3×3 convolutional kernel. A dimensionality reduction layer performs a linear projection operation on the feature map after the depthwise separable convolution operation using a 1×1 convolutional kernel. The 1×1 convolutional kernel of the dimensionality reduction layer does not contain an activation function. The depthwise separable convolutional layer performs independent convolution operations on each input channel using depthwise convolution, and maps the output of the depthwise convolution to a new channel space using a 1×1 pointwise convolution.

[0078] The SEBlock module is integrated into the low-level and high-level features of the EnhancedMobileNetV2 backbone network. The SEBlock module processes the extracted features using the following steps:

[0079] Step S7-1: Compress the input feature map in the spatial dimension by global average pooling to generate global statistics for each channel;

[0080] Step S7-2: Learn the relationship between channels through two convolutional layers; wherein, the first convolutional layer uses the reduction parameter to control the compression ratio to reduce the number of channels of the global statistics and uses the ReLU activation function to strengthen the channels; the second convolutional layer restores the number of channels of the global statistics to the initial size and uses the activation function to limit the channel weights to [0,1] to obtain the channel weights of each channel;

[0081] Step S7-3: Use the obtained channel weights to perform weighted processing on the input feature map;

[0082] The SeparableASPP module includes:

[0083] Step S8-1: Extract multi-scale features using depthwise separable convolution; wherein, depthwise separable convolution consists of depthwise convolution and pointwise convolution; wherein, depthwise convolution is performed on each input channel; pointwise convolution mixes information between channels using a 1×1 convolution kernel;

[0084] Step S8-2: Obtain global context information through global average pooling, and perform convolution and upsampling processing; wherein, by calculating the average value of all pixels in each channel, the spatial dimension information is compressed into a single value, thereby extracting global context information; after global average pooling, the input feature map retains only one average value for each output channel, and each average value represents the global information of the corresponding channel; after global average pooling, the features are processed by a 1×1 convolution kernel and upsampled to the initial size;

[0085] Step S8-3: Concatenate the multi-scale features extracted by depthwise separable convolution with the global context information obtained by global average pooling along the channel dimension to obtain the concatenated features; then fuse the concatenated features using a 1×1 convolution kernel.

[0086] The FPNFus ion feature fusion module is used to fuse the features output by the SeparableASPP module and the EnhancedMobileNetV2 module, and includes the following steps:

[0087] Step S9-1: Process the low-level features using a 1×1 convolution kernel to convert the number of channels of the low-level features from 24 to 48, so that the number of channels matches the number of channels of the high-level features;

[0088] Step S9-2: Upsample using bilinear interpolation to upsample high-level features to the same spatial resolution as low-level features; use a concatenation function to concatenate the high-level and low-level features to obtain the fused feature map;

[0089] Step S9-3: The fused feature maps are fused using a 3×3 convolution kernel to obtain feature maps with spatial details and semantic information.

[0090] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multispectral tree species classification method based on separable void ASPP, characterized in that, The classification method includes the following steps: Step S1: Obtain tree species images; perform data preprocessing on the tree species images to obtain label images; Step S2: Perform data augmentation on the labeled images to obtain a sample dataset; Step S3: Introduce the EnhancedMobileNetV2 module, SEBlock module, SeparableASPP module, and FPNFusion feature fusion module into the basic DeepLabV3+ model to obtain the optimized DeepLabV3+ model; input the sample dataset into the optimized DeepLabV3+ model to obtain the semantic segmentation result; Step S3 includes: The sample dataset is processed by the EnhancedMobileNetV2 module to extract low-level and high-level features; the SEBlock module processes the extracted features by adjusting the feature responses of each channel; the features processed by the SEBlock module are passed to the SeparableASPP module, which uses a preset depthwise separable dilated convolution to extract spatial context information; the FPNFusion feature fusion module generates a feature map by upsampling and fusing shallow and deep features layer by layer; the feature map is then processed by the DeepLabV3+ decoder to obtain the semantic segmentation result.

2. The multispectral tree species classification method based on separable void ASPP according to claim 1, characterized in that, Step S1 includes: Step S1-1: Collect tree species images using a drone, and stitch the tree species images collected by the drone together to obtain a complete tree species image in TIFF format; Step S1-2: The complete tree species image is preprocessed using a combination of principal component analysis and adaptive noise reduction techniques; Step S1-3: Crop the preprocessed complete tree species image according to the preset first size; Steps S1-4: Label the cropped complete tree species image with data, and convert the labeled complete tree species image from TIFF format to JSON format to obtain the label image.

3. The multispectral tree species classification method based on separable void ASPP according to claim 2, characterized in that, The preprocessing process for the complete tree species image includes: The data of the complete tree species image are standardized; the covariance between every two bands is calculated to obtain the covariance matrix, and the covariance matrix is ​​decomposed into eigenvalues ​​to obtain eigenvalues; Principal components with values ​​greater than the eigenvalues ​​are selected for data reconstruction to remove noise, so that the complete tree species image is displayed in near-infrared mode.

4. The multispectral tree species classification method based on separable void ASPP according to claim 1, characterized in that, Step S2 includes: Step S2-1: Using bilinear interpolation, the complete tree species image and label image cropped according to the first size are further cropped according to the preset second size. The complete tree species image and label image cropped according to the second size are simultaneously converted to JPG and PNG formats to obtain scaled image data and label data. Step S2-2: Perform linear weighted synthesis of the image data and the label data using Mixup technology to obtain a synthesized image; perform geometric transformation operation on the synthesized image to obtain a sample dataset.

5. The multispectral tree species classification method based on separable void ASPP according to claim 1, characterized in that, The EnhancedMobileNetV2 module includes: Low-level and high-level features are extracted using an inverse residual structure. The inverse residual structure expands the input feature map using a 1×1 convolutional kernel through an expansion layer, then performs a depthwise separable convolution operation on the expanded feature map using a 3×3 convolutional kernel. A dimensionality reduction layer performs a linear projection operation on the feature map after the depthwise separable convolution operation using a 1×1 convolutional kernel. The 1×1 convolutional kernel of the dimensionality reduction layer does not contain an activation function. The depthwise separable convolutional layer performs independent convolution operations on each input channel using depthwise convolution, and maps the output of the depthwise convolution to a new channel space using a 1×1 pointwise convolution.

6. The multispectral tree species classification method based on separable void ASPP according to claim 1, characterized in that, The SEBlock module is integrated into the low-level and high-level features of the EnhancedMobileNetV2 module backbone network; the SEBlock module processes the extracted features in the following steps: Step S7-1: Compress the input feature map in the spatial dimension by global average pooling to generate global statistics for each channel; Step S7-2: Learn the relationship between channels through two convolutional layers; wherein, the first convolutional layer uses the reduction parameter to control the compression ratio to reduce the number of channels of the global statistics and uses the ReLU activation function to strengthen the channels; the second convolutional layer restores the number of channels of the global statistics to the initial size and uses the activation function to limit the channel weights to [0,1] to obtain the channel weights of each channel; Step S7-3: Use the obtained channel weights to perform weighted processing on the input feature map.

7. The multispectral tree species classification method based on separable void ASPP according to claim 1, characterized in that, The SeparableASPP module includes: Step S8-1: Extract multi-scale features using depthwise separable convolution; wherein, depthwise separable convolution consists of depthwise convolution and pointwise convolution; wherein, depthwise convolution is performed on each input channel; pointwise convolution mixes information between channels using a 1×1 convolution kernel; Step S8-2: Obtain global context information through global average pooling, and perform convolution and upsampling processing; wherein, by calculating the average value of all pixels in each channel, the spatial dimension information is compressed into a single value, thereby extracting global context information; after global average pooling, the input feature map retains only one average value for each output channel, and each average value represents the global information of the corresponding channel; after global average pooling, the features are processed by a 1×1 convolution kernel and upsampled to the initial size; Step S8-3: Concatenate the multi-scale features extracted by depthwise separable convolution with the global context information obtained by global average pooling along the channel dimension to obtain the concatenated features; then fuse the concatenated features using a 1×1 convolution kernel.

8. The multispectral tree species classification method based on separable void ASPP according to claim 1, characterized in that, The FPNFusion feature fusion module is used to fuse the features output by the SeparableASPP module and the EnhancedMobileNetV2 module, including the following steps: Step S9-1: Process the low-level features using a 1×1 convolution kernel to convert the number of channels of the low-level features from 24 to 48, so that the number of channels matches the number of channels of the high-level features; Step S9-2: Upsample using bilinear interpolation to upsample high-level features to the same spatial resolution as low-level features; use a concatenation function to concatenate the high-level and low-level features to obtain the fused feature map; Step S9-3: The fused feature maps are fused using a 3×3 convolution kernel to obtain feature maps with spatial details and semantic information.

Images