A method for detecting diseased trees in remote sensing satellite images based on DeeplabV3 TC-SVM technology

Abstract

The present invention relates to a method for detecting diseased trees in remote sensing satellite images based on the DeeplabV3 TC-SVM technology, which solves the defect of difficult to improve the detection accuracy and effect of diseased trees compared with the prior art. The present invention includes the following steps: acquisition and preprocessing of remote sensing satellite images; construction of a diseased tree detection model for remote sensing satellite images; training of a diseased tree detection model for remote sensing satellite images; acquisition of remote sensing satellite image data; and obtaining of the diseased tree detection results. The present invention improves DeeplabV3 by an improved feature pyramid ASSP module and integrating a multi-head attention mechanism, strengthens the feature extraction ability of diseased trees in remote sensing satellite images, and finally improves the detection efficiency and accuracy of diseased trees through an SVM classifier, laying a foundation for improving the accuracy and efficiency of pest identification in the later stage, saving costs while improving the accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of satellite remote sensing image processing, and specifically to a method for detecting diseased trees in remote sensing satellite images based on DeeplabV3 TC-SVM technology. Background Art

[0002] Pine trees are one of the tree species with the largest area and the widest distribution in forest resources, and pine wood nematodes are their biggest natural enemies. Pine wood nematodes mainly rely on insects such as Monochamus alternatus for transmission, and have the characteristics of fast transmission speed, strong infectivity, high lethality, etc., posing a huge threat to the survival of pine trees. Therefore, the detection of pine wood nematode discolored wood has become an urgent problem to be solved at present.

[0003] However, for the detection of large areas of forest diseased trees, simply using drones combined with automatic AI detection has low recognition efficiency, often delaying the best treatment time for pests and diseases, resulting in immeasurable losses. However, the characteristics of satellite remote sensing data such as low acquisition cost, short revisit period, and large monitoring range make satellite remote sensing technology have natural advantages in the census and monitoring tasks of pine wood nematode disease. Therefore, the detection of diseased trees based on satellite remote sensing images has become a research hotspot at present.

[0004] Especially for the detection of pine wood nematode discolored wood, the target of diseased trees is small, the characteristic information extraction is insufficient, it is difficult to achieve real-time monitoring and single-tree precise monitoring; the detection area is large, and the model detection speed is slow.

[0005] Therefore, how to improve the ability to extract characteristic information of small targets and improve the detection efficiency and accuracy of pine wood nematode diseased trees has become an urgent technical problem to be solved. Summary of the Invention

[0006] The purpose of the present invention is to solve the defect that it is difficult to improve the detection accuracy and effect of diseased trees in the prior art, and provide a method for detecting diseased trees in remote sensing satellite images based on DeeplabV3 TC-SVM technology to solve the above problems.

[0007] To achieve the above purpose, the technical solution of the present invention is as follows:

[0008] A method for detecting diseased trees in remote sensing satellite images based on DeeplabV3 TC-SVM technology, comprising the following steps:

[0009] Collection and preprocessing of remote sensing satellite images: Obtain high-definition multi-spectral satellite remote sensing images, and perform image mosaicking, cutting, atmospheric correction, and orthorectification preprocessing on the data to form a data set, and perform diseased tree annotation on the preprocessed multi-spectral remote sensing satellite images to form a preprocessed multi-spectral remote sensing satellite image data set SRC;

[0010] Construction of a detection model for diseased trees in remote sensing satellite images: Based on DeeplabV3TC combined with SVM, a detection model for diseased trees in remote sensing satellite images is constructed;

[0011] Training of the detection model for diseased trees in remote sensing satellite images: Input the dataset labeled with diseased trees into the detection model for diseased trees in remote sensing satellite images for training;

[0012] Obtaining of remote sensing satellite image data: Obtain remote sensing satellite image data and perform preprocessing;

[0013] Obtaining of the detection results of diseased trees: Input the preprocessed remote sensing satellite image data into the trained detection model for diseased trees in remote sensing satellite images to obtain the detection results of diseased trees.

[0014] The acquisition and preprocessing of the remote sensing satellite images include the following steps:

[0015] Obtain high-definition multispectral remote sensing satellite images, and perform preprocessing on the data including image mosaicking, cutting, atmospheric correction, and orthorectification;

[0016] Perform coordinate registration on the multispectral remote sensing satellite images and ground UAV remote sensing images to unify the longitude and latitude coordinates of the diseased trees;

[0017] Perform shearing and flipping operations on the multispectral remote sensing satellite images to generate multiple multispectral remote sensing satellite images, and perform annotation on the diseased trees.

[0018] The construction of the detection model for diseased trees in remote sensing satellite images includes the following steps:

[0019] Set the first part of the detection model for diseased trees in remote sensing satellite images as the DeeplabV3TC semantic segmentation module, and the second part as the SVM diseased-tree classification and recognition module;

[0020] Set the DeeplabV3TC semantic segmentation module to be improved based on the DeeplabV3 network, which includes: an improved feature pyramid ASSP module and a multi-head attention W-MSA module;

[0021] Set the first layer of the DeeplabV3TC semantic segmentation module as the feature extraction network, and the feature extraction network as the resnet50 residual network, which includes: an input layer: receiving the input image; convolutional layers and pooling layers: performing feature extraction and downsampling on the input image through convolutional operations and pooling operations; residual blocks: composed of multiple residual modules, and each residual module has a skip connection; a global average pooling layer: performing average pooling on the height and width of the feature map to obtain a feature map with a fixed size;

[0022] The residual module A is set to contain two convolutional layers and a skip connection. The output of the convolutional layer passes through the ReLU activation function, is added to the skip connection, and then passes through another ReLU activation function. The residual module B is set to contain three convolutional layers and a skip connection. The residual module C is set to contain four convolutional layers and a skip connection. The residual module D is set to contain six convolutional layers and a skip connection. The residual module E contains three convolutional layers and a skip connection.

[0023] The second layer of the DeeplabV3TC semantic segmentation module is set as the feature fusion layer, which is an improved feature pyramid ASSP module;

[0024] The improved feature pyramid ASSP module is composed of three branches, where branch A and branch B are the output of the image through the feature extraction part and then enter the two swin transformer blocks for feature learning.

[0025] The input feature information of branches A and B is normalized by the normalization layer LN and then enters the feature learning module, followed by a residual operation. Then, it is normalized again by LN and then output by a multi-layer perceptron MLP and a residual. Among them, the feature learning module of branch A is the window multi-head attention mechanism W-MSA, and the feature learning module of branch B is the sliding window multi-head attention mechanism SW-MSA. Branch C is the image output by the feature extraction part and then undergoes a pooling operation. Finally, the feature information output by the three branches is fused to obtain a high-level feature map.

[0026] Set the third layer of the DeeplabV3TC semantic segmentation module as the decoding layer,

[0027] Set the first layer of the decoding layer to bilinear interpolation, the second layer to fusion layer, the third layer to multi-head attention W-MSA module, the fourth layer to 3×3 convolution layer, the fifth layer to multi-head attention W-MSA module, and the sixth layer to 3×3 convolution layer;

[0028] Set up the multi-head attention W-MSA module;

[0029] The multi-head attention W-MSA module is set to multiply multiple weight matrices at input to generate multiple Q, K, and V feature matrices for self-attention operations.

[0030] Where V represents the vector of input features, Q and K are the feature vectors for calculating attention weights; multiple outputs are combined through the Concat operation and then output through a fully connected layer;

[0031] The specific calculation formula of the multi-head attention mechanism is as follows:

[0032] MultiHead(Q, K, V) = Concat(head1,..., head n )W o ,

[0033]

[0034] where: n is the number of self-attention mechanisms, Q, K, and V are three feature vectors obtained by multiplying the input data with weight matrices W Q , W k , W v respectively, W O is the weight matrix, d x is the channel dimension, T represents the transpose operation, Concat represents the fusion operation, Attention represents the attention function, and SoftMax represents the normalization exponential function;

[0035] Set up the SVM diseased tree classification and recognition module;

[0036] After the image is subjected to feature extraction and feature learning by the DeeplabV3TC network, the SVM, as a classifier, is connected to the output of the DeeplabV3TC network through a SoftMax operation, and the diseased trees are recognized by classifying the feature information.

[0037] The training of the remote sensing satellite image diseased tree detection model includes the following steps:

[0038] Generate a corresponding dataset from the preprocessed image and input it into the remote sensing satellite image diseased tree detection model for training until the loss function converges. During training, use the Adam optimizer with parameters set as β1 = 0.9, β2 = 0.999, and the initial learning rate is 1.0×10 -4 , and the learning rate for the next training cycle becomes 0.98 times that of the previous one, and a total of 200 cycles are learned;

[0039] The input is the preprocessed multi-spectral remote sensing satellite image dataset SRC;

[0040] Input SRC into the DeeplabV3TC semantic segmentation module. After feature extraction in the first layer of the DeeplabV3TC semantic segmentation module and passing through the resnet50 residual network, low-level features L with a resolution of 1 / 4 and high-level features H with a resolution of 1 / 16 are obtained;

[0041] After feature fusion in the second layer of the DeeplabV3TC semantic segmentation module, the high-level feature map H is input into the improved pyramid feature extraction module ASPP to enable the model to capture multi-scale feature information. Three multi-scale features A, B, and C are obtained through the three branches of the improved pyramid feature extraction module ASPP. Finally, feature fusion is performed on A, B, and C to obtain feature F;

[0042] After decoding in the third layer of the DeeplabV3TC semantic segmentation module, in the decoding part, bilinear interpolation is performed on the fused feature F for 4 times upsampling to obtain feature F'. A feature fusion module is designed to fuse the low-level feature L with a resolution of 1 / 4 in the Backbone network and the high-level feature F' to obtain the fused feature R. Then, it passes through the multi-head attention W-MSA module, 3×3 convolutional layer, multi-head attention W-MSA module, 3×3 convolutional layer, and then bilinear interpolation is performed for 4 times upsampling to output the original image size Image feature tensor;

[0043] Classification and recognition are performed through the SVM diseased tree classification and recognition module: the Image feature tensor is sent into the SVM classifier for classification and recognition of the diseased tree area.

[0044] Beneficial effects

[0045] A remote sensing satellite image diseased tree detection method based on DeeplabV3TC-SVM technology in the present invention, compared with the prior art, improves DeeplabV3 by improving the feature pyramid ASPP module and integrating the multi-head attention mechanism, strengthens the feature extraction ability of diseased trees in remote sensing satellite images. Finally, the SVM classifier improves the efficiency and accuracy of diseased tree detection, lays a foundation for improving the accuracy and efficiency of later pest identification, and saves costs while improving accuracy. Brief description of the drawings

[0046] Figure 1 It is the sequence diagram of the method in the present invention;

[0047] Figure 2 It is the structural diagram of the remote sensing satellite image diseased tree detection model involved in the present invention;

[0048] Figure 3 It is the structural diagram of the improved feature pyramid ASPP module in the detection model involved in the present invention;

[0049] Figure 4 It is the structural diagram of the swin-transformer module in the improved feature pyramid ASPP module in the detection model involved in the present invention;

[0050] Figure 5 It is the structural diagram of the multi-head attention W-MSA module in the detection model involved in the present invention;

[0051] Figure 6 It is the result map of diseased trees detected by using the present invention; Specific embodiments

[0052] To have a further understanding and recognition of the structural features and achieved effects of the present invention, the following is a detailed description in conjunction with preferred embodiments and drawings:

[0053] Such as Figure 1 shown, a method for detecting diseased trees in remote sensing satellite images based on DeeplabV3TC-SVM technology according to the present invention includes the following steps:

[0054] The first step, collection and preprocessing of remote sensing satellite images: Obtain high-definition multi-spectral satellite remote sensing images, and perform image mosaicking, cutting, atmospheric correction, and orthorectification preprocessing on the data to form a data set, and perform diseased tree annotation on the preprocessed multi-spectral remote sensing satellite images to form a preprocessed multi-spectral remote sensing satellite image data set SRC.

[0055] (1) Obtain high-definition multi-spectral remote sensing satellite images, and perform image mosaicking, cutting, atmospheric correction, and orthorectification preprocessing on the data.

[0056] (2) Perform coordinate registration on the multi-spectral remote sensing satellite images and ground unmanned aerial vehicle remote sensing images to unify the longitude and latitude coordinates of diseased trees.

[0057] (3) Perform shearing and flipping operations on the multi-spectral remote sensing satellite images to generate multiple multi-spectral remote sensing satellite images, and perform diseased tree annotation.

[0058] The second step, construction of a diseased tree detection model for remote sensing satellite images: Such as Figure 2 shown, construct a diseased tree detection model for remote sensing satellite images based on DeeplabV3TC combined with SVM.

[0059] Set the diseased tree detection model for remote sensing satellite images as the DeeplabV3TC-SVM joint diseased tree detection network. The DeeplabV3TC-SVM network modifies the Resnet50 residual network in the encoder part to remove the last Stage in the original network as the Backbone of the improved network. Improve the spatial pyramid pooling module, and use the swintransformer network with window sizes of 2 and 7 to replace the original dilated convolutional part with dilation rates of 6, 12, and 18. Enable the model to obtain a larger field of view and more context information. Introduce a multi-head attention mechanism module in the decoding part to strengthen the network's attention to the targets in the image and improve the network's detection ability for small targets. Finally, the SVM classifier classifies and identifies diseased trees.

[0060] In view of the characteristics of forest diseased trees in satellite remote sensing images, the overall design of the network in this invention is to use the multi-head attention mechanism and Transformer to fuse and improve the traditional DeeplabV3 network, and propose a method for detecting diseased trees in remote sensing images based on the DeeplabV3TC-SVM network. Through ablation experiments, it is verified that both of the two improved methods have certain improvements in the detection accuracy and efficiency compared with the original network. Using the same dataset to conduct comparative experiments on the improved network and other classic networks, it is verified that the improved network has certain improvements in both the extraction of diseased tree feature information and the accuracy of extracting smaller diseased trees, reducing the probability of model misdetection and missed detection, and can obtain more accurate diseased tree detection results.

[0061] The construction of the diseased tree detection model for remote sensing satellite images includes the following steps:

[0062] (1) Set the first part of the diseased tree detection model for remote sensing satellite images as the DeeplabV3TC semantic segmentation module, and the second part as the SVM diseased tree classification and recognition module.

[0063] (2) Set the DeeplabV3TC semantic segmentation module to be improved based on the DeeplabV3 network, which includes: an improved feature pyramid ASSP module and a multi-head attention W-MSA module.

[0064] A1) Set the first layer of the DeeplabV3TC semantic segmentation module as the feature extraction network, and the feature extraction network is the resnet50 residual network, which includes: an input layer: receiving the input image; convolutional layers and pooling layers: extracting features and downsampling the input image through convolutional operations and pooling operations; residual blocks: composed of multiple residual modules, and each residual module has a skip connection; a global average pooling layer: performing average pooling on the height and width of the feature map to obtain a feature map with a fixed size;

[0065] Set the residual module A to contain two convolutional layers and one skip connection. The output of the convolutional layer is added to the skip connection after passing through the ReLU activation function, and then passes through another ReLU activation function; set the residual module B to contain three convolutional layers and one skip connection; set the residual module C to contain four convolutional layers and one skip connection; set the residual module D to contain six convolutional layers and one skip connection; the residual module E contains three convolutional layers and one skip connection;

[0066] A2) Set the second layer of the DeeplabV3TC semantic segmentation module as the feature fusion layer, and this layer is an improved feature pyramid ASSP module.

[0067] In the original pyramid feature extraction module (ASPP) of the DeeplabV3+ network, it is mainly composed of multiple dilated convolutions with different dilation rates. Using dilated convolution can expand the receptive field of the network for images, and using dilated convolutions with multiple dilation rates can obtain more multi-scale information in the image. Since dilated convolution can only obtain context information from some points around the pixel points and cannot obtain dense context information, the improved ASPP module introduces more context information by leveraging the characteristics of the transformer network to improve the model's detection ability for diseased trees. The main work is to replace the original dilated convolution layer with swin transformer blocks networks with different window sizes. Small windows are used to capture local information in the image, and larger windows are used to capture global information.

[0068] As Figure 3 shown, it is set that the improved feature pyramid ASPP module consists of three branches. Among them, branches A and B are that after the image is output by the feature extraction part, they enter 2 swin transformer block parts in sequence for feature learning. The swin transformer block part is as Figure 4 shown.

[0069] The feature information input to branches A and B enters the feature learning module after being normalized by the normalization layer LN, then undergoes a residual operation, and then after being normalized by LN again, it passes through a multi-layer perceptron MLP and a residual to obtain the output. Among them, the feature learning module of branch A is the window multi-head attention mechanism W-MSA, and the feature learning module of branch B is the sliding window multi-head attention mechanism SW-MSA. Branch C is that after the image is output by the feature extraction part, it undergoes a pooling operation, and finally the feature information output by the three branches is fused to obtain a high-level feature map.

[0070] A3) It is set that the third layer of the DeeplabV3TC semantic segmentation module is the decoding layer.

[0071] It is set that the first layer of the decoding layer is bilinear interpolation, the second layer is the fusion layer, the third layer is the multi-head attention W-MSA module, the fourth layer is a 3×3 convolution layer, the fifth layer is the multi-head attention W-MSA module, and the sixth layer is a 3×3 convolution layer.

[0072] (3) Set up the multi-head attention W-MSA module. Multi-head attention module: The core of the multi-head attention mechanism is the self-attention mechanism. Using the multi-head attention mechanism can enable the model to pay attention to more aspects of information. As the number of heads increases, the number of model parameters will also increase, making model training difficult. The input data is multiplied by the weight matrices Wq, Wk, and Wv respectively to obtain three feature vectors Q (query), K (key), and V (value), where V represents the vector of input features, and Q and K are the feature vectors for calculating the attention weights. The three feature vectors are linearly transformed by the Liner layer and used as the input of the self-attention mechanism. In the self-attention part, the dot product of the transpose of Q and K is first performed to calculate the attention score. Then it is divided by the scaling factor, processed using the softmax function, and multiplied by the tensor V to obtain the final output.

[0073] like Figure 5 As shown in Figure 2, the multi-head attention W-MSA module is set to multiply multiple weight matrices at input to generate multiple Q, K, and V feature matrices for self-attention operations.

[0074] Where V represents the vector of input features, Q and K are the feature vectors for calculating attention weights; multiple outputs are combined through the Concat operation and then output through a fully connected layer;

[0075] The specific calculation formula of the multi-head attention mechanism is as follows:

[0076] MultiHead(Q,K,V)=Concat(head1,...,head n )W o ,

[0077] head i =Attention(QW i Q ,KW i k ,VW i v ),

[0078]

[0079] Where n is the number of self-attention mechanisms, Q, K, and V are the input data and the weight matrix W respectively. Q 、W k 、W v The three eigenvectors obtained by multiplication, W O is the weight matrix, d xis the channel dimension, T represents the transpose operation, Concat represents the fusion operation, Attention represents the attention function, and SoftMax represents the normalization exponential function.

[0080] (4) Set up the SVM diseased tree classification and recognition module;

[0081] The SVM diseased tree classification and recognition module is that after the image undergoes feature extraction and feature learning through the DeeplabV3TC network, the SVM as a classifier is connected to the output of the DeeplabV3TC network through a SoftMax operation, and the recognition of diseased trees is achieved by classifying the feature information.

[0082] The third step, training of the remote sensing satellite image diseased tree detection model: Input the dataset labeled with diseased trees into the remote sensing satellite image diseased tree detection model for training.

[0083] (1) Use the preprocessed image to generate the corresponding dataset and input it into the remote sensing satellite image diseased tree detection model for training until the loss function converges. When training, use the Adam optimizer with parameter settings of β1 = 0.9, β2 = 0.999, and the initial learning rate is 1.0×10 -4 , and the learning rate for the next training cycle becomes 0.98 times that of the previous one, and a total of 200 cycles are learned.

[0084] (2) The input is the preprocessed multi-spectral remote sensing satellite image dataset SRC.

[0085] (3) Input SRC into the DeeplabV3TC semantic segmentation module. After feature extraction by the first layer of the DeeplabV3TC semantic segmentation module and passing through the resnet50 residual network, low-level features L with a resolution of 1 / 4 and high-level features H with a resolution of 1 / 16 are obtained;

[0086] After feature fusion by the second layer of the DeeplabV3TC semantic segmentation module, the high-level feature map H is input into the improved pyramid feature extraction module ASPP to enable the model to capture multi-scale feature information. Three multi-scale features A, B, and C are obtained through the three branches of the improved pyramid feature extraction module ASPP, and finally, feature fusion is performed on A, B, and C to obtain feature F;

[0087] After decoding by the third layer of the DeeplabV3+TC semantic segmentation module, in the decoding part, bilinear interpolation is performed on the fused feature F to upsample it by 4 times to obtain the feature F'. A feature fusion module is designed to fuse the low-level feature L with a resolution of 1 / 4 in the Backbone network and the high-level feature F' to obtain the fused feature R. Then, through the multi-head attention W-MSA module, 3×3 convolutional layer, multi-head attention W-MSA module, 3×3 convolutional layer, and then bilinear interpolation is performed to upsample by 4 times to output the Image feature tensor of the original image size.

[0088] (4) Classification and recognition by the SVM diseased tree classification and recognition module: The Image feature tensor is sent into the SVM classifier for classification and recognition of the diseased tree area.

[0089] Step 4, Acquisition of remote sensing satellite image data: Obtain remote sensing satellite image data and perform preprocessing.

[0090] Step 5, Obtaining the diseased tree detection result: Input the preprocessed remote sensing satellite image data into the trained remote sensing satellite image diseased tree detection model to obtain the diseased tree detection result.

[0091] As Figure 6 shown, it is the diseased tree result map detected by the model of the present invention. As can be seen from Figure 6 it, there are a large number of diseased trees in the figure, and the sizes of the diseased trees are also different. The smaller diseased trees are well detected and recognized, and the overall diseased tree detection effect is also good, further verifying the superiority of the present invention in the detection and recognition of diseased trees in remote sensing satellite images.

[0092] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification is only the principle of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection required by the present invention is defined by the appended claims and their equivalents.

Claims

1. A remote sensing satellite imagery diseased tree detection method based on DeeplabV3TC-SVM technology, characterized in that: The following steps are involved: 11) Acquisition and preprocessing of remote sensing satellite images: Acquire high-definition multispectral satellite remote sensing images and perform image stitching, cropping, atmospheric correction, and orthorectification preprocessing on the data to form a dataset. The preprocessed multispectral remote sensing satellite images are then annotated to form a preprocessed multispectral remote sensing satellite image dataset SRC. 12) Construction of a diseased tree detection model using remote sensing satellite images: A diseased tree detection model using remote sensing satellite images was constructed based on DeeplabV3TC combined with SVM. The construction of the remote sensing satellite image diseased tree detection model includes the following steps: 121) The first part of the remote sensing satellite imagery diseased tree detection model is set to be the DeeplabV3TC semantic segmentation module, and the second part is the SVM diseased tree classification and recognition module; 122) Set the DeeplabV3TC semantic segmentation module to be improved based on the DeeplabV3 network, which includes: improved feature pyramid ASSP module and multi-head attention W-MSA module; 1221) Set the first layer of the DeeplabV3TC semantic segmentation module as a feature extraction network, and the feature extraction network is a resnet50 residual network, which includes: an input layer: receiving an input image; a convolution layer and a pooling layer: performing feature extraction and downsampling on the input image through convolution and pooling operations; a residual block: consisting of multiple residual modules, each of which has a skip connection; a global average pooling layer: performing average pooling on the height and width of the feature map to obtain a fixed-size feature map; The residual module A is set to contain two convolutional layers and a skip connection. The output of the convolutional layer passes through the ReLU activation function, is added to the skip connection, and then passes through another ReLU activation function. The residual module B is set to contain three convolutional layers and a skip connection. The residual module C is set to contain four convolutional layers and a skip connection. The residual module D is set to contain six convolutional layers and a skip connection. The residual module E contains three convolutional layers and a skip connection. 1222) Set the second layer of the DeeplabV3TC semantic segmentation module as the feature fusion layer, which is an improved feature pyramid ASSP module; The improved feature pyramid ASSP module is composed of three branches, where branch A and branch B are the output of the image through the feature extraction part and then enter the two swin transformer blocks for feature learning. The input feature information of branches A and B is normalized by the normalization layer LN and then enters the feature learning module, followed by a residual operation. Then, it is normalized again by LN and then output by a multi-layer perceptron MLP and a residual. Among them, the feature learning module of branch A is the window multi-head attention mechanism W-MSA, and the feature learning module of branch B is the sliding window multi-head attention mechanism SW-MSA. Branch C is the image output by the feature extraction part and then undergoes a pooling operation. Finally, the feature information output by the three branches is fused to obtain a high-level feature map. 1223) Set the third layer of the DeeplabV3TC semantic segmentation module as the decoding layer, Set the first layer of the decoding layer to bilinear interpolation, the second layer to fusion layer, the third layer to multi-head attention W-MSA module, the fourth layer to 3×3 convolution layer, the fifth layer to multi-head attention W-MSA module, and the sixth layer to 3×3 convolution layer; 123) Set up the multi-head attention W-MSA module; The multi-head attention W-MSA module is set to multiply multiple weight matrices at input to generate multiple Q, K, and V feature matrices for self-attention operations. Where V represents the vector of input features, Q and K are the feature vectors for calculating attention weights; multiple outputs are combined through the Concat operation and then output through a fully connected layer; The specific calculation formula of the multi-head attention mechanism is as follows: MultiHead(Q,K,V)=Concat(head1,...,head n )W o , Where n is the number of self-attention mechanisms, Q, K, and V are the input data and the weight matrix W respectively. Q 、W k 、W v The three eigenvectors obtained by multiplication, W O is the weight matrix, d x is the channel dimension, T represents the transpose operation, Concat represents the fusion operation, Attention represents the attention function, and SoftMax represents the normalized exponential function; 124) Setting up SVM disease tree classification and recognition module; The SVM diseased tree classification and recognition module is a module that extracts and learns features from the image through the DeeplabV3TC network. The SVM is used as a classifier and is connected to the output of the DeeplabV3TC network through a SoftMax operation. The diseased tree is identified by classifying the feature information. 13) Training of the diseased tree detection model using remote sensing satellite images: Input the labeled diseased tree dataset into the diseased tree detection model using remote sensing satellite images for training; 14) Acquisition of remote sensing satellite image data: Acquisition of remote sensing satellite image data and preprocessing; 15) Obtaining diseased tree detection results: Input the pre-processed remote sensing satellite image data into the trained remote sensing satellite image diseased tree detection model to obtain diseased tree detection results.

2. A remote sensing satellite imagery diseased tree detection method based on DeeplabV3TC-SVM technology according to claim 1, characterized in that, The acquisition and preprocessing of the remote sensing satellite images includes the following steps: 21) Acquire high-definition multispectral remote sensing satellite images and perform image stitching, cutting, atmospheric correction, and orthorectification preprocessing on the data; 22) Coordinate registration of multispectral remote sensing satellite images and ground-based UAV remote sensing images to unify the latitude and longitude coordinates of the diseased trees; 23) Perform cropping and flipping operations on multispectral remote sensing satellite images to generate multiple multispectral remote sensing satellite images and perform tree labeling.

3. A remote sensing satellite imagery diseased tree detection method based on DeeplabV3TC-SVM technology according to claim 1, characterized in that, The training of the remote sensing satellite image diseased tree detection model includes the following steps: 31) The corresponding dataset generated by the preprocessed image was input into the remote sensing satellite image pest detection model for training until the loss function reached convergence. The Adam optimizer was used in training with parameters set to β1 = 0.9, β2 = 0.999, and the initial learning rate was 1.0 × 10 -4 , the learning rate of the next training cycle becomes 0.98 times that of the previous one, and a total of 200 cycles are learned; 32) The input is the preprocessed multispectral remote sensing satellite image dataset SRC; 33) Input SRC into the DeeplabV3TC semantic segmentation module, perform feature extraction through the first layer of the DeeplabV3TC semantic segmentation module, and obtain low-level features L with a resolution of 1 / 4 and high-level features H with a resolution of 1 / 16 after passing through the resnet50 residual network; After feature fusion is performed on the second layer of the DeeplabV3TC semantic segmentation module, the high-level feature map H is input into the improved pyramid feature extraction module ASPP to enable the model to capture multi-scale feature information. The three branches of the improved pyramid feature extraction module ASPP obtain three multi-scale features A, B, and C. Finally, the features of A, B, and C are fused to obtain feature F. After decoding by the third layer of the DeeplabV3TC semantic segmentation module, bilinear interpolation is performed on the fused feature F to perform 4x upsampling to obtain feature F'. A feature fusion module is designed to fuse the low-level feature L with a resolution of 1 / 4 in the Backbone network with the high-level feature F' to obtain the fused feature R. After that, it passes through the multi-head attention W-MSA module, 3×3 convolution layer, multi-head attention W-MSA module, 3×3 convolution layer, and then bilinear interpolation is performed to perform 4x upsampling to output the original image size Image feature tensor; 34) Classification and identification by SVM diseased tree classification and identification module: The image feature tensor is sent to the SVM classifier to classify and identify the diseased tree area.

Images