Intelligent ground cover identification method based on remote sensing image

Abstract

The invention discloses a remote sensing image-based ground cover intelligent identification method, which comprises the following steps of: obtaining a target ground cover remote sensing image set, and preprocessing the target ground cover remote sensing image set to obtain a preprocessed image set; constructing a segmentation network model, and segmenting the preprocessed image set to obtain a segmented image set; constructing an enhancement network model, and enhancing the sample size of the segmented image set to obtain an enhanced image set; and constructing a recognition network model, optimizing the recognition network model by using the training set in the enhanced image set to obtain an optimal recognition network model, inputting the test set in the enhanced image set into the optimal recognition network model to obtain a recognition result, and completing intelligent recognition of the ground cover. According to the method, the perception capability of the image is enhanced, the fine segmentation capability of the image is effectively improved, and the recognition accuracy is improved.

Description

Technical Field

[0001] The invention belongs to the technical field of remote sensing information and artificial intelligence, and in particular relates to an intelligent recognition method for ground cover based on remote sensing images. Background Art

[0002] With the increasing frequency of human activities and the continuous development of society, the information of surface cover such as cities, villages, mountains, rivers, and vegetation is becoming more and more abundant. Timely grasping and understanding the distribution of land feature information plays a vital role in the rational development and utilization of land resources. In addition, these land feature information are also widely used in transportation network planning, map drawing, geological survey, disaster prevention and mitigation, and precision agriculture, providing strong guidance for human production activities and having huge economic value. However, accurate classification and labeling of the above-mentioned surface cover information is a prerequisite for its value. At present, the classification of surface cover is mainly completed by professionals through reading and analyzing remote sensing images, which not only requires relevant personnel to have rich experience and superb professional qualities, but also requires a large number of professionals to spend a long time to complete. However, there is a shortage of talent resources in surveying and mapping in my country, and the distribution of high-quality resources is uneven. The traditional processing and interpretation method relying on professionals is time-consuming, labor-intensive, inefficient, and easily affected by subjective factors such as personnel level. Based on this, using deep learning to develop an intelligent recognition method and device for surface cover based on remote sensing images will help reduce the work pressure of surveying and mapping personnel, improve classification efficiency and reduce the misjudgment rate, and has a very broad prospect in application. As we all know, deep learning requires a large number of labeled samples. However, there are several problems with the current method of using remote sensing images to classify objects. First, the number of surface cover image sets that have been labeled with high quality is small and the labeling cost is high; second, due to the uneven distribution of natural objects, the sample size of different categories of objects obtained by remote sensing photography varies greatly, resulting in an imbalance between remote sensing image classes; third, the objects that need to be paid attention to in remote sensing images are smaller and more subtle. This brings great challenges to the use of deep learning technology to realize the intelligent classification of object information, resulting in high cost, low efficiency and low accuracy of the current object classification methods. Summary of the invention

[0003] In order to solve the above technical problems, the present invention proposes a method for intelligent recognition of surface cover based on remote sensing images, which enhances the perception ability of images, effectively improves the image fine segmentation capability and improves the recognition accuracy.

[0004] To achieve the above object, the present invention provides a method for intelligently identifying ground cover based on remote sensing images, comprising:

[0005] Acquire a target surface cover remote sensing image set, preprocess the target surface cover remote sensing image set, and obtain a preprocessed image set;

[0006] Constructing a segmentation network model, segmenting the preprocessed image set, and obtaining a segmented image set;

[0007] Constructing an enhanced network model, enhancing the sample size of the segmented image set to obtain an enhanced image set, and dividing the enhanced image set into a training set and a test set;

[0008] Construct a recognition network model, use the training set in the enhanced image set to optimize the recognition network model to obtain the optimal recognition network model, input the test set in the enhanced image set into the optimal recognition network model, obtain the recognition result, and complete the intelligent recognition of surface cover.

[0009] Optionally, the target surface cover remote sensing image set is preprocessed to obtain the preprocessed image set, including: cropping, denoising, grayscale, histogram equalization, gamma correction, normalization and size unification of the target surface cover remote sensing image set to obtain the preprocessed image set.

[0010] Optionally, building a segmentation network model includes:

[0011] Based on the supervised segmentation network Unet, a linear deformable convolution module and a multi-head attention refinement module are added to complete the construction of the segmentation network model.

[0012] Optionally, segmenting the preprocessed image set to obtain the segmented image set includes:

[0013] Inputting the preprocessed image set into the segmentation network model, performing downsampling to obtain a feature map, performing size enlargement processing on the feature map by upsampling, and classifying the size enlarged feature map to obtain a binary classification image of characteristic characteristics;

[0014] Using the linear deformable convolution module to convolve the binary image of the characteristic characteristics to obtain a convolved image;

[0015] The multi-head attention refinement module is used to weight encode the convolved image to obtain a segmented image set.

[0016] Optionally, when optimizing the recognition network model, a loss function is used for convergence, and the loss function is constructed by integrating a cross entropy loss function and a Tversky loss function:

[0017] Loss = λLoss Seg +(1-λ)Loss Tversky

[0018] LossSeg =y ture log(y predict )+(1-y ture )log(1-y ture ),

[0019]

[0020] Among them, λ is the weight coefficient, α and β are the coefficients for controlling false negatives and false positives respectively, and Loss seg is the cross entropy loss function, Loss Tversky is the Tversky loss function, y ture represents the true label, y predict Represents the predicted label, A represents the Ground truth segmentation image, and B represents the predicted segmentation image.

[0021] Optionally, building an enhanced network model includes:

[0022] The enhanced network model is CycleGAN composed of two mirrored generative adversarial networks, including two generators and two discriminators, and the loss function used includes generator G and discriminator D Y Loss between generator F and discriminator D X loss between them and cycle consistency loss.

[0023] Optionally, the learning-perceived image similarity block is incorporated into the CycleGAN loss function to obtain a composite loss function expressed as:

[0024] L(G,F,D x ,D y )=L GAN (G,D Y ,X,Y)+L GAN (F,D X ,Y,X)+λL cyc (G,F)+βL LPIPS (X,F(G(X)),

[0025] Among them, L GAN (G,D Y ,X,Y) are the generator G and the discriminator D Y Loss between, L GAN (F,D X ,Y,X) generator F and discriminator D X Loss between, L cyc (G, F) is the cycle consistency loss, L LPIPS (X,F(G(X)) is the perceptual loss, and λ and β are weight coefficients.

[0026] Optionally, building a recognition network model includes:

[0027] Build a CNN network model including convolutional layers, pooling layers and fully connected layers;

[0028] Introducing a multi-head attention mechanism into the CNN network model;

[0029] The hyperparameters of the CNN network model that introduces the multi-head attention mechanism are optimized according to the adaptive whale optimization algorithm to complete the construction of the recognition network model.

[0030] Optionally, optimizing the hyperparameters of the CNN network model that introduces the multi-head attention mechanism according to the adaptive whale optimization algorithm includes:

[0031] In the stage of surrounding prey, the whale gradually surrounds the prey by identifying the location of the prey and constantly adjusting its own position;

[0032] During the bubble attack phase, the whale's position is updated in a spiral manner;

[0033] During the prey search phase, whales move based on the locations of other whales and search for prey.

[0034] Optionally, the method further includes visually displaying the recognition result.

[0035] Technical effect of the present invention: The present invention discloses an intelligent recognition method for surface cover based on remote sensing images. The designed LDConvMHAR-Unet network is a supervised remote sensing image segmentation network. Compared with the unsupervised remote sensing image segmentation network, the segmentation accuracy and generalization ability of the present invention are superior. Compared with other supervised image segmentation technologies, the present invention uses LDConv to dynamically capture the global features of remote sensing images and enhance the perception ability of images on the one hand; on the other hand, MHAR is used to focus on the features of tiny objects in the image, effectively improving the image's refined segmentation ability. The LPIPS-CycleGAN sample enhancement network designed by the present invention incorporates learning-perceived image similarity blocks that are sensitive to tiny features. Compared with the traditional remote sensing image enhancement method of CycleGAN, the present invention can generate a large number of high-resolution image sets with obvious tiny object features, which can meet the requirements in both quantity and quality. The deep learning network designed by the present invention can not only mine the spatial distribution characteristics of objects, but also extract the key features between different object categories, which is helpful to improve the recognition accuracy; at the same time, the network introduces a multi-head attention mechanism to improve the training speed of the network model. The present invention uses an adaptive whale optimization algorithm to optimize the network structure, hyperparameters, attention weights, etc., and can obtain a network model with the best classification performance in a relatively short time, thereby improving the training speed and efficiency of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0037] Figure 1 A schematic diagram of an implementation process of a method for intelligently identifying ground cover based on remote sensing images according to an embodiment of the present invention;

[0038] Figure 2 A schematic diagram of a flow chart of a method for intelligently identifying ground cover based on remote sensing images according to an embodiment of the present invention;

[0039] Figure 3 A structural diagram of the supervised LDConvMHAR-Unet segmentation network model designed for an embodiment of the present invention;

[0040] Figure 4 A schematic diagram of the model training and testing process provided by an embodiment of the present invention;

[0041] Figure 5 A schematic diagram of a confusion matrix showing the comparison results of the embodiments of the present invention with other intelligent swarm optimization algorithms;

[0042] Figure 6 A schematic diagram of a recognition result visualization interface developed based on Python provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0043] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0044] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0045] like Figure 1-Figure 2 As shown, this embodiment provides a method for intelligently identifying ground cover based on remote sensing images, including:

[0046] Preprocess the existing labeled land cover remote sensing image set;

[0047] Construct the LDConvMHAR-Unet segmentation network to perform segmentation on the preprocessed image set;

[0048] Construct the LPIPS-CycleGAN enhancement network to increase the sample size of the image set;

[0049] The recognition network model of AWOA-CNN-MAHM was built, and the adaptive whale algorithm was used to optimize the parameters in the CNN network to obtain the best network model. The test set was sent to the best network model for intelligent recognition of surface cover.

[0050] Furthermore, the recognition results are visualized on the terminal platform.

[0051] In this embodiment, first, the existing labeled surface cover remote sensing image set is preprocessed, mainly by cropping, denoising, graying, histogram equalization, gamma correction, normalization and resizing, and finally a 512×512 grayscale image is obtained.

[0052] Furthermore, the preprocessed image set is segmented, that is, the Linear Deformable Convolution (LDConv) module and the Multi-head Attention Refinement (MHAR) module are introduced on the basis of the supervised segmentation network Unet, and the LDConvMHAR-Unet segmentation network is proposed. The network first uses LDConv to dynamically capture the global features of the remote sensing image of the ground object, enhances the global perception ability of the image, and then uses MHAR to improve the ability to segment subtle ground object features. Specifically:

[0053] First, a Unet network consisting of an encoder-decoder structure and a jump connection is built. The encoder structure is located in the left half and consists of a convolution layer, a pooling layer, and an activation function to implement downsampling operations; the decoder structure is located in the right half, and the feature image is enlarged through a series of upsampling operations. Finally, Softmax is used for classification operations to output the binary classification of feature characteristics. Then, LDConv is used to replace the standard convolution operation in the left half of Unet, and convolution kernels of different numbers and sizes are used to perform convolution operations on the remote sensing image of the surface cover to enhance the network's perception of the image. Finally, the MHAR module is added between the encoder and the decoder to encode the weight of the entire image in the horizontal and vertical directions, enhance the weight of the ground object image, and suppress the weight of the background pixels, so that the final feature map is obtained:

[0054]

[0055] In the formula, F is the input feature, P h is the horizontal pooling operation, P w is the vertical pooling operation, P c It is the channel average pooling operation.

[0056] Since the shapes and sizes of ground covers are different and randomly distributed, the present invention proposes a new loss function that integrates the cross entropy loss function and the Tversky loss function. The specific formula is shown below.

[0057] Loss = λLoss Seg +(1-λ)Loss Tversky

[0058] Loss Seg =y ture log(y predict )+(1-y ture )log(1-y ture )

[0059]

[0060] In the formula, λ is the weight coefficient, α and β are the coefficients for controlling false negatives and false positives respectively, and Loss seg is the cross entropy loss function, Loss Tversky is the Tversky loss function, y ture represents the true label, y predict Represents the predicted label, A represents the Ground truth segmentation image, and B represents the predicted segmentation image.

[0061] Furthermore, a sample enhancement network based on LPIPS-CycleGAN is constructed to increase the number of image sets. Specifically:

[0062] Build a CycleGAN consisting of two mirrored generative adversarial networks, which includes two generators and two discriminators. The loss function used includes the generator G and the discriminator D Y Loss between generator F and discriminator D X However, due to the presence of a large number of tiny objects in remote sensing images, the resolution of images generated by traditional CycleGAN cannot meet the analysis requirements. To address this, the present invention incorporates the learned perceptual image patch similarity (LPIPS) into the CycleGAN loss function and proposes a new high-resolution composite loss function. The specific formula is shown below.

[0063] L(G,F,D x ,D y )=L GAN (G,D Y ,X,Y)+L GAN (F,D X ,Y,X)+λL cyc(G,F)+βL LPIPS (X,F(G(X)),

[0064] Where, L GAN (G,D Y ,X,Y) are the generator G and the discriminator D Y Loss between, L GAN (F,D X ,Y,X) generator F and discriminator D X Loss between, L cyc (G, F) is the cycle consistency loss, L LPIPS (X,F(G(X)) is the perceptual loss, and λ and β are weight coefficients.

[0065] Furthermore, the AWOA-CNN-MHAM network model is constructed, specifically:

[0066] First, a CNN network consisting of convolutional layers, pooling layers, and fully connected layers is built to extract the features of segmented remote sensing images.

[0067] Then, a multi-head attention mechanism is introduced to extract key core features from the extracted features and remove redundant information. The whole process is:

[0068] Q=XW Q ,

[0069] K=XW K ,

[0070] V=XW V ,

[0071]

[0072] Multihead(Q,K,V)=concat(head1,L,head h )W o ,

[0073] Where X is the output of BiMGRU; W Q , W K , W V They are three trainable parameter matrices respectively; d k is the dimension; W o is the output weight matrix; Softmax() means normalizing the scores.

[0074] Finally, a new adaptive whale optimization algorithm is proposed to optimize the hyperparameters of the CNN-MHAM network model. The optimization algorithm process includes:

[0075] In the stage of surrounding prey, the whale identifies the position of the prey and continuously updates its own position to gradually surround the prey. In this stage, the whale's position is adaptively updated using the following formula:

[0076]

[0077] In the formula, t represents the number of iterations, X * (t) represents the current best whale position, X(t) represents the current whale position, D represents the distance between the current whale position and the best whale position, X(t+1) represents the t+1 generation whale position, ω is an adaptive weight coefficient designed by the present invention to improve the global and local search ability and convergence ability of the algorithm, ω st represents the initial weight at the beginning of the iteration, ω end Represents the weight at the end of the iteration, M iter represents the maximum number of iterations, k represents the nonlinear adjustment coefficient, and r1,r2∈[0,1] are random numbers.

[0078] During the bubble attack phase, the whale's position is updated in a spiral manner as shown below:

[0079]

[0080] In the formula, b is a constant, l is a random number between [-1, 1], and ω is an adaptive weight coefficient designed by the present invention.

[0081] In the prey search phase, the whale moves according to the positions of other whales and searches for prey. The position update in this phase uses the following formula:

[0082]

[0083] Where, X rand (t) is the position of the randomly selected individual, D″ is the distance between the randomly selected whale individual and the whale individual, and ω is the adaptive weight coefficient designed by the present invention.

[0084] The proposed AWOA algorithm is used to optimize the CNN network structure, hyperparameters, and attention weights to obtain the optimal network structure; then the test set is sent to the trained network model for testing.

[0085] Furthermore, a terminal user operation interface is developed to visualize the analysis results on the terminal platform: specifically, first, a terminal user platform embedded with a trained deep learning network model is developed based on Python, and the model is saved in a storage unit with computing functions; then, the computing storage unit is connected to the processor; finally, the user can use the operation interface to visualize the surface cover identification results in real time on the terminal.

[0086] Figure 3 It is a new type of supervised LDConvMHAR-Unet segmentation network provided by an embodiment of the present invention. The network adds LDConv module and MHAR module on the basis of traditional Unet, and the segmented image can refine the local features of tiny objects. The loss function adopts a new composite function that integrates the cross entropy loss function and the Tversky loss function. The Epochs of the training process is set to 180, the Batch_size is set to 6, and the initial learning rate is set to 0.0002. Table 1 shows the quantitative comparison results of the accuracy (Acc), specificity (Spe), sensitivity (Sen), precision (Pre) and intersection over union (IoU) of the tiny object of grassland segmented by the segmentation algorithm proposed by the present invention and other common segmentation algorithms.

[0087] Table 1

[0088] method Acc Spe Sen Pre IoU LDConvMHAR-Unet 98.5％ 98.0％ 99.0％ 98.0％ 97.0％ Unet 81.5％ 78.0％ 85.0％ 79.4％ 69.7％ R2-Unet 89.5％ 89.0％ 90.0％ 89.1％ 81.1％ ScSE-Unet 92.0％ 90.0％ 94.0％ 90.4％ 85.5％ AMICNet 85.0％ 86.0％ 88.0％ 86.3％ 77.2％

[0089] As can be seen from Table 1, compared with Unet, R2-Unet, ScSE-Unet and AMICNet, the segmentation method proposed in the present invention has improved accuracy by 17.0%, 9.0%, 6.5% and 13.5% respectively; specificity by 20.0%, 9.0%, 8.0% and 12.0% respectively; sensitivity by 14.0%, 9.0%, 5.0% and 11.0% respectively; precision by 18.6%, 8.9%, 7.6% and 11.7% respectively; intersection over union by 27.3%, 15.9%, 11.5% and 19.8% respectively. Therefore, the segmentation algorithm proposed in the present invention has the best effect on remote sensing ground object image segmentation.

[0090] Figure 4 The figure is a flow chart of the model training and testing provided by the embodiment of the present invention. The network consists of a CNN network and an MHAM model. The initial structure of the CNN includes a layer consisting of 3 convolutional layers, 3 pooling layers and 1 fully connected layer. The convolutional layer search range is 2 to 8, and the learning rate search range is set to 10. -5 ~10 -2 , Epoach = 300, the batch size search range is 5 to 20, the MGRU neuron number search range is set to less than 256, the MHAM model head number search range is 4 to 20, and each head dimension is 32. The AWOA parameters are set as follows: population size is 200, maximum number of iterations M iter is 500, nonlinear iteration coefficient k=0.9, r1=r2=0.5, b=0.5, l=0.4.

[0091] The AWOA algorithm is used to optimize the relevant parameters in the CNN-MHAM network training process. After obtaining the optimal network structure, the test set is used for testing and verification. The hyperparameters of the learning network after optimization are shown in Table 2.

[0092] Table 2

[0093] parameter Number of CNN convolutional layers Learning Rate Batch size MGRU neurons MHAM Head Count Value 6 0.00738 20 104 18

[0094] Figure 5 This is the comparison test result between the AWOA intelligent optimization algorithm adopted by the present invention and other intelligent optimization algorithms. Figure 5 Part (a) is a schematic diagram of the confusion matrix of the detection results after the network is optimized using the immune optimization algorithm (IA). Figure 5 Part (b) is a schematic diagram of the confusion matrix of the detection results after the network is optimized using the particle swarm algorithm (PSO). Figure 5 Part (c) is a schematic diagram of the confusion matrix of the detection results after the network is optimized using the ant colony algorithm (ACO). Figure 5 Part (d) is a schematic diagram of the confusion matrix of the detection results after the network is optimized using the Grey Wolf Algorithm (GWO). Figure 5 Part (e) is a diagram of the confusion matrix of the detection results after the standard whale (WOA) optimization network is used. Figure 5 Part (f) is a schematic diagram of the confusion matrix of the detection results after the AWOA optimization network of the present invention is used. It can be seen that:

[0095] Compared with the traditional WOA optimization algorithm, the final accuracy and F1-score of the AWOA optimization algorithm adopted by the present invention are 98.33% and 98.33%, respectively, which are increased by 4.83% and 4.81%, respectively, proving that the optimization performance of WOA is improved after the addition of the adaptive mechanism; compared with the commonly used intelligent optimization algorithms such as IA, PSO, ACO and GWO, the accuracy is increased by 12.66%, 10.00%, 11.66% and 7.33%, respectively, and the F1-score is increased by 12.64%, 10.01%, 11.64% and 7.27%, respectively. It can be seen that the performance of optimizing the classification network using the AWOA proposed by the present invention is the best.

[0096] Figure 6 It is a visualization interface of recognition results developed based on Python. The user interface platform here is developed based on Python language. The platform has functions such as result analysis and visualization display, data storage, etc. Through the development of this user interface platform, the surface cover recognition results are presented in an intuitive form.

[0097] From the description of the above embodiments, those skilled in the art can know that: the present invention provides a method for intelligently identifying surface cover. The present invention combines remote sensing technology, data enhancement and segmentation technology, and artificial intelligence technology to achieve accurate identification of surface cover, and provides valuable reference and basis for surveying and mapping personnel to mark objects. It can not only help relevant personnel improve efficiency, but also reduce the misjudgment rate of small object information and improve marking accuracy, and has a very broad application prospect.

[0098] The present invention also provides a ground cover intelligent identification device, comprising a memory and a processor, wherein: the memory stores computer instructions;

[0099] The processor is data-connected to the memory, and the processor executes the above-mentioned intelligent identification method of the ground cover category by executing the computer instructions.

[0100] An embodiment of the present invention also provides a storage medium on which one or more programs readable by a computing device are stored. The one or more programs include instructions, which, when executed by the computing device, enable the computing device to perform intelligent identification of a surface cover category in Example 1.

[0101] Examples of computer-readable storage media include: read-only memory (ROM), random-access programmable read-only memory (PROM), electrically erasable programmable read-only memory (EEPROM), random-access memory (RAM), dynamic random-access memory (DRAM), static random-access memory (SRAM), flash memory, non-volatile memory, CD-ROM, DVD-ROM, Blu-ray or optical disk storage, hard disk drive (HDD), solid-state drive (SSD), card-type memory (such as a multimedia card, a secure digital (SD) card, or an extreme digital (XD) card), magnetic tape, floppy disk, magneto-optical data storage device, optical data storage device, hard disk, solid-state disk, and any other device configured to store a computer program and any associated data, data files, and data structures in a non-transitory manner and provide the computer program and any associated data, data files, and data structures to a processor or computer so that the processor or computer can execute the computer program. The computer program in the above-mentioned computer-readable storage medium can be executed in an environment deployed in a computer device such as a client, a host, an agent device, a server, etc. In addition, in one example, the computer program and any associated data, data files, and data structures are distributed on networked computer devices, so that the computer program and any associated data, data files, and data structures are stored, accessed, and executed in a distributed manner by one or more processors or computers.

[0102] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, devices or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0103] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for intelligently identifying ground cover based on remote sensing images, characterized in that: include: Acquire a target surface cover remote sensing image set, preprocess the target surface cover remote sensing image set, and obtain a preprocessed image set; Constructing a segmentation network model, segmenting the preprocessed image set, and obtaining a segmented image set; Constructing an enhanced network model, enhancing the sample size of the segmented image set to obtain an enhanced image set, and dividing the enhanced image set into a training set and a test set; Construct a recognition network model, use the training set in the enhanced image set to optimize the recognition network model to obtain the optimal recognition network model, input the test set in the enhanced image set into the optimal recognition network model, obtain the recognition result, and complete the intelligent recognition of surface cover.

2. The method for intelligently identifying ground cover based on remote sensing images according to claim 1, characterized in that: Preprocessing the target surface cover remote sensing image set to obtain the preprocessed image set includes: cropping, denoising, graying, histogram equalization, gamma correction, normalization and size unification of the target surface cover remote sensing image set to obtain the preprocessed image set.

3. The method for intelligently identifying ground cover based on remote sensing images according to claim 1, characterized in that: Building a segmentation network model includes: Based on the supervised segmentation network Unet, a linear deformable convolution module and a multi-head attention refinement module are added to complete the construction of the segmentation network model.

4. The method for intelligently identifying ground cover based on remote sensing images according to claim 3, characterized in that: Segmenting the preprocessed image set to obtain a segmented image set includes: Inputting the preprocessed image set into the segmentation network model, performing downsampling to obtain a feature map, performing size enlargement processing on the feature map by upsampling, and classifying the size enlarged feature map to obtain a binary classification image of characteristic characteristics; Using the linear deformable convolution module to convolve the binary image of the characteristic characteristics to obtain a convolved image; The multi-head attention refinement module is used to weight encode the convolved image to obtain a segmented image set.

5. The method for intelligently identifying ground cover based on remote sensing images according to claim 1, characterized in that: When optimizing the recognition network model, a loss function is used for convergence, and the loss function is constructed by integrating the cross entropy loss function and the Tversky loss function: Loss=λLoss Seg +(1-λ)Loss Tversky Loss Seg =and ture log(y predict )+(1-and ture )log(1-y ture ), Among them, λ is the weight coefficient, α and β are the coefficients for controlling false negatives and false positives respectively, and Loss seg is the cross entropy loss function, Loss Tversky is the Tversky loss function, y ture represents the true label, y predict Represents the predicted label, A represents the Groundtruth segmentation image, and B represents the predicted segmentation image.

6. The method for intelligently identifying ground cover based on remote sensing images according to claim 1, characterized in that: Building an enhanced network model includes: The enhanced network model is CycleGAN composed of two mirrored generative adversarial networks, including two generators and two discriminators, and the loss function used includes generator G and discriminator D Y Loss between generator F and discriminator D X loss between them and cycle consistency loss.

7. The method for intelligently identifying ground cover based on remote sensing images according to claim 6, characterized in that: By integrating the learning-perceived image similarity block into the CycleGAN loss function, we obtain a composite loss function, which is expressed as: L(G,F,D x ,D y )=L GAN (G,D Y ,X,Y)+L GAN (F,D X ,Y,X)+λL cyc (G,F)+βL LPIPS (X,F(G(X)), Among them, L GAN (G,D Y ,X,Y) are the generator G and the discriminator D Y Loss between, L GAN (F,D X ,Y,X) generator F and discriminator D X Loss between, L cyc (G, F) is the cycle consistency loss, L LPIPS (X,F(G(X)) is the perceptual loss, and λ and β are weight coefficients.

8. The method for intelligently identifying ground cover based on remote sensing images according to claim 1, characterized in that: Building a recognition network model includes: Build a CNN network model including convolutional layers, pooling layers and fully connected layers; Introducing a multi-head attention mechanism into the CNN network model; The hyperparameters of the CNN network model that introduces the multi-head attention mechanism are optimized according to the adaptive whale optimization algorithm to complete the construction of the recognition network model.

9. The method for intelligently identifying ground cover based on remote sensing images according to claim 1, characterized in that: The optimization of the hyperparameters of the CNN network model with multi-head attention mechanism based on the adaptive whale optimization algorithm includes: In the stage of surrounding prey, the whale gradually surrounds the prey by identifying the location of the prey and constantly adjusting its own position; During the bubble attack phase, the whale's position is updated in a spiral manner; During the prey search phase, whales move based on the locations of other whales and search for prey.

10. The method for intelligently identifying ground cover based on remote sensing images according to claim 1, characterized in that: The method also includes visually displaying the recognition result.

Images