A remote sensing image building extraction method fusing convolutional neural network and transformer

Abstract

This invention provides a method for building extraction from remote sensing images that integrates convolutional neural networks (CNNs) and Transformers, relating to the intersection of remote sensing image segmentation and computer vision technologies. This method involves creating a remote sensing image dataset from acquired remote sensing images, labeling each image individually, and dividing the dataset into training, validation, and test sets. The preprocessed remote sensing images are then preprocessed to increase data diversity. Feature extraction is performed on the preprocessed images to collect feature maps containing building information and obtain global features of the image. This method can significantly reduce false positives and false negatives for small target buildings, improve the completeness of segmentation for large target buildings, and mitigate boundary blurring caused by insufficient extraction of target building edge information.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of remote sensing image segmentation and computer vision, specifically to a method for extracting buildings from remote sensing images by integrating convolutional neural networks and Transformers. Background Technology

[0002] The extraction of buildings from remotely sensed images can essentially be viewed as an image segmentation process. Currently, most deep learning-based image segmentation models utilize Convolutional Neural Networks (CNNs). In 2015, Jonathan Long et al. proposed FCN, implementing the first complete end-to-end semantic segmentation framework. It learns image features layer by layer through convolution and pooling operations, replacing fully connected layers with fully convolutional layers, and generating segmentation maps of the original image pixel by pixel through skip connections and upsampling. However, convolution operations, based on the size of their kernels, only perform local computations and cannot capture long-range dependencies. Furthermore, multiple downsampling operations after pooling lead to the loss of detailed image information. These factors contribute to the insufficient feature extraction capabilities of existing deep learning-based methods for buildings, resulting in low segmentation accuracy. Summary of the Invention

[0003] (a) Technical problems to be solved

[0004] To address the shortcomings of existing technologies, this invention provides a remote sensing image building extraction method that integrates convolutional neural networks and Transformers. This method can significantly reduce the problems of false detection and missed detection of small target buildings, improve the integrity of segmentation of large target buildings, and improve the boundary blurring problem caused by insufficient extraction of target building edge information.

[0005] (II) Technical Solution

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for extracting buildings from remote sensing images by integrating convolutional neural networks and Transformers, comprising:

[0007] The acquired remote sensing images are made into a remote sensing image dataset and labeled one by one. The remote sensing image dataset is divided into training set, validation set and test set.

[0008] Preprocessing remote sensing images obtained after processing remote sensing images increases the diversity of data.

[0009] Feature extraction is performed on the preprocessed remote sensing images to collect feature maps containing building information; and global features of the images are obtained.

[0010] The feature maps containing building information collected and the global features of the obtained images are fused to obtain semantic features containing both local and global information, and feature maps containing rich contextual information are obtained.

[0011] After preprocessing the divided training and validation sets, the model is trained and iterated for several rounds until it converges. The trained model weights are then saved.

[0012] The predefined test set is segmented and classified to obtain the model's output visualization results for evaluation.

[0013] Preferably, the feature map containing building information is collected using ResNet Block, and the specific calculation steps of ResNet Block are as follows:

[0014] Given a feature map A, it first passes through a 1×1 convolution to exchange channel information, then through a 3×3 convolution for feature extraction, and finally through a 1×1 convolution to adjust the number of channels. The extracted feature C is then added to the input feature map A via a residual connection. A 1×1 convolution is added to the residual connection to downsample the input feature map A. The mathematical expression is as follows:

[0015] C = f 1×1 (f 3×3 (f 1×1 (A)))+f 1×1 (A)

[0016] Here, f1×1 and f3×3 represent convolutions with kernels of 1×1 and 3×3, respectively.

[0017] Preferably, the acquisition of global features of the image is achieved through Swin Transformer Block, and the specific steps of Swin Transformer Block design are as follows:

[0018] The Swin Transformer Block comprises alternating window multi-head self-attention layers, offset window multi-head self-attention layers, and multilayer perceptrons. Each multi-head self-attention layer, offset window multi-head self-attention layer, and multilayer perceptron is preceded by a normalization layer. Before each normalization layer, a residual connection connects the features passed through the multi-head self-attention layer, offset window multi-head self-attention layer, or multilayer perceptron. The mathematical expression is as follows:

[0019]

[0020]

[0021]

[0022]

[0023] Where W-MSA represents a windowed multi-head self-attention layer, SW-MSA represents an offset windowed multi-head self-attention layer, MLP represents a multilayer perceptron, LN represents a normalization layer, and x l-1 This represents the output sequence of the previous encoder. This represents the output sequence of the l-th W-MSA block. Let x represent the output sequence of the (l+1)th SW-MSA block. l and x l+1 These represent the output sequences of the l-th and (l+1)-th MLP blocks, respectively.

[0024] Preferably, the semantic features containing local and global information are obtained through an adaptive attention module, and the calculation steps of the adaptive attention module are as follows:

[0025] For the extracted features C i With T i A concat operation is performed, followed by information exchange via a 1×1 convolution. Global average pooling compresses the spatial information to obtain a 1×1×C vector. This vector is then passed through a Sigmoid non-linear activation function to obtain channel feature weights. These weights are then combined with the C vectors after their respective 1×1 convolutions. i With T i Matrix multiplication is performed to obtain the feature map F with channel attention. Ci and F Ti F Ci This indicates a focus on the features extracted by the CNN, F Ti This indicates a focus on the features extracted by the Transformer; finally, the two feature vectors are added together to obtain the fused feature information F. i The mathematical expression is as follows:

[0026]

[0027]

[0028] F i =F Ci +F Ti i = 1, 2, 3, 4

[0029] Where i represents the four stages of network feature extraction, f 1×1 This indicates a convolution with a 1×1 kernel. represents matrix multiplication, Avgpool represents global average pooling, and Concat represents concatenating different feature maps.

[0030] Preferably, the acquisition of feature maps containing rich contextual information is achieved through a multi-scale feature fusion module, and the calculation steps of the multi-scale feature fusion module are as follows:

[0031] Input four feature maps F at different scales i The feature map resolution is adjusted by performing 3×3 convolutions to obtain feature maps of uniform resolution. Then, the four feature maps are concatenated using Concat to obtain image features W containing rich contextual information. The mathematical transformation is shown below:

[0032] W = Concat[f 3×3 (F1),f 3×3 (F2),f 3×3 (F3),f 3×3 (F4)]

[0033] Among them, f 3×3 This indicates a 3×3 convolution kernel, and Concat means concatenating different feature maps, i = 1, 2, 3, 4.

[0034] Preferably, the step of segmenting and classifying the pre-defined test set to obtain the model's output visualization results for model evaluation specifically includes:

[0035] The algorithm performance is evaluated using Accuracy, Precision, Record, F1 score, and mIoU metrics, which are mathematically described as follows:

[0036]

[0037]

[0038]

[0039]

[0040]

[0041] Where Accuracy represents the proportion of correctly classified pixels, Precision represents the proportion of correctly classified positive pixels to all predicted positive pixels, Record represents the proportion of correctly classified positive pixels to all positive pixels, F1 represents the combined performance of Precision and Record, mIoU is the average intersection-union ratio, used to represent the accuracy of image segmentation, and TP, TN, FP, and FN represent the number of true positive, true negative, false positive, and false negative pixels, respectively.

[0042] In a second aspect, a computer-readable storage medium is provided for storing one or more programs, the one or more programs including instructions that, when executed by a computing device, cause the computing device to perform any of the methods described.

[0043] Thirdly, a computing device is provided, comprising:

[0044] One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs include instructions for performing any of the methods described.

[0045] (III) Beneficial Effects

[0046] This invention presents a method for extracting buildings from remote sensing images by integrating convolutional neural networks (CNNs) and Transformers. It addresses the limitation that extracting features from input remote sensing images by reducing feature map size through convolution and pooling operations in neural networks can lead to the loss of detailed information. Furthermore, convolutional operations cannot capture long-range dependencies during feature extraction. This invention combines the powerful local detail feature extraction capabilities of CNNs with the global modeling capabilities of Transformers to extract building information from remote sensing images. A dual-branch network of CNN and Transformer is used to extract image features in parallel. Secondly, an adaptive feature fusion module is designed to exchange information between the feature maps obtained from each stage of the two branches. Finally, a multi-scale fusion module obtains multi-scale feature maps containing rich contextual information. Applying the proposed algorithm to remote sensing image building extraction tasks can significantly reduce the problems of false positives and false negatives for small target buildings, improve the completeness of segmentation for large target buildings, and alleviate the boundary blurring problem caused by insufficient extraction of target building edge information. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the overall process of the remote sensing image building extraction method of the present invention;

[0048] Figure 2 This is a schematic diagram of the overall network structure for building extraction from remote sensing images according to the present invention;

[0049] Figure 3 This is a schematic diagram of the structure of the Swin Transformer Block provided by the present invention;

[0050] Figure 4 This is a schematic diagram of the ResNet Block structure provided by the present invention;

[0051] Figure 5 A schematic diagram of the adaptive attention module structure provided by the present invention;

[0052] Figure 6 This is a schematic diagram of the multi-scale feature fusion module structure provided by the present invention;

[0053] Figure 7 This is a schematic diagram illustrating the building extraction effect of the remote sensing image building extraction method of the present invention. Detailed Implementation

[0054] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0055] Example

[0056] like Figure 1 As shown, this embodiment of the invention provides a method for extracting buildings from remote sensing images by fusing convolutional neural networks and Transformers, including:

[0057] The overall structure of the model is as follows Figure 2 As shown, a parallel dual-branch network of CNN and Transformer is used to extract features from remote sensing images, with each branch containing four feature extraction stages. Then, an adaptive attention module (T) interacts the features extracted in each stage to obtain four different scale features containing both local and global information. These features are then processed by a multi-scale feature fusion module to obtain multi-scale feature maps with rich contextual information. Finally, upsampling is used to restore the feature maps to their original resolution for pixel-level segmentation.

[0058] The specific steps are as follows:

[0059] Step 1: Creating a dataset

[0060] Acquire high-resolution remote sensing imagery, create a dataset containing multiple categories, and label each category. Divide the remote sensing images into 512×512 pixel segments. Randomly partition the dataset into training, validation, and test sets in a 7:2:1 ratio.

[0061] Step 2: Data Preprocessing

[0062] The input images are preprocessed. Preprocessing involves using image enhancement methods to increase data diversity, thereby improving the model's generalization ability. These image enhancements include random rotations from 1° to 360°, random horizontal flips, random vertical flips, random scaling from 0.5 to 1.0x, and random contrast changes. During training, any of these image enhancement methods will be randomly used to input training samples into the designed network model.

[0063] Step 3: Extract features in parallel using a dual-branch network of CNN and Transformer.

[0064] The image from step two is then processed using a CNN and a Transformer for feature extraction. The CNN branch comprises four feature extraction stages, each using a ResNet Block to extract image features. With each ResNet Block, the image resolution is halved, while the number of channels doubles. The ResNet Block structure is as follows: Figure 3 As shown. Given a feature map A, it first passes through a 1×1 convolution for channel information exchange, then a 3×3 convolution for feature extraction, and finally a 1×1 convolution for channel adjustment. The extracted feature C is then added to the input feature map A via a residual connection. A 1×1 convolution is added to the residual connection to downsample the input feature map A. The mathematical expression is as follows:

[0065] C = f 1×1 (f 3×3 (f 1×1 (A)))+f 1×1 (A)

[0066] In the above formula, f 1×1 and f 3×3 These represent convolutions with kernels of 1×1 and 3×3, respectively.

[0067] The parallel branch uses a Transformer to extract image features and also includes four feature extraction stages. Each stage contains a Patch Merging layer and a Swin Transformer Block. The Patch Merging layer downsamples the image to reduce its resolution. Then, the Swin Transformer Block extracts features from the image. The Swin Transformer divides the image into a series of windows, calculates the correlation between each pixel within the window using a self-attention mechanism, and then uses a sliding window to allow information exchange between the independent windows, thereby obtaining the global features of the image.

[0068] Swin Transformer Black Figure 4 As shown, it mainly includes alternating windowed multi-head self-attention layers (W-MSA), offset windowed multi-head self-attention layers (SW-MSA), and multilayer perceptrons (MLP). Each W-MSA, SW-MSA, and MLP is preceded by a normalization layer (LN). Before each LN layer, a residual connection connects the features passed through the W-MSA, SW-MSA, or MLP. The mathematical calculation expression is as follows:

[0069]

[0070]

[0071]

[0072]

[0073] Where W-MSA represents a windowed multi-head self-attention layer, SW-MSA represents an offset windowed multi-head self-attention layer, MLP represents a multilayer perceptron, LN represents a normalization layer, and x l-1 This represents the output sequence of the previous encoder. This represents the output sequence of the l-th W-MSA block. Let x represent the output sequence of the (l+1)th SW-MSA block. l and x l+1 These represent the output sequences of the l-th and (l+1)-th MLP blocks, respectively.

[0074] Step 4: Merge the information from the two branches

[0075] The features extracted from each stage of CNN and Transformer in step three are interacted through an adaptive attention module. First, regarding the features C extracted by CNN and Transformer...i With T i A concat operation is performed, and information exchange between the CNN and Transformer is achieved through a 1×1 convolution. Then, global average pooling (AvgPooling) is used to compress the spatial information into a 1×1×C vector. This vector is then passed through a Sigmad non-linear activation function to obtain channel feature weights. Finally, these weights are combined with the C vectors that have undergone 1×1 convolutions. i With T i Matrix multiplication is performed to obtain the feature map F with channel attention. Ci and F Ti F Ci This indicates a focus on the features extracted by the CNN, F Ti This indicates a focus on the features extracted by the Transformer. Finally, the two feature vectors are added together to obtain the fused feature information F. i Its structure is as follows: Figure 5 As shown. The mathematical expression is as follows:

[0076]

[0077]

[0078] F i =F Ci +F Ti i = 1, 2, 3, 4

[0079] In the above formula, i represents the four stages of network feature extraction, f 1×1 This indicates a convolution with a 1×1 kernel. represents matrix multiplication, Avgpool represents global average pooling, and Concat represents concatenating different feature maps.

[0080] The four feature maps at different scales obtained after the interaction between the CNN and the Transformer are processed by a multi-scale feature fusion module to obtain rich contextual semantic feature information. The structure of the multi-scale feature fusion module is as follows: Figure 6 As shown. Input four feature maps F at different scales. i (i = 1, 2, 3, 4) are each subjected to 3×3 convolution to adjust the feature map resolution, resulting in feature maps of uniform resolution. Then, the four feature maps are concatenated using Concat to obtain image feature W containing rich contextual information. Its structure is as follows: Figure 6 As shown. The mathematical transformation is as follows:

[0081] W = Concat[f 3×3 (F1),f 3×3 (F2),f 3×3 (F3),f3×3 (F4)]

[0082] In the above formula, f 3×3 This indicates a convolution with a 3×3 kernel, and Concat means concatenating different feature maps.

[0083] Afterwards, the upsampling module uses bilinear interpolation to upsample the merged feature map, restoring the feature map size to its original resolution. Finally, the segmented image is output after a 1×1 convolution.

[0084] Step 5: Model Training

[0085] After the data preprocessing described in step two, the training and validation sets obtained in step one are input into the designed network for training. The network weights are updated using backpropagation. Before model training, appropriate hyperparameters are set and initialized: the number of iterations is set to 500 rounds, the number of remote sensing images input to the network model each time is 4, the Adam optimizer is selected, the initial learning rate is set to 0.002, and a cosine decay strategy is used during training to prevent the model from getting trapped in local optima. The cross-entropy loss function is used to measure the closeness between the input and output.

[0086] Step Six: Model Evaluation

[0087] The model for building extraction from remote sensing imagery, which integrates convolutional neural networks and Transformers, is used to segment the test set defined in step one, yielding the visualization results of the model output. (See attached image.) Figure 7 As shown, where Figure 7 (a) shows the input remote sensing image. Figure 7 In the middle (b), the output segmented image is shown.

[0088] This invention uses metrics such as Accuracy, Precision, Record, F1, and mIoU to evaluate algorithm performance. Their mathematical descriptions are as follows:

[0089]

[0090]

[0091]

[0092]

[0093]

[0094] In the above formula, Accuracy represents the proportion of correctly classified pixels, Precision represents the proportion of correctly classified positive pixels to all predicted positive pixels, Record represents the proportion of correctly classified positive pixels to all positive pixels, F1 represents the combined performance of Precision and Record, mIoU is the average intersection-union ratio, used to represent the accuracy of image segmentation, and TP, TN, FP, and FN represent the number of true positive, true negative, false positive, and false negative pixels, respectively.

[0095] Experiments have demonstrated that this method is highly effective for building extraction from remote sensing images. Specific metrics include: Accuracy of 0.98, Precision of 0.93, Recalculation of 0.96, F1 score of 0.94, and mIoU of 0.81.

[0096] Embodiments of this application may be provided as methods or computer program products. Therefore, this application may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application 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. The solutions in the embodiments of this application may be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0097] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0098] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The function specified in one or more boxes.

[0099] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0100] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

Claims

1. A method for building extraction from remote sensing images by fusing convolutional neural networks and Transformers, characterized in that, The application relates to a remote sensing image building information extraction method and device. The remote sensing image is processed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; To the extracted feature C i With T i The information interaction through the Concat operation and a 1x1 convolution, the spatial information compression through the global average pooling to get a 1x1xC vector, and the channel feature weight through a Sigmoid nonlinear activation function, the obtained weight and C i With T i The matrix multiplication to get the feature map F with channel attention Ci And F Ti , F Ci Indicates that the extracted features of CNN are focused on, and F Ti Indicates that the extracted features of Transformer are focused on; finally, the two feature vectors are added to get the fused feature information F i , the mathematical expression is as follows: F i = F Ci + F Ti i = 1, 2, 3, 4 where i represents four stages of network feature extraction, f 1×1 denotes a convolution with a 1x1 convolution kernel, denotes matrix multiplication, Avgpool denotes global average pooling, and Concat denotes concatenating different feature maps; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; Input four different scale feature maps F i Resolutions of the feature maps are adjusted by 3x3 convolution respectively, and feature maps with uniform resolution size are obtained. Then, four feature maps are spliced by Concat to obtain image features W containing rich context information. The mathematical transformation is shown as follows: W = Concat[f 3×3 (F1), f 3×3 (F2), f 3×3 (F3), f 3×3 (F4)] wherein f 3×3 denotes a convolution with a 3x3 kernel, Concat denotes concatenation of different feature maps, and i = 1, 2, 3, 4. 2.The method of claim 1, wherein the method comprises: The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; C = f 1×1 (f 3×3 (f 1×1 (A)))+f 1×1 (A) The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; 3.The method of claim 2, wherein the method comprises: The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; where W-MSA denotes a windowed multi-head self-attention layer, SW-MSA denotes a shifted windowed multi-head self-attention layer, MLP denotes a multi-layer perceptron, LN denotes a normalization layer, x l-1 denotes the output sequence of the previous encoder, denotes the output sequence of the l-th block of W-MSA, denotes the output sequence of the (l+1)-th block of SW-MSA, x l and x l+1 denote the output sequences of the l-th and (l+1)-th blocks of MLP, respectively. 4.The method of claim 1, wherein the method further comprises: The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a remote sensing image, and the diversity of data is increased; The remote sensing image is preprocessed to obtain a Wherein, the Accuracy represents the proportion of correctly classified pixels, the Precision represents the proportion of correctly classified positive pixels in all predicted positive pixels, the Recall represents the proportion of correctly classified positive pixels in all positive pixels, the F1 represents an index of comprehensive Precision and Recall performance, the mIoU is an average intersection over union, and is used to represent the accuracy of image segmentation, and the TP, TN, FP and FN represent the number of pixel points of true positive, true negative, false positive and false negative, respectively.

5. A computer-readable storage medium storing one or more programs, the one or more programs comprising instructions that when executed by a computer cause the computer to perform a method comprising: The one or more programs include instructions, which when executed by a computing device, cause the computing device to perform any of the methods of claims 1-4.

6. A computing device, comprising: Comprise: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs comprising instructions for performing any of the methods of claims 1-4.

Images