A method, system and program product for extracting components of an alpine wetland landscape

Abstract

The application belongs to the technical field of machine learning, and specifically discloses a plateau wetland landscape component extraction method, system and program product. Through the design of a spectrum-space double-branch interaction model, the spectral features of the image elements are extracted through the image element dimension increasing module and the ViT architecture Transformer encoder in the spectrum branch, the neighborhood spatial features of the image elements are extracted through the multi-scale feature extraction module and the feature enhancement in the space branch, then the features of the two branches are integrated through the cross attention fusion module, and the image element classification result is output, so as to determine the plateau wetland landscape component regions in the remote sensing image according to the image element classification results, and the efficient and reliable extraction of the plateau wetland landscape components can be realized. The application can effectively alleviate the salt and pepper noise problem existing in the traditional pixel-by-pixel classification method, improve the classification accuracy of the plateau wetland landscape components, ensure the lightweight of the network, and provide more accurate technical support for the long-time sequence plateau wetland landscape component extraction.

Description

Technical Field

[0001] This invention belongs to the field of machine learning technology, specifically relating to a method, system, and program product for extracting components of plateau wetland landscapes. Background Technology

[0002] Wetlands are important ecosystems formed through complex hydrological activities. As a special type of wetland, plateau wetlands play a vital role in water conservation, biodiversity protection, and climate regulation. However, accelerated climate change and intensive human activities have led to significant changes in wetland composition. Due to their unique geological structure and meteorological and phenological variations, the wetland ecosystem is fragile and difficult to restore. Therefore, monitoring changes in the landscape components of plateau wetlands is crucial for protecting their sustainable development.

[0003] Remote sensing technology, with its advantages of large-scale and multi-temporal coverage, is an effective means of acquiring information about the Earth's surface. Remote sensing imagery can effectively characterize the distribution of wetland landscape elements, and extracting and analyzing landscape components from multi-temporal wetland remote sensing images has become an important method for wetland dynamic monitoring. However, the construction of sample datasets and the selection of effective wetland component extraction methods are the main obstacles to accurately monitoring long-term dynamic changes in wetlands.

[0004] For extracting wetland landscape components from long-term wetland imagery, traditional object-oriented semantic segmentation methods suffer from difficulties in sample preparation, high manpower requirements, and the accumulation of errors and noise. Pixel-by-pixel classification methods typically consider only pixel spectral information while ignoring spatial relationships between pixels, leading to discrete isolated points (salt-and-pepper phenomenon) and insufficient classification accuracy. Therefore, existing wetland landscape component extraction methods are ineffective at extracting complex and variable long-term wetland landscape components, necessitating a more efficient and accurate method for extracting components from plateau wetlands. Summary of the Invention

[0005] The purpose of this invention is to provide a method, system, and program product for extracting components of plateau wetland landscapes, in order to solve the above-mentioned problems existing in the prior art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] Firstly, a method for extracting components of plateau wetland landscapes is provided, including:

[0008] Acquire feature-fused remote sensing image data of the target plateau wetland area, wherein the feature-fused remote sensing image data contains a pixel sequence composed of several pixels;

[0009] Feature-fused remote sensing image data is input into a pre-set spectral-spatial bi-branch interactive model for classification of plateau wetland landscape components, resulting in classified remote sensing image data containing each pixel and its classification results. The spectral-spatial bi-branch interactive model includes a spectral feature extraction branch, a spatial feature extraction branch, and a cross-attention fusion module. The spectral feature extraction branch is used to extract the fused spectral features of each pixel, the spatial feature extraction branch is used to extract the neighborhood spatial features of each pixel, and the cross-attention fusion module is used to combine the fused spectral features and neighborhood spatial features of each pixel and input them into a fully connected layer to classify plateau wetland landscape components for each pixel, obtaining the classification results for each pixel.

[0010] The pixel regions corresponding to each plateau wetland landscape component in the classified remote sensing image data were determined, and the pixel regions corresponding to each plateau wetland landscape component were extracted.

[0011] In one possible design, the feature-fused remote sensing image data is multi-level feature-fused remote sensing image data containing 6 spectral bands, 2 spectral features, and 4 spatial features. The 6 spectral bands include blue spectral band, green spectral band, red spectral band, near-infrared spectral band, first shortwave infrared spectral band, and second shortwave infrared spectral band. The 2 spectral features include normalized vegetation index (NVI) and enhanced water index. The 4 spatial features include slope topography, aspect topography, texture features, and spatial statistical distribution features.

[0012] In one possible design, before inputting the feature-fused remote sensing image data into a pre-set spectral-spatial bibranch interaction model for classification of plateau wetland landscape components, the method further includes: constructing a spectral-spatial bibranch interaction model, and using a training set, a validation set, and a test set to train, validate, and test the spectral-spatial bibranch interaction model, wherein the training set, validation set, and test set each contain several pixel samples after the plateau wetland landscape component classification and labeling.

[0013] In one possible design, the spectral feature extraction branch includes a pixel upscaling module and a Transformer module. The pixel upscaling module is used to transform the input pixels into corresponding two-dimensional feature images using the Grami angle field method, the Markov transition field method, and the recursive graph method, respectively. Then, the two-dimensional feature images are merged into a three-channel feature image through a concat operation, and then bilinear interpolation is performed on the three-channel feature image to obtain the upscaled feature image. The Transformer module is used to extract the multi-dimensional fused spectral features of the upscaled feature image. The Transformer module includes a weighted average ViT structure Transformer encoder, and the Transformer encoder consists of 4 Transformer layers, each with the same weights. The Transformer layers are sequentially set with a layer normalization unit, a multi-head self-attention unit, a layer normalization unit, and a multilayer perceptron unit in the computation direction, and skip residual connections are performed after the multi-head self-attention unit and the multilayer perceptron unit.

[0014] In one possible design, the spectral feature extraction branch also includes a cropping and mapping module located between the pixel upscaling module and the Transformer module. The cropping and mapping module is used to crop the upscaling feature image into blocks and then perform linear mapping. Then, block embedding and position embedding are performed on each block after linear mapping to obtain multiple sets of one-dimensional embedding sequences. Each set of one-dimensional embedding sequences is then input into the corresponding Transformer layer.

[0015] In one possible design, the spatial feature extraction branch includes a multi-scale feature extraction module and a feature enhancement module. The multi-scale feature extraction module first extracts the neighborhood pixel group of the corresponding pixel as input using a 3×3 size mask. Then, it performs 1×1 group convolution and 3×3 group convolution on the input respectively. Then, it concatenates the 1×1 group convolution results and 3×3 group convolution results through a concat operation to obtain joint features. Finally, it performs deformable convolution on the joint features to obtain multi-scale neighborhood spatial features. The feature enhancement module is used to enhance the multi-scale neighborhood spatial features to obtain enhanced neighborhood spatial features.

[0016] In one possible design, the feature enhancement module first performs global average pooling on the input multi-scale neighborhood spatial features, then uses a 1×1 convolution kernel with ReLU activation function and a 1×1 convolution kernel with Sigmoid activation function to continuously capture the non-linear dependencies between channels in the channel dimension, and then uses a 3×3 convolution kernel to perform context modeling in the spatial dimension to obtain the enhanced neighborhood spatial features.

[0017] In one possible design, the cross-attention fusion module uses a multi-head cross-attention mechanism to associate the fused spectral features of the spectral feature extraction branch with the neighborhood spatial features of the spatial feature extraction branch to obtain spectral-spatial joint features. Then, after processing the spectral-spatial joint features through three fully connected layers, the classification result of the corresponding pixel is output by the Softmax function.

[0018] Secondly, a system for extracting components from plateau wetland landscapes is provided, including:

[0019] Memory, used to store instructions;

[0020] The processor is configured to read instructions stored in the memory and execute any one of the plateau wetland landscape component extraction methods described in the first aspect above, according to the instructions.

[0021] Thirdly, a computer-readable storage medium is provided, on which instructions are stored, which, when executed on a computer, cause the computer to perform any one of the plateau wetland landscape component extraction methods described in the first aspect. Simultaneously, a computer program product is also provided, which, when executed on a computer, performs any one of the plateau wetland landscape component extraction methods described in the first aspect.

[0022] Beneficial Effects: This invention designs a spectral-spatial dual-branch interactive model. In the spectral branch, spectral features of pixels are extracted through a pixel upscaling module and a ViT architecture Transformer encoder. In the spatial branch, neighborhood spatial features of pixels are extracted through a multi-scale feature extraction module and feature enhancement. Then, the features from both branches are integrated through a cross-attention fusion module to output pixel classification results. Based on the classification results of each pixel, the regions of each plateau wetland landscape component in remote sensing images can be determined, achieving efficient and reliable extraction of plateau wetland landscape components. This invention can effectively alleviate the salt-and-pepper noise problem of traditional pixel-by-pixel classification methods, improving the classification accuracy of plateau wetland landscape components while ensuring the lightweight nature of the network, providing more accurate technical support for the extraction of long-term plateau wetland landscape components. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating the method in Embodiment 1 of the present invention;

[0025] Figure 2 This is a schematic diagram of the structure of the spectral-spatial bibranch interaction model;

[0026] Figure 3 This is a schematic diagram of the pixel upscaling module.

[0027] Figure 4 This is a structural diagram of the Transformer module;

[0028] Figure 5 This is a schematic diagram of the multi-scale feature extraction module.

[0029] Figure 6 This is a schematic diagram of the feature enhancement module.

[0030] Figure 7 Example figure showing the wetland landscape component extraction results of the spectral-spatial bibranch interaction model;

[0031] Figure 8 A comparative diagram showing the landscape component extraction results of different models;

[0032] Figure 9 A comparative diagram showing the extraction results of landscape components in the ablation experiment;

[0033] Figure 10 This is a schematic diagram of the system configuration in Embodiment 2 of the present invention. Detailed Implementation

[0034] It should be noted that the descriptions of these embodiments are intended to aid in understanding the invention and do not constitute a limitation thereof. The specific structural and functional details disclosed herein are merely for describing exemplary embodiments of the invention. However, the invention may be embodied in many alternative forms and should not be construed as being limited to the embodiments described herein.

[0035] It should be understood that, unless otherwise explicitly specified and limited, the corresponding terms should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be a connection within two components. Those skilled in the art can understand the specific meaning of the above terms in the embodiments according to the specific circumstances.

[0036] Specific details are provided in the following description to provide a complete understanding of the exemplary embodiments. However, those skilled in the art will understand that the exemplary embodiments can be implemented without these specific details. For example, apparatus may be shown in block diagrams to avoid obscuring the examples with unnecessary details. In other embodiments, well-known processes, structures, and techniques may be omitted with non-essential details to avoid obscuring the embodiments.

[0037] Example 1:

[0038] This embodiment provides a method for extracting components of plateau wetland landscapes, which can be applied to corresponding remote sensing image data processing terminals, such as... Figure 1 As shown, the method includes the following steps:

[0039] S1. Acquire feature-fused remote sensing image data of the target plateau wetland area, wherein the feature-fused remote sensing image data contains a pixel sequence composed of several pixels.

[0040] In practice, initial remote sensing image data of the target plateau wetland area is collected and preprocessed accordingly to obtain multi-level feature fusion remote sensing image data containing 6 spectral bands, 2 spectral features, and 4 spatial features. The feature fusion remote sensing image data can be resolved into a pixel sequence composed of several pixels. The 6 spectral bands include blue (SR_B2), green (SR_B3), red (SR_B4), near-infrared (SR_B5), shortwave infrared 1 (SR_B6), and shortwave infrared 2 (SR_B7) bands. The 2 spectral features include normalized vegetation index (NVI) and enhanced water index. The 4 spatial features include slope topography, aspect topography, texture, and spatial statistical distribution.

[0041] S2. Input the feature-fused remote sensing image data into a pre-set spectral-spatial bi-branch interactive model to classify plateau wetland landscape components, obtaining classified remote sensing image data containing each pixel and its classification results. The spectral-spatial bi-branch interactive model includes a spectral feature extraction branch, a spatial feature extraction branch, and a cross-attention fusion module. The spectral feature extraction branch is used to extract the fused spectral features of each pixel, the spatial feature extraction branch is used to extract the neighborhood spatial features of each pixel, and the cross-attention fusion module is used to combine the fused spectral features and neighborhood spatial features of each pixel and input them into a fully connected layer to classify plateau wetland landscape components of each pixel, obtaining the classification results of each pixel.

[0042] In specific implementation, a spectral-spatial bi-branch interaction model is pre-constructed on the processing terminal, and a training set, a validation set, and a test set are used to train, validate, and test the spectral-spatial bi-branch interaction model. The training set, validation set, and test set each contain several pixel samples after the classification and labeling of plateau wetland landscape components.

[0043] For example, the training, validation, and test sets can be derived from the Landsat 8 / 9 atmospheric corrected surface reflectance dataset provided by the Google Earth Engine platform. Multiple initial remote sensing images of the target plateau wetland area during a specified time period are selected, and then preprocessed, including histogram matching for color balancing, image mosaicking, band selection, edge cropping, cloud removal using the "QA_PIXEL" and "ST_CDIST" bands, multispectral panchromatic image fusion, and manual secondary cloud removal and hole repair, to obtain high-quality cloud-free remote sensing images. Then, for the cloud-free remote sensing image data, six spectral bands—blue (SR_B2), green (SR_B3), red (SR_B4), near-infrared (SR_B5), shortwave infrared 1 (SR_B6), and shortwave infrared 2 (SR_B7)—are selected to form multispectral image data. On one hand, the Normalized Difference Vegetation Index (NDVI) and Enhanced Water Index (EDI) were calculated as spectral features. On the other hand, 12.5m resolution digital elevation model (DEM) data of the corresponding region was downloaded from the data center, resampled to 15m resolution, and the slope and aspect topographic features were calculated. Furthermore, the first principal component was extracted using principal component analysis to achieve feature dimensionality reduction of the multispectral image data. The mean statistic of its gray-level co-occurrence matrix was calculated as a texture feature. Then, the multi-scale segmentation vector of the multispectral image data was calculated, and the regional mean statistic was used as a spatial statistical distribution feature. Finally, a feature-fused remote sensing image sample containing 6 spectral bands, 2 spectral features, and 4 spatial features, totaling 12 dimensions, was constructed.

[0044] The wetland landscape elements in the target plateau wetland area were then categorized into nine types: river-lake, marsh, marshy meadow, meadow, bare peatland, meadow-shrubland, sandy meadow, bare sandy gravel, and building-road. Following the criteria of balanced quantity and spatial distribution, and avoiding mixed pixels, the feature-fused remote sensing image samples were labeled, resulting in a total of 26,449 labeled pixels. 80% of these pixels were randomly selected for the training set, 20% for the validation set, and an additional 1,452 pixels were labeled for the test set. Finally, the spectral-spatial bibranch interaction model was trained, validated, and tested using the training, validation, and test sets, yielding a satisfactory spectral-spatial bibranch interaction model.

[0045] like Figure 2As shown, the spectral-spatial dual-branch interaction model includes a spectral feature extraction branch, a spatial feature extraction branch, and a cross-attention fusion module. The spectral feature extraction branch is used to extract the fused spectral features of each pixel, the spatial feature extraction branch is used to extract the neighborhood spatial features of each pixel, and the cross-attention fusion module is used to combine the fused spectral features and neighborhood spatial features of each pixel and input them into a fully connected layer to classify the plateau wetland landscape components of each pixel, thereby obtaining the classification results of each pixel.

[0046] The spectral feature extraction branch includes a pixel upscaling module and a Transformer module. The structure of the pixel upscaling module is as follows: Figure 3 As shown, for the input cell x i ∈X={x1,x2,...,x n The Gramian Angular Field (GAF), Markov Transition Field (MTF), and Recurrence Plot (RP) methods were used to transform the data into two-dimensional feature images.

[0047] Grami angle field rescales the given sequence x according to equation (1). i To the interval [0,1]:

[0048]

[0049] Then, using the encoded value as the angle cosine and the time series stamp as the radius, the scaled sequence is transformed to a polar coordinate system:

[0050]

[0051] Among them, t i It is a time series stamp. This is a constant factor used to regularize the span of the polar coordinate system. Finally, the sum of angles between any two points (Gramian Summation Angular Field, GASF) is calculated to construct a symmetric two-dimensional matrix:

[0052]

[0053] Where I is a unit row vector.

[0054] Markov transition fields discretize sequential data into Q quantile intervals according to their value ranges, and assign each pixel x... i Assign the corresponding interval q j (j∈[1,Q]), and then construct a weighted adjacency matrix W of size Q×Q, which can be expressed as:

[0055]

[0056] Among them W i,j From quantile q j The middle data is followed by the quantile q i The frequency of adjacent data points determines the final Markov transition field, as shown in equation (5).

[0057]

[0058] The recursive graph reconstructs the phase space of the input sequence to obtain a high-dimensional state vector, which can be represented as:

[0059] Y(i)=[x i ,x i+τ ,...,x i +(m-1) τ (6)

[0060] Where m is the embedding dimension and τ is the time delay, the similarity between two points is determined by calculating whether the Euclidean distance between the points is less than the preset threshold ε, thereby generating a recursive graph. The calculation process is shown in equation (7):

[0061] R(i,j)=Θ(ε-||Y(i)-Y(j)||) (7)

[0062] The original 12×1×1 input pixels are transformed into 1×8×8 two-dimensional feature maps using three different dimensionality upscaling methods. These feature maps are then merged into a 3×8×8 three-channel feature image using a concat operation. Finally, bilinear interpolation is used to obtain a 3×64×64 upscaling feature image that conforms to the input size of the subsequent network. The calculation process is shown in equations (8) and (9).

[0063] I concat =Concat(I GAF ,I MTF ,I RP (8)

[0064]

[0065] Where x and y are the target coordinates, and w is the weight obtained from the normalized offset calculation.

[0066] The structure of the Transformer module is as follows: Figure 4As shown, the ViT structure Transformer encoder includes a weighted average, and the Transformer encoder consists of four Transformer layers, each with the same weights. Each Transformer layer sequentially includes a layer normalization unit, a multi-head self-attention unit, another layer normalization unit, and a multilayer perceptron unit along the computation direction. Skip residual connections are performed after computation by the multi-head self-attention unit and the multilayer perceptron unit. ViT performs contextual modeling of global pixels in the input image, enabling long-distance spatial dependency capture.

[0067] For the feature map after pixel upscaling, ViT can adaptively focus on the weight ratio of each dimension of the pixel, enhancing effective spectral-spatial features and suppressing invalid information. This method designs a weighted average ViT structure Transformer encoder, with each layer having the same weights. The Transformer receives a one-dimensional embedding sequence as input. Therefore, a corresponding cropping mapping module needs to be set between the pixel upscaling module and the Transformer module. The cropping mapping module is used to crop the upscaled feature image into blocks and perform linear mapping. Then, block embedding and position embedding are performed on each linearly mapped block to obtain multiple sets of one-dimensional embedding sequences. Each set of one-dimensional embedding sequences is then input into the corresponding Transformer layer. For example, the cropping mapping module inputs the feature X∈R... 3×H×W Cut evenly into Group of pixel blocks X∈R 3×8×8 After flattening the pixel blocks, they are mapped to a D-dimensional embedding space that is invariant to the Transformer layer, and then e is embedded in each block. i Add learnable positional embedding p i Form a one-dimensional embedding sequence Z = {e1+p1,e2+p2,...,e L +p L}, Z∈R L×C , where L is the sequence length.

[0068] The Transformer encoder consists of four Transformer layers. Each Transformer layer comprises multi-head self-attention (MSA), a multi-layer perceptron (MLP), and layer normalization (LN). After LN+MSA and LN+MLP computation, skip residual connections are used to preserve shallow layer information to avoid gradient vanishing. The computation process can be represented as follows:

[0069] z = x + MSA(LN(x)) (10)

[0070] y=z+MLP(LN(z)) (11)

[0071] The outputs of the four Transformer layers are collected and weighted to average them, so as to retain both the fine-grained information of the shallow layers and the abstract semantic information of the deep layers.

[0072] The spatial feature extraction branch includes a multi-scale feature extraction module and a feature enhancement module. The structure of the multi-scale feature extraction module is as follows: Figure 5 As shown, a 3×3 mask is used to extract the neighborhood pixel group X∈R of a pixel. C×3×3 (C is the number of input channels) is used as input, and each channel is processed by a 1×1 grouped convolution f. 1×1 and 3×3 grouped convolution f 3×3 Extracting spatial adjacency relationships between pixels at different scales. Grouped convolution divides the output channel number D into g groups, ensuring that each group's convolution kernel is associated with only a portion of the input features, significantly reducing the number of convolution parameters. The input is processed by GELU(f)... 1×1 (X)) and GELU(f 3×3 (X) (GELU is the activation function) yields spatial features F at different scales. i×i =[F 1×1 ,F 3×3 ]∈R 3×3×D Then, the joint features F∈R are obtained by concatenating them through the concat operation. 3×3×2D Finally, deformable convolution is performed on the joint features to obtain multi-scale neighborhood spatial features. By adding an offset to the receptive field, the ability to detect edge changes of ground objects is enhanced. The calculation process is shown in Equation (12):

[0073]

[0074] Where p0 is the coordinate position of the input feature map, p n It is a relative coordinate, Δp n It is the learnable offset for each sampling point.

[0075] The structure of the feature enhancement module is as follows: Figure 6 As shown, for the input multi-scale neighborhood spatial features F∈R C×H×W (Where C is the number of channels, H is the feature height, and W is the feature width) First, global average pooling is performed. Then, in the channel dimension, a 1×1 convolution kernel with ReLU activation function and a 1×1 convolution kernel with Sigmoid activation function are used to continuously capture the non-linear dependencies between channels. Then, a 3×3 convolution kernel is used to perform context modeling in the spatial dimension to obtain the enhanced neighborhood spatial features, thereby achieving adaptive calibration of the feature map in the spatial dimension.

[0076] The cross-attention fusion module employs a multi-head cross-attention mechanism to associate the fused spectral features of the spectral feature extraction branch with the neighborhood spatial features of the spatial feature extraction branch, thereby obtaining joint spectral-spatial features. Where B is the batch size, D is the fixed dimension in the Transformer layer, and P is the side length of the Transformer encoder embedding block. The cross-attention calculation formula is shown in equation (13):

[0077]

[0078] Where Q is the query vector matrix of the first sequence, K and V are the key-value query matrices of the second sequence, and d k This refers to the dimension of the key vector. First, the spectral branch features are linearly mapped to a Q vector, and the spatial branch features are linearly mapped to K and V vectors. Then, Q, K, and V are divided into h independent head vectors. i =QW i Q K i =KW i K V i =VW i V W i Q W i K W i V This is the linear projection matrix of each head. Finally, after parallel computation on multiple attention heads and merging, the spectral-spatial joint feature F is obtained. The multi-head attention computation can be expressed as:

[0079]

[0080] MultiHeadCross=Concat(head1,...,head h (15)

[0081] The spectral-spatial joint feature F is then processed through three fully connected layers, and the classification result of the corresponding pixel is output by the Softmax function.

[0082] After the spectral-spatial bibranch interaction model is constructed, trained, validated, and tested, the processing terminal inputs the feature-fused remote sensing image data into the spectral-spatial bibranch interaction model to classify the plateau wetland landscape components, thereby obtaining classified remote sensing image data containing each pixel and its classification results.

[0083] S3. Determine the pixel regions corresponding to each plateau wetland landscape component in the classified remote sensing image data, and extract the pixel regions corresponding to each plateau wetland landscape component.

[0084] In practice, the processing terminal can determine the pixel region corresponding to each plateau wetland landscape component in the remote sensing image based on the classified remote sensing image data, and extract the pixel region corresponding to each plateau wetland landscape component.

[0085] To evaluate the differences between the classification results of the proposed spectral-spatial bibranch interaction model and the actual land cover, overall accuracy (OA), macro-average precision (mAP), average recall (AR), mean intersection over union (mIoU), and macro-average F1 score (Macro-F1) were selected as evaluation metrics to comprehensively measure the classification performance. Comparative experiments were conducted using AlexNet, UNet, and ViT classification models as benchmark methods to extract wetland landscape components from a plateau wetland region. The wetland landscape component extraction results of this method's spectral-spatial bibranch interaction model are shown below. Figure 7 As shown in Table 1 below, the comparative experimental results are as follows:

[0086] Table 1

[0087]

[0088] Experimental results show that, while achieving high levels of training accuracy (Train Acc) and validation accuracy (Val Acc) for all methods, our proposed method outperforms other models in all metrics. Furthermore, the difference in test accuracy compared to training and validation accuracy is smaller, indicating that our method effectively alleviates overfitting and achieves better generalization. Notably, ViT and AlexNet outperform UNet in all metrics, while UNet, with its multi-scale skip connection mechanism in its encoder-decoder structure, effectively captures spectral-spatial features across different dimensions. Our proposed method integrates a spatial convolution branch while extracting features using the Transformer encoder, achieving collaborative modeling of spectral-spatial features. OA, AR, and mIoU are improved by 3.67%, 2.54%, and 5.39% respectively compared to UNet, demonstrating that our model provides more accurate landscape component category detection and clearer boundary segmentation. A comparison of landscape extraction results from different methods is shown in Table 2 below.

[0089] Table 2

[0090]

[0091]

[0092] Further landscape component analysis in Table 2 shows that this method achieves the best results in both F1 and IoU indices, except for the swamp category. This phenomenon may stem from the high spatial correlation between swamps and adjacent land features such as marshy meadows and river-lake areas, leading to the model being overly sensitive to local texture features. The landscape component extraction results of different methods are shown below. Figure 8 As shown in the first row of images, compared to the contrasting methods which exhibit fragmentation in the river-lake segmentation, our model achieves accurate segmentation of continuous water bodies through spectral-feature enhancement. Furthermore, in the bare peatland area in the lower right corner of the image, our model avoids misclassification caused by local spectral confusion in other methods through a multi-scale neighborhood feature extraction module. In the second row of images, AlexNet and ViT exhibit blurring at the boundaries between bare peatland and marshy meadow categories, which is related to their limitations in relying on single-scale spectral upscaling. While UNet has relatively clear landscape component boundaries, insufficient feature fusion in the decoder leads to severe underclassification of bare sand and gravel. Our model, through the cross-channel interaction mechanism of the Transformer encoder, can deeply mine spectral features of different dimensions, and combined with the spatial features after feature enhancement, significantly enhances spectral discrimination capabilities and improves boundary segmentation accuracy. In the third and fourth rows, other methods suffer from confusion between bare sand and gravel and buildings / roads, and significant salt-and-pepper effect. Our model effectively alleviates this problem by capturing and enhancing neighborhood spatial information through a multi-scale feature extraction module.

[0093] To verify the effectiveness of the spectral feature extraction branch and the spatial feature extraction branch in the proposed method model, ablation experiments were conducted. Classification and accuracy evaluation were performed using only the spectral feature extraction branch (a) and the spectral feature extraction branch combined with the multi-scale feature extraction module (b). The results were then compared with those obtained from the complete network. The ablation experiment results are shown in Table 3 below.

[0094] Table 3

[0095]

[0096]

[0097] Compared to model a, adding module b improved OA, mAP, AR, mIoU, and Macro-F1 scores by 2.42, 2.05, 2.12, 3.76, and 2.16 percentage points, respectively, demonstrating that the addition of the neighborhood pixel space effectively improves classification accuracy. Further adding the feature enhancement module (c) to the complete model improved OA and mIoU by 1.04 and 0.61 percentage points, respectively, proving that the feature enhancement module can enhance ground feature boundaries and improve boundary detection accuracy. The landscape component extraction results from the ablation experiment are as follows: Figure 9As shown, although the spectral feature extraction branch can basically identify various landscape components, it often mistakenly mixes isolated noise points from other categories in each category due to the lack of spatial context constraints, and the boundaries are not complete and continuous enough. The addition of the multi-scale feature extraction module in a+b can effectively capture the spatial correlation of the neighborhood, effectively alleviate the salt-and-pepper phenomenon, and greatly improve the classification accuracy. After adding the feature enhancement module in a+b+c, the boundaries of land features are clearer and the division of the transition area between different categories is more accurate.

[0098] In summary, the spectral-spatial bi-branch interaction model provided by this method can achieve efficient and reliable extraction of plateau wetland landscape components. It can effectively alleviate the salt-and-pepper noise problem of traditional pixel-by-pixel classification methods, improve the classification accuracy of plateau wetland landscape components while ensuring the lightweight nature of the network, and provide more accurate technical support for the extraction of plateau wetland landscape components over long time periods.

[0099] Example 2:

[0100] This embodiment provides a system for extracting components from plateau wetland landscapes, such as... Figure 10 As shown, at the hardware level, it includes:

[0101] The data interface is used to establish data communication between the processor and external data terminals;

[0102] Memory, used to store instructions;

[0103] The processor is used to read the instructions stored in the memory and execute the plateau wetland landscape component extraction method in Example 1 according to the instructions.

[0104] Optionally, the system also includes an internal bus, through which the processor, memory, and data interface can be interconnected. This internal bus can be a PCIe (Peripheral Component Interconnect Express) bus, which can be divided into an address bus, a data bus, a control bus, etc. The memory can include, but is not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Flash Memory, First Input First Output (FIFO), and / or First In Last Out (FILO). The processor can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0105] Example 3:

[0106] This embodiment provides a computer-readable storage medium storing instructions. When these instructions are executed on a computer, the computer performs the plateau wetland landscape component extraction method described in Embodiment 1. The computer-readable storage medium refers to a data storage medium, which may include, but is not limited to, floppy disks, optical disks, hard disks, flash memory, USB flash drives, and / or Memory Sticks. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices.

[0107] This embodiment also provides a computer program product that, when run on a computer, executes the plateau wetland landscape component extraction method of Embodiment 1. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device.

[0108] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for extracting components from plateau wetland landscapes, characterized in that, include: Acquire feature-fused remote sensing image data of the target plateau wetland area, wherein the feature-fused remote sensing image data contains a pixel sequence composed of several pixels; Feature-fused remote sensing image data is input into a pre-defined spectral-spatial bi-branch interactive model for classification of plateau wetland landscape components, resulting in classified remote sensing image data containing each pixel and its classification results. The spectral-spatial bi-branch interactive model includes a spectral feature extraction branch, a spatial feature extraction branch, and a cross-attention fusion module. The spectral feature extraction branch extracts the fused spectral features of each pixel, and includes a pixel upscaling module and a Transformer module. The pixel upscaling module converts the input pixels into corresponding two-dimensional feature images using the Grami angle field method, Markov transition field method, and recursive graph method, respectively. These two-dimensional feature images are then merged into a three-channel feature image using a concat operation, followed by bilinear interpolation to obtain the upscaled feature image. The Transformer module extracts the multi-dimensional fused spectral features of the upscaled feature image, and includes a weighted average ViT structure Transformer. The encoder, specifically the Transformer encoder, consists of four Transformer layers, each with the same weights. Each Transformer layer sequentially includes a layer normalization unit, a multi-head self-attention unit, another layer normalization unit, and a multilayer perceptron unit along the computation direction. Skip residual connections are performed after computation by the multi-head self-attention unit and the multilayer perceptron unit. The spatial feature extraction branch extracts the neighborhood spatial features of each pixel. This branch includes a multi-scale feature extraction module and a feature enhancement module. The multi-scale feature extraction module first extracts the neighborhood pixel group of the corresponding pixel using a 3×3 mask as input. Then, it performs 1×1 grouped convolution and 3×3 grouped convolution on the input, respectively. The concatenated 1×1 and 3×3 grouped convolution results are then concatenated using a concat operation to obtain joint features. Finally, deformable convolution is performed on the joint features to obtain multi-scale neighborhood spatial features. The feature enhancement module enhances the multi-scale neighborhood spatial features to obtain enhanced neighborhood spatial features. The cross-attention fusion module is used to combine the fused spectral features and neighborhood spatial features of each pixel and input them into the fully connected layer to classify the plateau wetland landscape components of each pixel and obtain the classification results of each pixel. The pixel regions corresponding to each plateau wetland landscape component in the classified remote sensing image data were determined, and the pixel regions corresponding to each plateau wetland landscape component were extracted.

2. The method for extracting components from plateau wetland landscapes according to claim 1, characterized in that, The feature-fused remote sensing image data is a multi-level feature-fused remote sensing image data containing 6 spectral bands, 2 spectral features, and 4 spatial features. The 6 spectral bands include blue spectral band, green spectral band, red spectral band, near-infrared spectral band, first shortwave infrared spectral band, and second shortwave infrared spectral band. The 2 spectral features include normalized vegetation index feature and enhanced water body index feature. The 4 spatial features include slope topography feature, slope aspect topography feature, texture feature, and spatial statistical distribution feature.

3. The method for extracting components from plateau wetland landscapes according to claim 1, characterized in that, Before inputting the feature-fused remote sensing image data into a pre-set spectral-spatial bibranch interaction model for classification of plateau wetland landscape components, the method further includes: constructing a spectral-spatial bibranch interaction model, and using a training set, a validation set, and a test set to train, validate, and test the spectral-spatial bibranch interaction model, wherein the training set, validation set, and test set each contain several pixel samples after the plateau wetland landscape component classification and labeling.

4. The method for extracting components from plateau wetland landscapes according to claim 1, characterized in that, The spectral feature extraction branch also includes a cropping and mapping module located between the pixel upscaling module and the Transformer module. The cropping and mapping module is used to crop the upscaling feature image into blocks and then perform linear mapping. Then, block embedding and position embedding are performed on each block after linear mapping to obtain multiple sets of one-dimensional embedding sequences. Each set of one-dimensional embedding sequences is then input into the corresponding Transformer layer.

5. The method for extracting components from plateau wetland landscapes according to claim 1, characterized in that, The feature enhancement module first performs global average pooling on the input multi-scale neighborhood spatial features, then uses a 1×1 convolution kernel with ReLU activation function and a 1×1 convolution kernel with Sigmoid activation function to continuously capture the non-linear dependencies between channels in the channel dimension, and then uses a 3×3 convolution kernel to perform context modeling in the spatial dimension to obtain the enhanced neighborhood spatial features.

6. The method for extracting components from plateau wetland landscapes according to claim 1, characterized in that, The cross-attention fusion module uses a multi-head cross-attention mechanism to associate the fused spectral features of the spectral feature extraction branch with the neighborhood spatial features of the spatial feature extraction branch to obtain spectral-spatial joint features. Then, after processing the spectral-spatial joint features through three fully connected layers, the classification result of the corresponding pixel is output by the Softmax function.

7. A system for extracting components from plateau wetland landscapes, characterized in that, include: Memory, used to store instructions; A processor is configured to read instructions stored in the memory and execute the plateau wetland landscape component extraction method according to any one of claims 1-6.

8. A computer program product, characterized in that, When the computer program product is run on a computer, it executes the plateau wetland landscape component extraction method according to any one of claims 1-6.

Images