Hyperspectral wetland vegetation classification method, medium and system
Through linear unbiased randomization network and dual-adversarial regularization training methods, the problems of cross-domain generalization error and poor interpretation effect in wetland vegetation classification are solved, and the precise classification of hyperspectral wetland vegetation is realized, which is suitable for vegetation monitoring in complex environments. .
Patent Information
- Application Number
- CN202411418579.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-12
- Publication Date
- 2024-11-12
AI Technical Summary
Deep learning has problems such as large cross-domain generalization error and poor interpretation effect in wetland vegetation classification. Especially in areas with complex coastal wetland environments and many types of vegetation, it is difficult to obtain comprehensive training data, which limits the generalization and classification of the model. Accuracy.
A linear unbiased randomization network is adopted to superpixel segment the hyperspectral image through a linear iterative clustering algorithm, and a generator and discriminator with linear response and style randomization are designed to realize data generation and domain invariant characterization, using dual The adversarial regularization training method generates simulated images with a certain similarity to the training image but with spectral offset, improving the generalization and classification efficiency of the model.
It achieves better generalization and classification accuracy in cross-satellite hyperspectral image classification, reduces calculation costs, improves the reliability of the model and interpretability of classification results, and is suitable for vegetation classification in complex environments such as coastal wetlands.
Smart Images

Figure CN118941966A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of wetland vegetation classification, and in particular relates to a hyperspectral wetland vegetation classification method, medium and system. Background Art
[0002] In recent years, hyperspectral imaging technology has developed rapidly. With the launch of a series of high-resolution satellites, environmental and disaster reduction small satellites, etc., the data obtained by the imaging spectrometers independently developed by my country have been applied to scientific research and production activities. Hyperspectral remote sensing has been increasingly widely used in vegetation ecological monitoring in my country.
[0003] Deep learning algorithms have become an important means of accurately identifying wetland vegetation. However, due to factors such as the satellite's hyperspectral payload resolution, frequency band, atmospheric conditions during shooting, and seasonal changes, the spectral characteristics of the same object change, that is, there is a spectral shift, and deep learning methods require a large amount of data for training, etc., and collecting new ground truth data usually requires professional knowledge and a lot of manpower, especially in coastal wetlands, where the environment is complex and there are many types of vegetation. It is difficult to enter the field survey in most areas and obtain comprehensive training data, which limits the scale and diversity of vegetation classification model training data, resulting in large generalization errors and poor interpretation effects of deep learning in cross-domain classification tasks, hindering model sharing between satellite hyperspectral sensors and information interpretation of wetland scenes. Summary of the invention
[0004] The purpose of the present invention is to solve the deficiencies of the above-mentioned technologies and to provide a hyperspectral wetland vegetation classification method, medium and system.
[0005] In a first aspect, the present invention provides a hyperspectral wetland vegetation classification method, comprising the following steps:
[0006] S10: Obtain pixels of the hyperspectral vegetation image, intercept the three-dimensional data of a cube at the center of each pixel as spatial feature data, perform spectral feature extraction on the spatial feature data, and obtain spectral feature data of different scales;
[0007] S20: fusing the spectral feature data and the spatial feature data, and then performing secondary feature extraction to obtain fused feature data;
[0008] S30: Use global average pooling to flatten the fused feature data mapping into a feature vector, perform forward propagation training on the feature vector, and obtain a separation probability result;
[0009] S40: recalculate the loss function and back-propagate to update the parameter model, then clear all parameter gradients, recalculate the gradient of the loss function relative to the model parameters, and update the model parameters according to the recalculated gradients;
[0010] S50: loop the steps of S40 until the training reaches a convergence state and the loss function is minimized;
[0011] S60: After training, patches are constructed from the hyperspectral image, and finally the CNN-Transforme algorithm is used to obtain the wetland vegetation classification results.
[0012] In a second aspect, the present invention provides a computer-readable storage medium having program instructions stored therein for the above-mentioned hyperspectral wetland vegetation classification method.
[0013] In a third aspect, the present invention provides a hyperspectral wetland vegetation classification system, comprising the above-mentioned computer-readable storage medium.
[0014] The present invention provides a hyperspectral wetland vegetation classification method, which has the following beneficial effects:
[0015] The present invention uses a linear unbiased randomized network to solve the cross-domain hyperspectral image classification problem. Before inputting into the network, the hyperspectral image is first segmented into superpixels by the linear iterative clustering (SLIC) algorithm, and the obtained superpixel blocks are input into the network. The network includes a generator for data generation and a discriminator for learning domain-invariant representations. In the training stage, the generator and the discriminator are mutually optimized to finally train a generalized discriminator. In the testing stage, after extracting features through the generator, the discriminator is directly used for classification. In order to achieve the generalization of the network, the network proposes a generator with linear response and style randomization, which can generate simulated images with a certain degree of similarity to the training image but with spectral offset, thereby improving the generation efficiency and reliability. In addition, the dual adversarial regularization training method of the generator and the discriminator realizes non-redundant expansion, and realizes the similarity and offset between the training image and the simulated image at a low computational cost. BRIEF DESCRIPTION OF THE DRAWINGS
[0016] Figure 1 It is a structural diagram of the hyperspectral wetland vegetation classification method based on a linear unbiased randomized network of the present invention;
[0017] Figure 2 is a pseudo-color image of the GF-5 training data of the present invention;
[0018] Figure 3 It is a pseudo-color image of the XG-003 test data of the present invention;
[0019] FIG4 (a) is a visual classification result image of the present invention using the support vector machine algorithm on the GF-5 image;
[0020] FIG4( b ) is a visualized classification result image of the present invention using the K-nearest neighbor algorithm on the GF-5 image;
[0021] FIG4( c ) is a visualization classification result image of the present invention using the dual-branch CNN-Transformer algorithm on the GF-5 image;
[0022] FIG5 (a) is a visual classification result image of the XG-003 image using the support vector machine algorithm of the present invention;
[0023] FIG5( b ) is a visualized classification result image of the XG-003 image using the K-nearest neighbor algorithm of the present invention;
[0024] FIG5( c ) is a visualization classification result image of the present invention using the dual-branch CNN-Transformer algorithm on the XG-003 image;
[0025] FIG6 (a) is a visualization result image of the semi-supervised hyperspectral image classification using the K-nearest neighbor algorithm of the present invention;
[0026] FIG6 (b) is a visualization result image of the semi-supervised hyperspectral image classification using the support vector machine algorithm of the present invention;
[0027] FIG6 (c) is a visualization result image of the semi-supervised hyperspectral image classification using the dual-branch CNN-Transformer algorithm of the present invention. DETAILED DESCRIPTION
[0028] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments to help understand the content of the present invention. The methods used in the present invention are all conventional methods unless otherwise specified; the raw materials and devices used are all conventional commercial products unless otherwise specified.
[0029] 1. Network Framework
[0030] like Figure 1As shown, the present invention designs a linear unbiased randomized network to solve the cross-domain hyperspectral image classification problem. Before inputting into the network, the hyperspectral image is first segmented into superpixels by the linear iterative clustering (SLIC) algorithm, and the obtained superpixel blocks are input into the network. The network includes a generator for data generation and a discriminator for learning domain-invariant representations. In the training stage, the generator and the discriminator are mutually optimized to finally train a generalized discriminator. In the testing stage, after extracting features through the generator, the discriminator is directly used for classification. In order to achieve the generalization of the network, the network proposes a generator with linear response and style randomization, which can generate simulated images with a certain degree of similarity to the training image but with spectral offset, thereby improving the generation efficiency and reliability. In addition, the dual adversarial regularization training method of the generator and the discriminator realizes non-redundant expansion, and realizes the similarity and offset between the training image and the simulated image at a low computational cost.
[0031] 1.1 Generator
[0032] The generator is an asymmetric, masked autoencoder structure that takes into account spectral diversity. In the generator, pseudo data generated using training images are called simulated images. The generator generates simulated images by reconstructing the learned domain discriminative representations so that the simulated images contain spatial-spectral features with spectral shifts relative to the training images.
[0033] Encoder: Includes linear resampling and feature extraction based on 3D convolution. After superpixel segmentation, the hyperspectral image is segmented into superpixel image blocks of different scales. The pixels within each superpixel have the same label, so that each superpixel is used as a new sample instead of each pixel as a sample. The obtained hyperspectral image block is subjected to spectral channel scaling and spatial-spectral feature extraction through 1×1 2D convolution and 3×3×3 3D convolution to obtain the image block feature z.
[0034] Unbiased randomization: Calculate the mean and variance of the image block feature z to generate a style map x, and then randomly generate a style map y. By adaptively learning the affine transformation parameters, style map y is migrated to style map x:
[0035]
[0036] Among them, μ represents the mean of the image block feature, and σ represents the variance of the image block feature;
[0037] The core of the AdaIN (Adaptive Instance Normalization) method is a new adaptive instance normalization layer, which aligns the mean and variance of content features (here, the features of the source domain image) with the mean and variance of style features (here, the features of the target domain image). The basic idea is to normalize features from two aspects: instance normalization and adaptive adjustment.
[0038] Instance normalization is to normalize each sample (usually one sample in each batch) separately. Given a feature map, its shape is C×H×W, where C represents the number of channels, H represents the height, and W represents the width. The formula for instance normalization is as follows:
[0039]
[0040] Among them, μ(x) represents the mean on C, σ 2 (x) represents the variance, i and j represent the pixels corresponding to the i-th row and j-th column respectively. The normalized feature map X norm for:
[0041]
[0042] Here, o' is an infinitesimal positive number used to avoid division by zero.
[0043] Adaptive adjustment refers to adjusting the normalized feature map through the learned scale and offset parameters. The target domain image is input into a pre-trained neural network (such as VGG), and high-level features are extracted through several layers of the network. Statistical features such as mean and covariance matrix are calculated from the feature maps extracted from these layers. These statistical features reflect the style characteristics of the style image. These statistical features are then concatenated into a vector m as the style vector. The formula for adaptive adjustment is as follows:
[0044]
[0045] Among them, γ and β are the scale and offset parameters calculated from the style vector m, which can be learned from the style vector m through a simple fully connected layer.
[0046] Decoder: The decoding process is the inverse process of the encoding process and consists of deconvolution, which is used to upsample and restore the image block size.
[0047] To optimize the generator, the following loss function is used as a constraint for generating simulated images:
[0048]
[0049] Among them, class c is used as a positive sample, and other classes are used as negative samples. The distance within the class is continuously narrowed, and the distance between classes is increased. The effectiveness of generating simulated images is enhanced through intra-class contrast learning. I represents the Ith sample, n c represents the number of samples in the cth class, P represents the samples in the positive sample set, Pc(I) and Nc(I) represent the Ith sample in the positive sample set of the cth class and the Ith sample in the negative sample set of the cth class, respectively, and Z I represents the feature vector of the I-th sample, Zp + represents the positive sample, a represents the sample in the negative sample set, Zn - represents negative samples, and τ represents the temperature coefficient.
[0050] 1.2 Dual Discriminator
[0051] The discriminator consists of classification projection, feature projection and feature extractor, and is mainly used to extract and classify features of training images and simulated images. During the training phase, the discriminator is used to classify samples of training images, simulated images and mixed images (mixed images are weighted results of training images and simulated images), and then generate prediction results and feature projections. The generator and discriminator perform contrastive adversarial learning by generating simulated image samples and learning invariant representations.
[0052] In order to fully extract global and local spatial-spectral features, the feature extractor uses a dual-branch model. The model mainly consists of two parts, a multi-scale spectral feature extraction module and a dual-branch module based on a three-dimensional convolutional neural network (CNN) and a Transformer encoder. In order to fully extract spectral features, a multi-scale spectral feature extraction module is proposed. The multi-scale spectral feature extraction module enables the model to obtain different receptive fields to extract spectral features at different scales, enriches the extracted features, and improves the performance of the model. Based on the original Transformer encoder, an improved Transformer encoder module with convolution operation is designed, and a convolutional spectral projection unit and a convolutional multi-head self-attention unit are proposed to extract spatial and global spectral features. This module can fully integrate spatial and local-global spectral features while having low computational complexity.
[0053] Multi-scale spectral feature extraction module: This module extracts features from the original training image, the simulated image output by the generator, and the mixed image obtained by the linear combination of the two. First, a 1×1×7 three-dimensional convolution layer is used for dimensionality reduction. Then, it is divided into four parts along the channel size, which are features of different scales. Three-dimensional deep separable convolutions with different convolution kernel sizes are applied to each part of the features. Each three-dimensional deep separable convolution is followed by a ReLU function to extract spectral features of different scales. Finally, the multi-scale features are connected on the channel dimension to obtain a feature map.
[0054] Based on 3D CNN and Transformer encoder module: This module performs joint extraction of spatial and spectral features. The two branches are CNN and Transformer encoder improved by convolution operation. The Transformer encoder branch improved by convolution operation consists of convolution layer, projection layer with convolution and convolution multi-layer perceptron. By using the long-distance dependency modeling capability of Transformer, the convolution multi-head self-attention mechanism is applied to each spectral band, which can capture better global spectral features from multiple spectral bands; then the global spatial-spectral features of Transformer encoder are output through a 3×3×1 3D convolution layer of one channel. The CNN branch is relatively simple, which is calculated through 16 channels of 3×3×3 3D depth separable convolution and 1×1×1 3D point-by-point convolution, and then through a batch normalization layer, followed by a ReLU function, to output spatial and local spectral features.
[0055] Classification projection uses the Softmax function to obtain the classification result of each superpixel. Feature projection is used to calculate the loss function, and the following loss function is used as the constraint of the discriminator to optimize the discriminator:
[0056]
[0057] Among them, the training image P (I) is a positive sample, and the simulated image N (I) is a negative sample. The similarity between the sample and the training image is increased, and the similarity with the simulated image is reduced, so that the discriminator learns domain-invariant representation.
[0058] 2. Experimental Setup
[0059] 2.1 Experimental Data
[0060] The experimental area of this experiment was selected from some typical vegetation distribution areas of the Yellow River Delta Nature Reserve. In the nature reserve, Suaeda salsa, Tamarix chinensis, Phragmites australis, Spartina alterniflora and other plants grow mixedly and interlaced, and most of the plants are short and sparsely distributed. Therefore, the key identification targets of the experiment are Suaeda salsa, Spartina alterniflora, Phragmites australis and Tamarix chinensis.
[0061] The experiment was conducted on the hyperspectral images of Gaofen-5 and Xiamen Science and Technology Satellite No. 1. Gaofen-5 (GF-5) is a hyperspectral satellite used for land environment monitoring in my country's major project of high-resolution earth observation system. The GF-5 hyperspectral sensor AHSI uses two different receiving units, including 150 visible near-infrared bands (VNIR) and 180 mid-infrared bands (SWIR); Xiamen Science and Technology Satellite No. 1 (XG-003) is a city customized service satellite, a hyperspectral earth remote sensing satellite used for ecological environment, land resources and other needs. The hyperspectral camera on board works in the visible-near infrared band, with 150 channels. For technical parameters, see Table 1:
[0062] Table 1 Comparison of GF-5 and XG-003 hyperspectral imaging technology parameters
[0063]
[0064] First, the band ranges of GF-5 and XG-003 were compared, and the classification results of GF-5 wetland features were used as the true value dataset. 98 bands in the common band range (433nm-848nm) were retained. The labels corresponding to the training data included 9 categories: Spartina alterniflora, Phragmites australis, Tamarix salsa, Suaeda salsa, Phragmites australis on tidal flats, bare tidal flats, saline-alkali flats, water bodies and others. The GF-5 true value map was downsampled to 763×734 according to the image resolution as training data, as shown in Figure 2. Figure 2 As shown in the figure, the XG-003 image has a spatial resolution of 557×242, and 98 bands are retained after synchronization with the training data, as shown in the figure. Figure 3 shown.
[0065] 2.2 Experimental methods
[0066] For each pixel of the hyperspectral image, a three-dimensional data cube with a shape of P×P×S is intercepted with the pixel as the center as a sample and sent to the multi-scale spectral feature extraction module to extract spectral features. The spectral features of different scales are obtained and connected as feature maps, and then the spatial features and spectral features are jointly extracted and fused through the dual-branch CNN-Transformer module; the feature map is flattened into a vector using global average pooling, and it is forward propagated and fed into the linear layer to obtain the classification probability result, and then the loss function is calculated and the model parameters are updated by backpropagation: the gradients of all parameters are cleared, the gradients of the loss function relative to the model parameters are calculated, and the model parameters are updated according to the calculated gradients. Through this process, the model will continuously adjust the parameters in multiple iterations to minimize the loss function and finally reach a convergence state. After the training is completed, the patch is constructed from the hyperspectral image and fed into the proposed network using the dual-branch CNN-Transformer algorithm to obtain the classification result.
[0067] Some images are selected from the data in Section 2.1 for model training and testing. OA (Overall Accuracy) and AA (Average Accuracy) are used as quantitative evaluation indicators, and support vector machine is selected as the representative of classic algorithms for comparison. Since there are a small amount of labeled data and a large amount of unlabeled data in the data, the k-nearest neighbor algorithm can better predict the labels of unlabeled data points. At the same time, the k-nearest neighbor algorithm is selected for comparison of experimental results.
[0068] 3. Experimental Results and Analysis
[0069] 3.1 Supervised Classification Experiment
[0070] As shown in Table 2, the support vector machine, K-nearest neighbor algorithm and dual-branch CNN-Transformer algorithm are used to learn and classify the GF-5 and XG-003 images. The support vector machine and K-nearest neighbor algorithm are trained with the entire GF-5 image and XG-003 image respectively, and then tested with the entire GF-5 image and XG-003 image. The classification results are as follows:
[0071] Table 2 Classification performance of different algorithms on training data and GF-5 test images
[0072]
[0073] Table 3 Classification performance of different algorithms on training data and XG-003 test images
[0074]
[0075] From Table 2, Figure 4 (a)-Figure 4 (c), Figure 5 (a)-Figure 5 (c), GF-5 images and XG-003 images, the classification performance of the dual-branch CNN-Transformer algorithm is the best, followed by the K-nearest neighbor algorithm. On the GF-5 image, the three methods have the highest classification accuracy for the first category of Spartina alterniflora, and the lowest classification accuracy for the fifth category of tidal flat reed. This is mainly because the characteristics of the first category of Spartina alterniflora samples are highly distinguishable from other categories in the training data, making it easy for the model to learn these distinguishing features, while the characteristics of the fifth category of tidal flat reed samples are similar or overlapping with the second category of reed spectra, and the model may be difficult to accurately distinguish, resulting in poor classification results. The second poor classification result is the ninth category of other categories. The reason is that the number of samples in the ninth category is small, and the model is biased towards the majority class during the training process, resulting in poor classification results for the minority class.
[0076] Table 4 Legend and pixel number of coastal wetland features
[0077]
[0078] There is also an imbalanced sample distribution problem in the XG-003 image. The number of samples in categories 1, 3, 4, and 5 is much smaller than that in other categories, so the classification accuracy of these four categories is low. This problem is more obvious in the SVM algorithm. When the number of samples in different categories varies greatly, SVM tends to maximize the overall classification interval, which may lead to inaccurate classification boundaries for minority class samples because they contribute less to the interval maximization. The boundary of the minority class may be dominated by the majority class, which weakens the model's prediction ability for the minority class. The K-nearest neighbor algorithm has a good classification accuracy for each category because it is an instance-based learning method that does not build any explicit classification model, but directly determines the category of the test sample based on its nearest neighbors. This means that as long as there are samples belonging to the minority class among the neighbors of the test sample, the K-nearest neighbor algorithm can better identify the category, even if the number of samples in this category is small. The dual-branch CNN-Transformer algorithm can learn and integrate global and local spatial-spectral features, and has the best comprehensive classification performance.
[0079] 3.2 Semi-supervised classification experiments
[0080] Since the generalization of the model obtained by using only source domain images as training data is poor, semi-supervised hyperspectral image classification is further proposed in the cross-satellite object classification experiment. All training data and 70% of the test data of GF-5 and XG-003 images are used for training, and then 30% of the samples of XG-003 are tested. Compared with the image classification experiment in 3.1, the learning of additional training images in this experiment improves the average classification accuracy. The experimental results are shown in Figure 6 (a)-Figure 6 (c).
[0081] Table 5 Classification performance of different algorithms on test images
[0082]
[0083] As shown in Table 5, there is a data conflict between the training images and the test images. The performance of the SVM and K-nearest neighbor algorithms on this task is very poor. Considering that the spectral offset of water bodies in different satellite data in training and testing is small, while the spectral offset of other vegetation is large. As a result, only the classification accuracy is high in categories with a large number of samples, such as the eighth category of water bodies, reaching 0.50. The classification accuracy of other categories is extremely low, and the dual-branch CNN-Transformer algorithm has the best performance.
[0084] Randomly cropped parts of the XG-003 image were used to classify wetland features using the proposed method, and the visualization results are shown in the figure below. The image is located in the Yellow River Estuary Nature Reserve, which is mainly water-based. The tidal flat vegetation is sparse, with sporadic distribution of Spartina alterniflora on the sea side, and mixed coexistence of reeds, Tamarix, and Suaeda salsa on both sides of the estuary. The bare beach area is small, and the wetland vegetation community is healthy.
[0085] IV. Conclusion
[0086] The present invention proposes a hyperspectral wetland vegetation classification method. The cross-satellite experimental results on GF-5 and XG-003 show that the classification accuracy exceeds the existing classical methods and has good generalization. In addition, the multi-scale spectral feature extraction module added to the method in this paper obtains different receptive fields to extract spectral features at different scales, which is conducive to the extraction of fine-scale information of small-patch wetland vegetation. In subsequent work, it is planned to increase field sampling experiments to test the cross-domain classification effect of this method, and optimize the loss function according to the characteristics of wetland vegetation to improve classification accuracy. In addition, the method in this paper can be extended to the application between other small satellite hyperspectral images in my country, and can also be extended to the use of cross-satellite multispectral data. Hyperspectral wetland vegetation classification can reduce the cost of field sampling, is cost-effective, and provides refined data support for my country's wetland ecological monitoring.
[0087] The present invention provides a computer-readable storage medium, in which program instructions are stored for the above-mentioned hyperspectral wetland vegetation classification method.
[0088] The present invention provides a hyperspectral wetland vegetation classification system, comprising the above-mentioned computer-readable storage medium.
[0089] However, what is described above is only a specific embodiment of the present invention, and should not be used to limit the scope of implementation of the present invention. Therefore, the replacement of equivalent components, or equivalent changes and modifications made according to the patent protection scope of the present invention should still fall within the scope of protection of the present invention.
Claims
1. A hyperspectral wetland vegetation classification method, characterized in that: The following steps are involved: S10: Obtain pixels of the hyperspectral vegetation image, intercept the three-dimensional data of a cube at the center of each pixel as spatial feature data, extract spectral features from the spatial feature data, and obtain spectral feature data of different scales; S20: fusing the spectral feature data and the spatial feature data, and then performing secondary feature extraction to obtain fused feature data; S30: Use global average pooling to flatten the fused feature data mapping into a feature vector, perform forward propagation training on the feature vector, and obtain a separation probability result; S40: recalculate the loss function and back-propagate to update the parameter model, then clear all parameter gradients, recalculate the gradient of the loss function relative to the model parameters, and update the model parameters according to the recalculated gradients; S50: loop the steps of S40 until the training reaches a convergence state and the loss function is minimized; S60: After training, patches are constructed from the hyperspectral image, and finally the CNN-Transforme algorithm is used to obtain the wetland vegetation classification results.
2. A hyperspectral wetland vegetation classification method according to claim 1, characterized in that: The pixels of the hyperspectral vegetation image are obtained, and the three-dimensional data of the cube is intercepted at the center of each pixel as the spatial feature data. The spectral feature of the spatial feature data is extracted to obtain the spectral feature data of different scales. The specific steps are as follows: Establish a linear unbiased randomized network model, which includes a generator and a discriminator. The generator is used to extract feature data, and the discriminator is used to classify feature data. After the hyperspectral vegetation image is segmented into superpixels using a linear iterative clustering algorithm, the superpixel blocks are input into a linear unbiased randomized network model.
3. A hyperspectral wetland vegetation classification method according to claim 2, characterized in that: The generator includes an encoder and a decoder. The encoder performs spectral channel scaling and spatial feature extraction on the superpixel block to obtain the image block feature z; Calculate the mean and variance of the image block feature z to generate a style map x, then randomly generate a style map y, and migrate the random style map y to the previous style map x by adaptively learning the radial transformation parameters: ; Among them, μ represents the mean of the image block feature, and σ represents the variance of the image block feature; The mean μ and variance σ of the image block features are normalized, and the shape of the style map x is defined as C×H×W, where C represents the number of channels, H represents the height, and W represents the width: ; ; Among them, μ(x) represents the mean on C, σ 2 (x) represents the variance, i and j represent the pixel points corresponding to the i-th row and j-th column respectively; Normalized feature map X norm for: ; Where o' is an infinitesimal positive number.
4. A hyperspectral wetland vegetation classification method according to claim 3, characterized in that: The decoder consists of deconvolutions, where the following loss function is used as a constraint to generate simulated images: ; Among them, class c is regarded as a positive sample, other classes are regarded as negative samples, I represents the Ith sample, n c represents the number of samples in the cth class, P represents the samples in the positive sample set, Pc(I) and Nc(I) represent the Ith sample in the positive sample set of the cth class and the Ith sample in the negative sample set of the cth class, respectively, and Z I represents the feature vector of the I-th sample, Zp + represents the positive sample, a represents the sample in the negative sample set, Zn - represents negative samples, and τ represents the temperature coefficient.
5. A hyperspectral wetland vegetation classification method according to claim 4, characterized in that: The discriminator includes a multi-scale spectral feature extraction module and a Transformer encoder module. The multi-scale spectral feature extraction module uses a 1×1×7 three-dimensional convolution layer for dimensionality reduction and divides the channel size into four parts; Each 3D depth-wise separable convolution is followed by a ReLU function to extract spectral features of different scales.
6. A hyperspectral wetland vegetation classification method according to claim 5, characterized in that: The classification projection of the Transformer encoder module uses the Softmax function to obtain the classification result of each superpixel, and uses the following loss function as the constraint of the discriminator to optimize the discriminator: ; Among them, the training image P (I) is a positive sample, and the simulated image N (I) is a negative sample. The similarity between the sample and the training image is increased, and the similarity with the simulated image is reduced, so that the discriminator learns domain-invariant representation.
7. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores program instructions for executing the hyperspectral wetland vegetation classification method according to any one of claims 1 to 6.
8. A hyperspectral wetland vegetation classification system, characterized in that: A computer readable storage medium comprising claim 7.
Citation Information
Patent Citations
Small sample hyperspectral image classification method, system and device and storage medium
CN116977693A
Hyperspectral image classification method and system based on lightweight multi-scale Transform network
CN117710818A
Hyperspectral image tree species classification method and device
CN117876793A
Hyperspectral and laser radar multilayer fusion classification method based on adversarial learning
CN117934978A
Cross-scene hyperspectral image classification method and device, equipment and storage medium
CN118247576A
Cited By
Crop growth monitoring method and system based on agricultural Internet of Things
CN120429737A
Vegetation species identification method for enhancing AI computing power and time sequence tracing
CN121033530A
